qmd/src/store.ts

/**
 * QMD Store - Core data access and retrieval functions
 *
 * This module provides all database operations, search functions, and document
 * retrieval for QMD. It returns raw data structures that can be formatted by
 * CLI or MCP consumers.
 *
 * Usage:
 *   const store = createStore("/path/to/db.sqlite");
 *   // or use default path:
 *   const store = createStore();
 */

import { openDatabase, loadSqliteVec } from "./db.js";
import type { Database } from "./db.js";
import picomatch from "picomatch";
import { createHash } from "crypto";
import { realpathSync, statSync, mkdirSync } from "node:fs";
import {
  LlamaCpp,
  getDefaultLlamaCpp,
  formatQueryForEmbedding,
  formatDocForEmbedding,
  type RerankDocument,
  type ILLMSession,
} from "./llm.js";
import {
  findContextForPath as collectionsFindContextForPath,
  addContext as collectionsAddContext,
  removeContext as collectionsRemoveContext,
  listAllContexts as collectionsListAllContexts,
  getCollection,
  listCollections as collectionsListCollections,
  addCollection as collectionsAddCollection,
  removeCollection as collectionsRemoveCollection,
  renameCollection as collectionsRenameCollection,
  setGlobalContext,
  loadConfig as collectionsLoadConfig,
  type NamedCollection,
} from "./collections.js";

// =============================================================================
// Configuration
// =============================================================================

const HOME = process.env.HOME || "/tmp";
export const DEFAULT_EMBED_MODEL = "embeddinggemma";
export const DEFAULT_RERANK_MODEL = "ExpedientFalcon/qwen3-reranker:0.6b-q8_0";
export const DEFAULT_QUERY_MODEL = "Qwen/Qwen3-1.7B";
export const DEFAULT_GLOB = "**/*.md";
export const DEFAULT_MULTI_GET_MAX_BYTES = 10 * 1024; // 10KB

// Chunking: 900 tokens per chunk with 15% overlap
// Increased from 800 to accommodate smart chunking finding natural break points
export const CHUNK_SIZE_TOKENS = 900;
export const CHUNK_OVERLAP_TOKENS = Math.floor(CHUNK_SIZE_TOKENS * 0.15);  // 135 tokens (15% overlap)
// Fallback char-based approximation for sync chunking (~4 chars per token)
export const CHUNK_SIZE_CHARS = CHUNK_SIZE_TOKENS * 4;  // 3600 chars
export const CHUNK_OVERLAP_CHARS = CHUNK_OVERLAP_TOKENS * 4;  // 540 chars
// Search window for finding optimal break points (in tokens, ~200 tokens)
export const CHUNK_WINDOW_TOKENS = 200;
export const CHUNK_WINDOW_CHARS = CHUNK_WINDOW_TOKENS * 4;  // 800 chars

// =============================================================================
// Smart Chunking - Break Point Detection
// =============================================================================

/**
 * A potential break point in the document with a base score indicating quality.
 */
export interface BreakPoint {
  pos: number;    // character position
  score: number;  // base score (higher = better break point)
  type: string;   // for debugging: 'h1', 'h2', 'blank', etc.
}

/**
 * A region where a code fence exists (between ``` markers).
 * We should never split inside a code fence.
 */
export interface CodeFenceRegion {
  start: number;  // position of opening ```
  end: number;    // position of closing ``` (or document end if unclosed)
}

/**
 * Patterns for detecting break points in markdown documents.
 * Higher scores indicate better places to split.
 * Scores are spread wide so headings decisively beat lower-quality breaks.
 * Order matters for scoring - more specific patterns first.
 */
export const BREAK_PATTERNS: [RegExp, number, string][] = [
  [/\n#{1}(?!#)/g, 100, 'h1'],     // # but not ##
  [/\n#{2}(?!#)/g, 90, 'h2'],      // ## but not ###
  [/\n#{3}(?!#)/g, 80, 'h3'],      // ### but not ####
  [/\n#{4}(?!#)/g, 70, 'h4'],      // #### but not #####
  [/\n#{5}(?!#)/g, 60, 'h5'],      // ##### but not ######
  [/\n#{6}(?!#)/g, 50, 'h6'],      // ######
  [/\n```/g, 80, 'codeblock'],     // code block boundary (same as h3)
  [/\n(?:---|\*\*\*|___)\s*\n/g, 60, 'hr'],  // horizontal rule
  [/\n\n+/g, 20, 'blank'],         // paragraph boundary
  [/\n[-*]\s/g, 5, 'list'],        // unordered list item
  [/\n\d+\.\s/g, 5, 'numlist'],    // ordered list item
  [/\n/g, 1, 'newline'],           // minimal break
];

/**
 * Scan text for all potential break points.
 * Returns sorted array of break points with higher-scoring patterns taking precedence
 * when multiple patterns match the same position.
 */
export function scanBreakPoints(text: string): BreakPoint[] {
  const points: BreakPoint[] = [];
  const seen = new Map<number, BreakPoint>();  // pos -> best break point at that pos

  for (const [pattern, score, type] of BREAK_PATTERNS) {
    for (const match of text.matchAll(pattern)) {
      const pos = match.index!;
      const existing = seen.get(pos);
      // Keep higher score if position already seen
      if (!existing || score > existing.score) {
        const bp = { pos, score, type };
        seen.set(pos, bp);
      }
    }
  }

  // Convert to array and sort by position
  for (const bp of seen.values()) {
    points.push(bp);
  }
  return points.sort((a, b) => a.pos - b.pos);
}

/**
 * Find all code fence regions in the text.
 * Code fences are delimited by ``` and we should never split inside them.
 */
export function findCodeFences(text: string): CodeFenceRegion[] {
  const regions: CodeFenceRegion[] = [];
  const fencePattern = /\n```/g;
  let inFence = false;
  let fenceStart = 0;

  for (const match of text.matchAll(fencePattern)) {
    if (!inFence) {
      fenceStart = match.index!;
      inFence = true;
    } else {
      regions.push({ start: fenceStart, end: match.index! + match[0].length });
      inFence = false;
    }
  }

  // Handle unclosed fence - extends to end of document
  if (inFence) {
    regions.push({ start: fenceStart, end: text.length });
  }

  return regions;
}

/**
 * Check if a position is inside a code fence region.
 */
export function isInsideCodeFence(pos: number, fences: CodeFenceRegion[]): boolean {
  return fences.some(f => pos > f.start && pos < f.end);
}

/**
 * Find the best cut position using scored break points with distance decay.
 *
 * Uses squared distance for gentler early decay - headings far back still win
 * over low-quality breaks near the target.
 *
 * @param breakPoints - Pre-scanned break points from scanBreakPoints()
 * @param targetCharPos - The ideal cut position (e.g., maxChars boundary)
 * @param windowChars - How far back to search for break points (default ~200 tokens)
 * @param decayFactor - How much to penalize distance (0.7 = 30% score at window edge)
 * @param codeFences - Code fence regions to avoid splitting inside
 * @returns The best position to cut at
 */
export function findBestCutoff(
  breakPoints: BreakPoint[],
  targetCharPos: number,
  windowChars: number = CHUNK_WINDOW_CHARS,
  decayFactor: number = 0.7,
  codeFences: CodeFenceRegion[] = []
): number {
  const windowStart = targetCharPos - windowChars;
  let bestScore = -1;
  let bestPos = targetCharPos;

  for (const bp of breakPoints) {
    if (bp.pos < windowStart) continue;
    if (bp.pos > targetCharPos) break;  // sorted, so we can stop

    // Skip break points inside code fences
    if (isInsideCodeFence(bp.pos, codeFences)) continue;

    const distance = targetCharPos - bp.pos;
    // Squared distance decay: gentle early, steep late
    // At target: multiplier = 1.0
    // At 25% back: multiplier = 0.956
    // At 50% back: multiplier = 0.825
    // At 75% back: multiplier = 0.606
    // At window edge: multiplier = 0.3
    const normalizedDist = distance / windowChars;
    const multiplier = 1.0 - (normalizedDist * normalizedDist) * decayFactor;
    const finalScore = bp.score * multiplier;

    if (finalScore > bestScore) {
      bestScore = finalScore;
      bestPos = bp.pos;
    }
  }

  return bestPos;
}

// Hybrid query: strong BM25 signal detection thresholds
// Skip expensive LLM expansion when top result is strong AND clearly separated from runner-up
export const STRONG_SIGNAL_MIN_SCORE = 0.85;
export const STRONG_SIGNAL_MIN_GAP = 0.15;
// Max candidates to pass to reranker — balances quality vs latency.
// 40 keeps rank 31-40 visible to the reranker (matters for recall on broad queries).
export const RERANK_CANDIDATE_LIMIT = 40;

/**
 * A typed query expansion result. Decoupled from llm.ts internal Queryable —
 * same shape, but store.ts owns its own public API type.
 *
 * - lex: keyword variant → routes to FTS only
 * - vec: semantic variant → routes to vector only
 * - hyde: hypothetical document → routes to vector only
 */
export type ExpandedQuery = {
  type: 'lex' | 'vec' | 'hyde';
  text: string;
};

// =============================================================================
// Path utilities
// =============================================================================

export function homedir(): string {
  return HOME;
}

/**
 * Check if a path is absolute.
 * Supports:
 * - Unix paths: /path/to/file
 * - Windows native: C:\path or C:/path
 * - Git Bash: /c/path or /C/path (C-Z drives, excluding A/B floppy drives)
 * 
 * Note: /c without trailing slash is treated as Unix path (directory named "c"),
 * while /c/ or /c/path are treated as Git Bash paths (C: drive).
 */
export function isAbsolutePath(path: string): boolean {
  if (!path) return false;
  
  // Unix absolute path
  if (path.startsWith('/')) {
    // Check if it's a Git Bash style path like /c/ or /c/Users (C-Z only, not A or B)
    // Requires path[2] === '/' to distinguish from Unix paths like /c or /cache
    if (path.length >= 3 && path[2] === '/') {
      const driveLetter = path[1];
      if (driveLetter && /[c-zC-Z]/.test(driveLetter)) {
        return true;
      }
    }
    // Any other path starting with / is Unix absolute
    return true;
  }
  
  // Windows native path: C:\ or C:/ (any letter A-Z)
  if (path.length >= 2 && /[a-zA-Z]/.test(path[0]!) && path[1] === ':') {
    return true;
  }
  
  return false;
}

/**
 * Normalize path separators to forward slashes.
 * Converts Windows backslashes to forward slashes.
 */
export function normalizePathSeparators(path: string): string {
  return path.replace(/\\/g, '/');
}

/**
 * Get the relative path from a prefix.
 * Returns null if path is not under prefix.
 * Returns empty string if path equals prefix.
 */
export function getRelativePathFromPrefix(path: string, prefix: string): string | null {
  // Empty prefix is invalid
  if (!prefix) {
    return null;
  }
  
  const normalizedPath = normalizePathSeparators(path);
  const normalizedPrefix = normalizePathSeparators(prefix);
  
  // Ensure prefix ends with / for proper matching
  const prefixWithSlash = !normalizedPrefix.endsWith('/') 
    ? normalizedPrefix + '/' 
    : normalizedPrefix;
  
  // Exact match
  if (normalizedPath === normalizedPrefix) {
    return '';
  }
  
  // Check if path starts with prefix
  if (normalizedPath.startsWith(prefixWithSlash)) {
    return normalizedPath.slice(prefixWithSlash.length);
  }
  
  return null;
}

export function resolve(...paths: string[]): string {
  if (paths.length === 0) {
    throw new Error("resolve: at least one path segment is required");
  }
  
  // Normalize all paths to use forward slashes
  const normalizedPaths = paths.map(normalizePathSeparators);
  
  let result = '';
  let windowsDrive = '';
  
  // Check if first path is absolute
  const firstPath = normalizedPaths[0]!;
  if (isAbsolutePath(firstPath)) {
    result = firstPath;
    
    // Extract Windows drive letter if present
    if (firstPath.length >= 2 && /[a-zA-Z]/.test(firstPath[0]!) && firstPath[1] === ':') {
      windowsDrive = firstPath.slice(0, 2);
      result = firstPath.slice(2);
    } else if (firstPath.startsWith('/') && firstPath.length >= 3 && firstPath[2] === '/') {
      // Git Bash style: /c/ -> C: (C-Z drives only, not A or B)
      const driveLetter = firstPath[1];
      if (driveLetter && /[c-zC-Z]/.test(driveLetter)) {
        windowsDrive = driveLetter.toUpperCase() + ':';
        result = firstPath.slice(2);
      }
    }
  } else {
    // Start with PWD or cwd, then append the first relative path
    const pwd = normalizePathSeparators(process.env.PWD || process.cwd());
    
    // Extract Windows drive from PWD if present
    if (pwd.length >= 2 && /[a-zA-Z]/.test(pwd[0]!) && pwd[1] === ':') {
      windowsDrive = pwd.slice(0, 2);
      result = pwd.slice(2) + '/' + firstPath;
    } else {
      result = pwd + '/' + firstPath;
    }
  }
  
  // Process remaining paths
  for (let i = 1; i < normalizedPaths.length; i++) {
    const p = normalizedPaths[i]!;
    if (isAbsolutePath(p)) {
      // Absolute path replaces everything
      result = p;
      
      // Update Windows drive if present
      if (p.length >= 2 && /[a-zA-Z]/.test(p[0]!) && p[1] === ':') {
        windowsDrive = p.slice(0, 2);
        result = p.slice(2);
      } else if (p.startsWith('/') && p.length >= 3 && p[2] === '/') {
        // Git Bash style (C-Z drives only, not A or B)
        const driveLetter = p[1];
        if (driveLetter && /[c-zC-Z]/.test(driveLetter)) {
          windowsDrive = driveLetter.toUpperCase() + ':';
          result = p.slice(2);
        } else {
          windowsDrive = '';
        }
      } else {
        windowsDrive = '';
      }
    } else {
      // Relative path - append
      result = result + '/' + p;
    }
  }
  
  // Normalize . and .. components
  const parts = result.split('/').filter(Boolean);
  const normalized: string[] = [];
  for (const part of parts) {
    if (part === '..') {
      normalized.pop();
    } else if (part !== '.') {
      normalized.push(part);
    }
  }
  
  // Build final path
  const finalPath = '/' + normalized.join('/');
  
  // Prepend Windows drive if present
  if (windowsDrive) {
    return windowsDrive + finalPath;
  }
  
  return finalPath;
}

// Flag to indicate production mode (set by qmd.ts at startup)
let _productionMode = false;

export function enableProductionMode(): void {
  _productionMode = true;
}

export function getDefaultDbPath(indexName: string = "index"): string {
  // Always allow override via INDEX_PATH (for testing)
  if (process.env.INDEX_PATH) {
    return process.env.INDEX_PATH;
  }

  // In non-production mode (tests), require explicit path
  if (!_productionMode) {
    throw new Error(
      "Database path not set. Tests must set INDEX_PATH env var or use createStore() with explicit path. " +
      "This prevents tests from accidentally writing to the global index."
    );
  }

  const cacheDir = process.env.XDG_CACHE_HOME || resolve(homedir(), ".cache");
  const qmdCacheDir = resolve(cacheDir, "qmd");
  try { mkdirSync(qmdCacheDir, { recursive: true }); } catch { }
  return resolve(qmdCacheDir, `${indexName}.sqlite`);
}

export function getPwd(): string {
  return process.env.PWD || process.cwd();
}

export function getRealPath(path: string): string {
  try {
    return realpathSync(path);
  } catch {
    return resolve(path);
  }
}

// =============================================================================
// Virtual Path Utilities (qmd://)
// =============================================================================

export type VirtualPath = {
  collectionName: string;
  path: string;  // relative path within collection
};

/**
 * Normalize explicit virtual path formats to standard qmd:// format.
 * Only handles paths that are already explicitly virtual:
 * - qmd://collection/path.md (already normalized)
 * - qmd:////collection/path.md (extra slashes - normalize)
 * - //collection/path.md (missing qmd: prefix - add it)
 *
 * Does NOT handle:
 * - collection/path.md (bare paths - could be filesystem relative)
 * - :linenum suffix (should be parsed separately before calling this)
 */
export function normalizeVirtualPath(input: string): string {
  let path = input.trim();

  // Handle qmd:// with extra slashes: qmd:////collection/path -> qmd://collection/path
  if (path.startsWith('qmd:')) {
    // Remove qmd: prefix and normalize slashes
    path = path.slice(4);
    // Remove leading slashes and re-add exactly two
    path = path.replace(/^\/+/, '');
    return `qmd://${path}`;
  }

  // Handle //collection/path (missing qmd: prefix)
  if (path.startsWith('//')) {
    path = path.replace(/^\/+/, '');
    return `qmd://${path}`;
  }

  // Return as-is for other cases (filesystem paths, docids, bare collection/path, etc.)
  return path;
}

/**
 * Parse a virtual path like "qmd://collection-name/path/to/file.md"
 * into its components.
 * Also supports collection root: "qmd://collection-name/" or "qmd://collection-name"
 */
export function parseVirtualPath(virtualPath: string): VirtualPath | null {
  // Normalize the path first
  const normalized = normalizeVirtualPath(virtualPath);

  // Match: qmd://collection-name[/optional-path]
  // Allows: qmd://name, qmd://name/, qmd://name/path
  const match = normalized.match(/^qmd:\/\/([^\/]+)\/?(.*)$/);
  if (!match?.[1]) return null;
  return {
    collectionName: match[1],
    path: match[2] ?? '',  // Empty string for collection root
  };
}

/**
 * Build a virtual path from collection name and relative path.
 */
export function buildVirtualPath(collectionName: string, path: string): string {
  return `qmd://${collectionName}/${path}`;
}

/**
 * Check if a path is explicitly a virtual path.
 * Only recognizes explicit virtual path formats:
 * - qmd://collection/path.md
 * - //collection/path.md
 *
 * Does NOT consider bare collection/path.md as virtual - that should be
 * handled separately by checking if the first component is a collection name.
 */
export function isVirtualPath(path: string): boolean {
  const trimmed = path.trim();

  // Explicit qmd:// prefix (with any number of slashes)
  if (trimmed.startsWith('qmd:')) return true;

  // //collection/path format (missing qmd: prefix)
  if (trimmed.startsWith('//')) return true;

  return false;
}

/**
 * Resolve a virtual path to absolute filesystem path.
 */
export function resolveVirtualPath(db: Database, virtualPath: string): string | null {
  const parsed = parseVirtualPath(virtualPath);
  if (!parsed) return null;

  const coll = getCollectionByName(db, parsed.collectionName);
  if (!coll) return null;

  return resolve(coll.pwd, parsed.path);
}

/**
 * Convert an absolute filesystem path to a virtual path.
 * Returns null if the file is not in any indexed collection.
 */
export function toVirtualPath(db: Database, absolutePath: string): string | null {
  // Get all collections from YAML config
  const collections = collectionsListCollections();

  // Find which collection this absolute path belongs to
  for (const coll of collections) {
    if (absolutePath.startsWith(coll.path + '/') || absolutePath === coll.path) {
      // Extract relative path
      const relativePath = absolutePath.startsWith(coll.path + '/')
        ? absolutePath.slice(coll.path.length + 1)
        : '';

      // Verify this document exists in the database
      const doc = db.prepare(`
        SELECT d.path
        FROM documents d
        WHERE d.collection = ? AND d.path = ? AND d.active = 1
        LIMIT 1
      `).get(coll.name, relativePath) as { path: string } | null;

      if (doc) {
        return buildVirtualPath(coll.name, relativePath);
      }
    }
  }

  return null;
}

// =============================================================================
// Database initialization
// =============================================================================


function createSqliteVecUnavailableError(reason: string): Error {
  return new Error(
    "sqlite-vec extension is unavailable. " +
    `${reason}. ` +
    "Install Homebrew SQLite so the sqlite-vec extension can be loaded, " +
    "and set BREW_PREFIX if Homebrew is installed in a non-standard location."
  );
}

function getErrorMessage(err: unknown): string {
  return err instanceof Error ? err.message : String(err);
}

export function verifySqliteVecLoaded(db: Database): void {
  try {
    const row = db.prepare(`SELECT vec_version() AS version`).get() as { version?: string } | null;
    if (!row?.version || typeof row.version !== "string") {
      throw new Error("vec_version() returned no version");
    }
  } catch (err) {
    const message = getErrorMessage(err);
    throw createSqliteVecUnavailableError(`sqlite-vec probe failed (${message})`);
  }
}

let _sqliteVecAvailable: boolean | null = null;

function initializeDatabase(db: Database): void {
  try {
    loadSqliteVec(db);
    verifySqliteVecLoaded(db);
    _sqliteVecAvailable = true;
  } catch {
    // sqlite-vec is optional — vector search won't work but FTS is fine
    _sqliteVecAvailable = false;
  }
  db.exec("PRAGMA journal_mode = WAL");
  db.exec("PRAGMA foreign_keys = ON");

  // Drop legacy tables that are now managed in YAML
  db.exec(`DROP TABLE IF EXISTS path_contexts`);
  db.exec(`DROP TABLE IF EXISTS collections`);

  // Content-addressable storage - the source of truth for document content
  db.exec(`
    CREATE TABLE IF NOT EXISTS content (
      hash TEXT PRIMARY KEY,
      doc TEXT NOT NULL,
      created_at TEXT NOT NULL
    )
  `);

  // Documents table - file system layer mapping virtual paths to content hashes
  // Collections are now managed in ~/.config/qmd/index.yml
  db.exec(`
    CREATE TABLE IF NOT EXISTS documents (
      id INTEGER PRIMARY KEY AUTOINCREMENT,
      collection TEXT NOT NULL,
      path TEXT NOT NULL,
      title TEXT NOT NULL,
      hash TEXT NOT NULL,
      created_at TEXT NOT NULL,
      modified_at TEXT NOT NULL,
      active INTEGER NOT NULL DEFAULT 1,
      FOREIGN KEY (hash) REFERENCES content(hash) ON DELETE CASCADE,
      UNIQUE(collection, path)
    )
  `);

  db.exec(`CREATE INDEX IF NOT EXISTS idx_documents_collection ON documents(collection, active)`);
  db.exec(`CREATE INDEX IF NOT EXISTS idx_documents_hash ON documents(hash)`);
  db.exec(`CREATE INDEX IF NOT EXISTS idx_documents_path ON documents(path, active)`);

  // Cache table for LLM API calls
  db.exec(`
    CREATE TABLE IF NOT EXISTS llm_cache (
      hash TEXT PRIMARY KEY,
      result TEXT NOT NULL,
      created_at TEXT NOT NULL
    )
  `);

  // Content vectors
  const cvInfo = db.prepare(`PRAGMA table_info(content_vectors)`).all() as { name: string }[];
  const hasSeqColumn = cvInfo.some(col => col.name === 'seq');
  if (cvInfo.length > 0 && !hasSeqColumn) {
    db.exec(`DROP TABLE IF EXISTS content_vectors`);
    db.exec(`DROP TABLE IF EXISTS vectors_vec`);
  }
  db.exec(`
    CREATE TABLE IF NOT EXISTS content_vectors (
      hash TEXT NOT NULL,
      seq INTEGER NOT NULL DEFAULT 0,
      pos INTEGER NOT NULL DEFAULT 0,
      model TEXT NOT NULL,
      embedded_at TEXT NOT NULL,
      PRIMARY KEY (hash, seq)
    )
  `);

  // FTS - index filepath (collection/path), title, and content
  db.exec(`
    CREATE VIRTUAL TABLE IF NOT EXISTS documents_fts USING fts5(
      filepath, title, body,
      tokenize='porter unicode61'
    )
  `);

  // Triggers to keep FTS in sync
  db.exec(`
    CREATE TRIGGER IF NOT EXISTS documents_ai AFTER INSERT ON documents
    WHEN new.active = 1
    BEGIN
      INSERT INTO documents_fts(rowid, filepath, title, body)
      SELECT
        new.id,
        new.collection || '/' || new.path,
        new.title,
        (SELECT doc FROM content WHERE hash = new.hash)
      WHERE new.active = 1;
    END
  `);

  db.exec(`
    CREATE TRIGGER IF NOT EXISTS documents_ad AFTER DELETE ON documents BEGIN
      DELETE FROM documents_fts WHERE rowid = old.id;
    END
  `);

  db.exec(`
    CREATE TRIGGER IF NOT EXISTS documents_au AFTER UPDATE ON documents
    BEGIN
      -- Delete from FTS if no longer active
      DELETE FROM documents_fts WHERE rowid = old.id AND new.active = 0;

      -- Update FTS if still/newly active
      INSERT OR REPLACE INTO documents_fts(rowid, filepath, title, body)
      SELECT
        new.id,
        new.collection || '/' || new.path,
        new.title,
        (SELECT doc FROM content WHERE hash = new.hash)
      WHERE new.active = 1;
    END
  `);
}


export function isSqliteVecAvailable(): boolean {
  return _sqliteVecAvailable === true;
}

function ensureVecTableInternal(db: Database, dimensions: number): void {
  if (!_sqliteVecAvailable) {
    throw new Error("sqlite-vec is not available. Vector operations require a SQLite build with extension loading support.");
  }
  const tableInfo = db.prepare(`SELECT sql FROM sqlite_master WHERE type='table' AND name='vectors_vec'`).get() as { sql: string } | null;
  if (tableInfo) {
    const match = tableInfo.sql.match(/float\[(\d+)\]/);
    const hasHashSeq = tableInfo.sql.includes('hash_seq');
    const hasCosine = tableInfo.sql.includes('distance_metric=cosine');
    const existingDims = match?.[1] ? parseInt(match[1], 10) : null;
    if (existingDims === dimensions && hasHashSeq && hasCosine) return;
    // Table exists but wrong schema - need to rebuild
    db.exec("DROP TABLE IF EXISTS vectors_vec");
  }
  db.exec(`CREATE VIRTUAL TABLE vectors_vec USING vec0(hash_seq TEXT PRIMARY KEY, embedding float[${dimensions}] distance_metric=cosine)`);
}

// =============================================================================
// Store Factory
// =============================================================================

export type Store = {
  db: Database;
  dbPath: string;
  close: () => void;
  ensureVecTable: (dimensions: number) => void;

  // Index health
  getHashesNeedingEmbedding: () => number;
  getIndexHealth: () => IndexHealthInfo;
  getStatus: () => IndexStatus;

  // Caching
  getCacheKey: typeof getCacheKey;
  getCachedResult: (cacheKey: string) => string | null;
  setCachedResult: (cacheKey: string, result: string) => void;
  clearCache: () => void;

  // Cleanup and maintenance
  deleteLLMCache: () => number;
  deleteInactiveDocuments: () => number;
  cleanupOrphanedContent: () => number;
  cleanupOrphanedVectors: () => number;
  vacuumDatabase: () => void;

  // Context
  getContextForFile: (filepath: string) => string | null;
  getContextForPath: (collectionName: string, path: string) => string | null;
  getCollectionByName: (name: string) => { name: string; pwd: string; glob_pattern: string } | null;
  getCollectionsWithoutContext: () => { name: string; pwd: string; doc_count: number }[];
  getTopLevelPathsWithoutContext: (collectionName: string) => string[];

  // Virtual paths
  parseVirtualPath: typeof parseVirtualPath;
  buildVirtualPath: typeof buildVirtualPath;
  isVirtualPath: typeof isVirtualPath;
  resolveVirtualPath: (virtualPath: string) => string | null;
  toVirtualPath: (absolutePath: string) => string | null;

  // Search
  searchFTS: (query: string, limit?: number, collectionName?: string) => SearchResult[];
  searchVec: (query: string, model: string, limit?: number, collectionName?: string, session?: ILLMSession, precomputedEmbedding?: number[]) => Promise<SearchResult[]>;

  // Query expansion & reranking
  expandQuery: (query: string, model?: string) => Promise<ExpandedQuery[]>;
  rerank: (query: string, documents: { file: string; text: string }[], model?: string) => Promise<{ file: string; score: number }[]>;

  // Document retrieval
  findDocument: (filename: string, options?: { includeBody?: boolean }) => DocumentResult | DocumentNotFound;
  getDocumentBody: (doc: DocumentResult | { filepath: string }, fromLine?: number, maxLines?: number) => string | null;
  findDocuments: (pattern: string, options?: { includeBody?: boolean; maxBytes?: number }) => { docs: MultiGetResult[]; errors: string[] };

  // Fuzzy matching and docid lookup
  findSimilarFiles: (query: string, maxDistance?: number, limit?: number) => string[];
  matchFilesByGlob: (pattern: string) => { filepath: string; displayPath: string; bodyLength: number }[];
  findDocumentByDocid: (docid: string) => { filepath: string; hash: string } | null;

  // Document indexing operations
  insertContent: (hash: string, content: string, createdAt: string) => void;
  insertDocument: (collectionName: string, path: string, title: string, hash: string, createdAt: string, modifiedAt: string) => void;
  findActiveDocument: (collectionName: string, path: string) => { id: number; hash: string; title: string } | null;
  updateDocumentTitle: (documentId: number, title: string, modifiedAt: string) => void;
  updateDocument: (documentId: number, title: string, hash: string, modifiedAt: string) => void;
  deactivateDocument: (collectionName: string, path: string) => void;
  getActiveDocumentPaths: (collectionName: string) => string[];

  // Vector/embedding operations
  getHashesForEmbedding: () => { hash: string; body: string; path: string }[];
  clearAllEmbeddings: () => void;
  insertEmbedding: (hash: string, seq: number, pos: number, embedding: Float32Array, model: string, embeddedAt: string) => void;
};

/**
 * Create a new store instance with the given database path.
 * If no path is provided, uses the default path (~/.cache/qmd/index.sqlite).
 *
 * @param dbPath - Path to the SQLite database file
 * @returns Store instance with all methods bound to the database
 */
export function createStore(dbPath?: string): Store {
  const resolvedPath = dbPath || getDefaultDbPath();
  const db = openDatabase(resolvedPath);
  initializeDatabase(db);

  return {
    db,
    dbPath: resolvedPath,
    close: () => db.close(),
    ensureVecTable: (dimensions: number) => ensureVecTableInternal(db, dimensions),

    // Index health
    getHashesNeedingEmbedding: () => getHashesNeedingEmbedding(db),
    getIndexHealth: () => getIndexHealth(db),
    getStatus: () => getStatus(db),

    // Caching
    getCacheKey,
    getCachedResult: (cacheKey: string) => getCachedResult(db, cacheKey),
    setCachedResult: (cacheKey: string, result: string) => setCachedResult(db, cacheKey, result),
    clearCache: () => clearCache(db),

    // Cleanup and maintenance
    deleteLLMCache: () => deleteLLMCache(db),
    deleteInactiveDocuments: () => deleteInactiveDocuments(db),
    cleanupOrphanedContent: () => cleanupOrphanedContent(db),
    cleanupOrphanedVectors: () => cleanupOrphanedVectors(db),
    vacuumDatabase: () => vacuumDatabase(db),

    // Context
    getContextForFile: (filepath: string) => getContextForFile(db, filepath),
    getContextForPath: (collectionName: string, path: string) => getContextForPath(db, collectionName, path),
    getCollectionByName: (name: string) => getCollectionByName(db, name),
    getCollectionsWithoutContext: () => getCollectionsWithoutContext(db),
    getTopLevelPathsWithoutContext: (collectionName: string) => getTopLevelPathsWithoutContext(db, collectionName),

    // Virtual paths
    parseVirtualPath,
    buildVirtualPath,
    isVirtualPath,
    resolveVirtualPath: (virtualPath: string) => resolveVirtualPath(db, virtualPath),
    toVirtualPath: (absolutePath: string) => toVirtualPath(db, absolutePath),

    // Search
    searchFTS: (query: string, limit?: number, collectionName?: string) => searchFTS(db, query, limit, collectionName),
    searchVec: (query: string, model: string, limit?: number, collectionName?: string, session?: ILLMSession, precomputedEmbedding?: number[]) => searchVec(db, query, model, limit, collectionName, session, precomputedEmbedding),

    // Query expansion & reranking
    expandQuery: (query: string, model?: string) => expandQuery(query, model, db),
    rerank: (query: string, documents: { file: string; text: string }[], model?: string) => rerank(query, documents, model, db),

    // Document retrieval
    findDocument: (filename: string, options?: { includeBody?: boolean }) => findDocument(db, filename, options),
    getDocumentBody: (doc: DocumentResult | { filepath: string }, fromLine?: number, maxLines?: number) => getDocumentBody(db, doc, fromLine, maxLines),
    findDocuments: (pattern: string, options?: { includeBody?: boolean; maxBytes?: number }) => findDocuments(db, pattern, options),

    // Fuzzy matching and docid lookup
    findSimilarFiles: (query: string, maxDistance?: number, limit?: number) => findSimilarFiles(db, query, maxDistance, limit),
    matchFilesByGlob: (pattern: string) => matchFilesByGlob(db, pattern),
    findDocumentByDocid: (docid: string) => findDocumentByDocid(db, docid),

    // Document indexing operations
    insertContent: (hash: string, content: string, createdAt: string) => insertContent(db, hash, content, createdAt),
    insertDocument: (collectionName: string, path: string, title: string, hash: string, createdAt: string, modifiedAt: string) => insertDocument(db, collectionName, path, title, hash, createdAt, modifiedAt),
    findActiveDocument: (collectionName: string, path: string) => findActiveDocument(db, collectionName, path),
    updateDocumentTitle: (documentId: number, title: string, modifiedAt: string) => updateDocumentTitle(db, documentId, title, modifiedAt),
    updateDocument: (documentId: number, title: string, hash: string, modifiedAt: string) => updateDocument(db, documentId, title, hash, modifiedAt),
    deactivateDocument: (collectionName: string, path: string) => deactivateDocument(db, collectionName, path),
    getActiveDocumentPaths: (collectionName: string) => getActiveDocumentPaths(db, collectionName),

    // Vector/embedding operations
    getHashesForEmbedding: () => getHashesForEmbedding(db),
    clearAllEmbeddings: () => clearAllEmbeddings(db),
    insertEmbedding: (hash: string, seq: number, pos: number, embedding: Float32Array, model: string, embeddedAt: string) => insertEmbedding(db, hash, seq, pos, embedding, model, embeddedAt),
  };
}

// =============================================================================
// Core Document Type
// =============================================================================

/**
 * Unified document result type with all metadata.
 * Body is optional - use getDocumentBody() to load it separately if needed.
 */
export type DocumentResult = {
  filepath: string;           // Full filesystem path
  displayPath: string;        // Short display path (e.g., "docs/readme.md")
  title: string;              // Document title (from first heading or filename)
  context: string | null;     // Folder context description if configured
  hash: string;               // Content hash for caching/change detection
  docid: string;              // Short docid (first 6 chars of hash) for quick reference
  collectionName: string;     // Parent collection name
  modifiedAt: string;         // Last modification timestamp
  bodyLength: number;         // Body length in bytes (useful before loading)
  body?: string;              // Document body (optional, load with getDocumentBody)
};

/**
 * Extract short docid from a full hash (first 6 characters).
 */
export function getDocid(hash: string): string {
  return hash.slice(0, 6);
}

/**
 * Handelize a filename to be more token-friendly.
 * - Convert triple underscore `___` to `/` (folder separator)
 * - Convert to lowercase
 * - Replace sequences of non-word chars (except /) with single dash
 * - Remove leading/trailing dashes from path segments
 * - Preserve folder structure (a/b/c/d.md stays structured)
 * - Preserve file extension
 */
export function handelize(path: string): string {
  if (!path || path.trim() === '') {
    throw new Error('handelize: path cannot be empty');
  }

  // Allow route-style "$" filenames while still rejecting paths with no usable content.
  const segments = path.split('/').filter(Boolean);
  const lastSegment = segments[segments.length - 1] || '';
  const filenameWithoutExt = lastSegment.replace(/\.[^.]+$/, '');
  const hasValidContent = /[\p{L}\p{N}$]/u.test(filenameWithoutExt);
  if (!hasValidContent) {
    throw new Error(`handelize: path "${path}" has no valid filename content`);
  }

  const result = path
    .replace(/___/g, '/')       // Triple underscore becomes folder separator
    .toLowerCase()
    .split('/')
    .map((segment, idx, arr) => {
      const isLastSegment = idx === arr.length - 1;

      if (isLastSegment) {
        // For the filename (last segment), preserve the extension
        const extMatch = segment.match(/(\.[a-z0-9]+)$/i);
        const ext = extMatch ? extMatch[1] : '';
        const nameWithoutExt = ext ? segment.slice(0, -ext.length) : segment;

        const cleanedName = nameWithoutExt
          .replace(/[^\p{L}\p{N}$]+/gu, '-')  // Keep route marker "$", dash-separate other chars
          .replace(/^-+|-+$/g, ''); // Remove leading/trailing dashes

        return cleanedName + ext;
      } else {
        // For directories, just clean normally
        return segment
          .replace(/[^\p{L}\p{N}$]+/gu, '-')
          .replace(/^-+|-+$/g, '');
      }
    })
    .filter(Boolean)
    .join('/');

  if (!result) {
    throw new Error(`handelize: path "${path}" resulted in empty string after processing`);
  }

  return result;
}

/**
 * Search result extends DocumentResult with score and source info
 */
export type SearchResult = DocumentResult & {
  score: number;              // Relevance score (0-1)
  source: "fts" | "vec";      // Search source (full-text or vector)
  chunkPos?: number;          // Character position of matching chunk (for vector search)
};

/**
 * Ranked result for RRF fusion (simplified, used internally)
 */
export type RankedResult = {
  file: string;
  displayPath: string;
  title: string;
  body: string;
  score: number;
};

/**
 * Error result when document is not found
 */
export type DocumentNotFound = {
  error: "not_found";
  query: string;
  similarFiles: string[];
};

/**
 * Result from multi-get operations
 */
export type MultiGetResult = {
  doc: DocumentResult;
  skipped: false;
} | {
  doc: Pick<DocumentResult, "filepath" | "displayPath">;
  skipped: true;
  skipReason: string;
};

export type CollectionInfo = {
  name: string;
  path: string;
  pattern: string;
  documents: number;
  lastUpdated: string;
};

export type IndexStatus = {
  totalDocuments: number;
  needsEmbedding: number;
  hasVectorIndex: boolean;
  collections: CollectionInfo[];
};

// =============================================================================
// Index health
// =============================================================================

export function getHashesNeedingEmbedding(db: Database): number {
  const result = db.prepare(`
    SELECT COUNT(DISTINCT d.hash) as count
    FROM documents d
    LEFT JOIN content_vectors v ON d.hash = v.hash AND v.seq = 0
    WHERE d.active = 1 AND v.hash IS NULL
  `).get() as { count: number };
  return result.count;
}

export type IndexHealthInfo = {
  needsEmbedding: number;
  totalDocs: number;
  daysStale: number | null;
};

export function getIndexHealth(db: Database): IndexHealthInfo {
  const needsEmbedding = getHashesNeedingEmbedding(db);
  const totalDocs = (db.prepare(`SELECT COUNT(*) as count FROM documents WHERE active = 1`).get() as { count: number }).count;

  const mostRecent = db.prepare(`SELECT MAX(modified_at) as latest FROM documents WHERE active = 1`).get() as { latest: string | null };
  let daysStale: number | null = null;
  if (mostRecent?.latest) {
    const lastUpdate = new Date(mostRecent.latest);
    daysStale = Math.floor((Date.now() - lastUpdate.getTime()) / (24 * 60 * 60 * 1000));
  }

  return { needsEmbedding, totalDocs, daysStale };
}

// =============================================================================
// Caching
// =============================================================================

export function getCacheKey(url: string, body: object): string {
  const hash = createHash("sha256");
  hash.update(url);
  hash.update(JSON.stringify(body));
  return hash.digest("hex");
}

export function getCachedResult(db: Database, cacheKey: string): string | null {
  const row = db.prepare(`SELECT result FROM llm_cache WHERE hash = ?`).get(cacheKey) as { result: string } | null;
  return row?.result || null;
}

export function setCachedResult(db: Database, cacheKey: string, result: string): void {
  const now = new Date().toISOString();
  db.prepare(`INSERT OR REPLACE INTO llm_cache (hash, result, created_at) VALUES (?, ?, ?)`).run(cacheKey, result, now);
  if (Math.random() < 0.01) {
    db.exec(`DELETE FROM llm_cache WHERE hash NOT IN (SELECT hash FROM llm_cache ORDER BY created_at DESC LIMIT 1000)`);
  }
}

export function clearCache(db: Database): void {
  db.exec(`DELETE FROM llm_cache`);
}

// =============================================================================
// Cleanup and maintenance operations
// =============================================================================

/**
 * Delete cached LLM API responses.
 * Returns the number of cached responses deleted.
 */
export function deleteLLMCache(db: Database): number {
  const result = db.prepare(`DELETE FROM llm_cache`).run();
  return result.changes;
}

/**
 * Remove inactive document records (active = 0).
 * Returns the number of inactive documents deleted.
 */
export function deleteInactiveDocuments(db: Database): number {
  const result = db.prepare(`DELETE FROM documents WHERE active = 0`).run();
  return result.changes;
}

/**
 * Remove orphaned content hashes that are not referenced by any active document.
 * Returns the number of orphaned content hashes deleted.
 */
export function cleanupOrphanedContent(db: Database): number {
  const result = db.prepare(`
    DELETE FROM content
    WHERE hash NOT IN (SELECT DISTINCT hash FROM documents WHERE active = 1)
  `).run();
  return result.changes;
}

/**
 * Remove orphaned vector embeddings that are not referenced by any active document.
 * Returns the number of orphaned embedding chunks deleted.
 */
export function cleanupOrphanedVectors(db: Database): number {
  // Check if vectors_vec table exists
  const tableExists = db.prepare(`
    SELECT name FROM sqlite_master WHERE type='table' AND name='vectors_vec'
  `).get();

  if (!tableExists) {
    return 0;
  }

  // Count orphaned vectors first
  const countResult = db.prepare(`
    SELECT COUNT(*) as c FROM content_vectors cv
    WHERE NOT EXISTS (
      SELECT 1 FROM documents d WHERE d.hash = cv.hash AND d.active = 1
    )
  `).get() as { c: number };

  if (countResult.c === 0) {
    return 0;
  }

  // Delete from vectors_vec first
  db.exec(`
    DELETE FROM vectors_vec WHERE hash_seq IN (
      SELECT cv.hash || '_' || cv.seq FROM content_vectors cv
      WHERE NOT EXISTS (
        SELECT 1 FROM documents d WHERE d.hash = cv.hash AND d.active = 1
      )
    )
  `);

  // Delete from content_vectors
  db.exec(`
    DELETE FROM content_vectors WHERE hash NOT IN (
      SELECT hash FROM documents WHERE active = 1
    )
  `);

  return countResult.c;
}

/**
 * Run VACUUM to reclaim unused space in the database.
 * This operation rebuilds the database file to eliminate fragmentation.
 */
export function vacuumDatabase(db: Database): void {
  db.exec(`VACUUM`);
}

// =============================================================================
// Document helpers
// =============================================================================

export async function hashContent(content: string): Promise<string> {
  const hash = createHash("sha256");
  hash.update(content);
  return hash.digest("hex");
}

const titleExtractors: Record<string, (content: string) => string | null> = {
  '.md': (content) => {
    const match = content.match(/^##?\s+(.+)$/m);
    if (match) {
      const title = (match[1] ?? "").trim();
      if (title === "📝 Notes" || title === "Notes") {
        const nextMatch = content.match(/^##\s+(.+)$/m);
        if (nextMatch?.[1]) return nextMatch[1].trim();
      }
      return title;
    }
    return null;
  },
  '.org': (content) => {
    const titleProp = content.match(/^#\+TITLE:\s*(.+)$/im);
    if (titleProp?.[1]) return titleProp[1].trim();
    const heading = content.match(/^\*+\s+(.+)$/m);
    if (heading?.[1]) return heading[1].trim();
    return null;
  },
};

export function extractTitle(content: string, filename: string): string {
  const ext = filename.slice(filename.lastIndexOf('.')).toLowerCase();
  const extractor = titleExtractors[ext];
  if (extractor) {
    const title = extractor(content);
    if (title) return title;
  }
  return filename.replace(/\.[^.]+$/, "").split("/").pop() || filename;
}

// =============================================================================
// Document indexing operations
// =============================================================================

/**
 * Insert content into the content table (content-addressable storage).
 * Uses INSERT OR IGNORE so duplicate hashes are skipped.
 */
export function insertContent(db: Database, hash: string, content: string, createdAt: string): void {
  db.prepare(`INSERT OR IGNORE INTO content (hash, doc, created_at) VALUES (?, ?, ?)`)
    .run(hash, content, createdAt);
}

/**
 * Insert a new document into the documents table.
 */
export function insertDocument(
  db: Database,
  collectionName: string,
  path: string,
  title: string,
  hash: string,
  createdAt: string,
  modifiedAt: string
): void {
  db.prepare(`
    INSERT INTO documents (collection, path, title, hash, created_at, modified_at, active)
    VALUES (?, ?, ?, ?, ?, ?, 1)
    ON CONFLICT(collection, path) DO UPDATE SET
      title = excluded.title,
      hash = excluded.hash,
      modified_at = excluded.modified_at,
      active = 1
  `).run(collectionName, path, title, hash, createdAt, modifiedAt);
}

/**
 * Find an active document by collection name and path.
 */
export function findActiveDocument(
  db: Database,
  collectionName: string,
  path: string
): { id: number; hash: string; title: string } | null {
  const row = db.prepare(`
    SELECT id, hash, title FROM documents
    WHERE collection = ? AND path = ? AND active = 1
  `).get(collectionName, path) as { id: number; hash: string; title: string } | undefined;
  return row ?? null;
}

/**
 * Update the title and modified_at timestamp for a document.
 */
export function updateDocumentTitle(
  db: Database,
  documentId: number,
  title: string,
  modifiedAt: string
): void {
  db.prepare(`UPDATE documents SET title = ?, modified_at = ? WHERE id = ?`)
    .run(title, modifiedAt, documentId);
}

/**
 * Update an existing document's hash, title, and modified_at timestamp.
 * Used when content changes but the file path stays the same.
 */
export function updateDocument(
  db: Database,
  documentId: number,
  title: string,
  hash: string,
  modifiedAt: string
): void {
  db.prepare(`UPDATE documents SET title = ?, hash = ?, modified_at = ? WHERE id = ?`)
    .run(title, hash, modifiedAt, documentId);
}

/**
 * Deactivate a document (mark as inactive but don't delete).
 */
export function deactivateDocument(db: Database, collectionName: string, path: string): void {
  db.prepare(`UPDATE documents SET active = 0 WHERE collection = ? AND path = ? AND active = 1`)
    .run(collectionName, path);
}

/**
 * Get all active document paths for a collection.
 */
export function getActiveDocumentPaths(db: Database, collectionName: string): string[] {
  const rows = db.prepare(`
    SELECT path FROM documents WHERE collection = ? AND active = 1
  `).all(collectionName) as { path: string }[];
  return rows.map(r => r.path);
}

export { formatQueryForEmbedding, formatDocForEmbedding };

export function chunkDocument(
  content: string,
  maxChars: number = CHUNK_SIZE_CHARS,
  overlapChars: number = CHUNK_OVERLAP_CHARS,
  windowChars: number = CHUNK_WINDOW_CHARS
): { text: string; pos: number }[] {
  if (content.length <= maxChars) {
    return [{ text: content, pos: 0 }];
  }

  // Pre-scan all break points and code fences once
  const breakPoints = scanBreakPoints(content);
  const codeFences = findCodeFences(content);

  const chunks: { text: string; pos: number }[] = [];
  let charPos = 0;

  while (charPos < content.length) {
    // Calculate target end position for this chunk
    const targetEndPos = Math.min(charPos + maxChars, content.length);

    let endPos = targetEndPos;

    // If not at the end, find the best break point
    if (endPos < content.length) {
      // Find best cutoff using scored algorithm
      const bestCutoff = findBestCutoff(
        breakPoints,
        targetEndPos,
        windowChars,
        0.7,
        codeFences
      );

      // Only use the cutoff if it's within our current chunk
      if (bestCutoff > charPos && bestCutoff <= targetEndPos) {
        endPos = bestCutoff;
      }
    }

    // Ensure we make progress
    if (endPos <= charPos) {
      endPos = Math.min(charPos + maxChars, content.length);
    }

    chunks.push({ text: content.slice(charPos, endPos), pos: charPos });

    // Move forward, but overlap with previous chunk
    // For last chunk, don't overlap (just go to the end)
    if (endPos >= content.length) {
      break;
    }
    charPos = endPos - overlapChars;
    const lastChunkPos = chunks.at(-1)!.pos;
    if (charPos <= lastChunkPos) {
      // Prevent infinite loop - move forward at least a bit
      charPos = endPos;
    }
  }

  return chunks;
}

/**
 * Chunk a document by actual token count using the LLM tokenizer.
 * More accurate than character-based chunking but requires async.
 */
export async function chunkDocumentByTokens(
  content: string,
  maxTokens: number = CHUNK_SIZE_TOKENS,
  overlapTokens: number = CHUNK_OVERLAP_TOKENS,
  windowTokens: number = CHUNK_WINDOW_TOKENS
): Promise<{ text: string; pos: number; tokens: number }[]> {
  const llm = getDefaultLlamaCpp();

  // Use moderate chars/token estimate (prose ~4, code ~2, mixed ~3)
  // If chunks exceed limit, they'll be re-split with actual ratio
  const avgCharsPerToken = 3;
  const maxChars = maxTokens * avgCharsPerToken;
  const overlapChars = overlapTokens * avgCharsPerToken;
  const windowChars = windowTokens * avgCharsPerToken;

  // Chunk in character space with conservative estimate
  let charChunks = chunkDocument(content, maxChars, overlapChars, windowChars);

  // Tokenize and split any chunks that still exceed limit
  const results: { text: string; pos: number; tokens: number }[] = [];

  for (const chunk of charChunks) {
    const tokens = await llm.tokenize(chunk.text);

    if (tokens.length <= maxTokens) {
      results.push({ text: chunk.text, pos: chunk.pos, tokens: tokens.length });
    } else {
      // Chunk is still too large - split it further
      // Use actual token count to estimate better char limit
      const actualCharsPerToken = chunk.text.length / tokens.length;
      const safeMaxChars = Math.floor(maxTokens * actualCharsPerToken * 0.95); // 5% safety margin

      const subChunks = chunkDocument(chunk.text, safeMaxChars, Math.floor(overlapChars * actualCharsPerToken / 2), Math.floor(windowChars * actualCharsPerToken / 2));

      for (const subChunk of subChunks) {
        const subTokens = await llm.tokenize(subChunk.text);
        results.push({
          text: subChunk.text,
          pos: chunk.pos + subChunk.pos,
          tokens: subTokens.length,
        });
      }
    }
  }

  return results;
}

// =============================================================================
// Fuzzy matching
// =============================================================================

function levenshtein(a: string, b: string): number {
  const m = a.length, n = b.length;
  if (m === 0) return n;
  if (n === 0) return m;
  const dp: number[][] = Array.from({ length: m + 1 }, () => Array(n + 1).fill(0));
  for (let i = 0; i <= m; i++) dp[i]![0] = i;
  for (let j = 0; j <= n; j++) dp[0]![j] = j;
  for (let i = 1; i <= m; i++) {
    for (let j = 1; j <= n; j++) {
      const cost = a[i - 1] === b[j - 1] ? 0 : 1;
      dp[i]![j] = Math.min(
        dp[i - 1]![j]! + 1,
        dp[i]![j - 1]! + 1,
        dp[i - 1]![j - 1]! + cost
      );
    }
  }
  return dp[m]![n]!;
}

/**
 * Normalize a docid input by stripping surrounding quotes and leading #.
 * Handles: "#abc123", 'abc123', "abc123", #abc123, abc123
 * Returns the bare hex string.
 */
export function normalizeDocid(docid: string): string {
  let normalized = docid.trim();

  // Strip surrounding quotes (single or double)
  if ((normalized.startsWith('"') && normalized.endsWith('"')) ||
      (normalized.startsWith("'") && normalized.endsWith("'"))) {
    normalized = normalized.slice(1, -1);
  }

  // Strip leading # if present
  if (normalized.startsWith('#')) {
    normalized = normalized.slice(1);
  }

  return normalized;
}

/**
 * Check if a string looks like a docid reference.
 * Accepts: #abc123, abc123, "#abc123", "abc123", '#abc123', 'abc123'
 * Returns true if the normalized form is a valid hex string of 6+ chars.
 */
export function isDocid(input: string): boolean {
  const normalized = normalizeDocid(input);
  // Must be at least 6 hex characters
  return normalized.length >= 6 && /^[a-f0-9]+$/i.test(normalized);
}

/**
 * Find a document by its short docid (first 6 characters of hash).
 * Returns the document's virtual path if found, null otherwise.
 * If multiple documents match the same short hash (collision), returns the first one.
 *
 * Accepts lenient input: #abc123, abc123, "#abc123", "abc123"
 */
export function findDocumentByDocid(db: Database, docid: string): { filepath: string; hash: string } | null {
  const shortHash = normalizeDocid(docid);

  if (shortHash.length < 1) return null;

  // Look up documents where hash starts with the short hash
  const doc = db.prepare(`
    SELECT 'qmd://' || d.collection || '/' || d.path as filepath, d.hash
    FROM documents d
    WHERE d.hash LIKE ? AND d.active = 1
    LIMIT 1
  `).get(`${shortHash}%`) as { filepath: string; hash: string } | null;

  return doc;
}

export function findSimilarFiles(db: Database, query: string, maxDistance: number = 3, limit: number = 5): string[] {
  const allFiles = db.prepare(`
    SELECT d.path
    FROM documents d
    WHERE d.active = 1
  `).all() as { path: string }[];
  const queryLower = query.toLowerCase();
  const scored = allFiles
    .map(f => ({ path: f.path, dist: levenshtein(f.path.toLowerCase(), queryLower) }))
    .filter(f => f.dist <= maxDistance)
    .sort((a, b) => a.dist - b.dist)
    .slice(0, limit);
  return scored.map(f => f.path);
}

export function matchFilesByGlob(db: Database, pattern: string): { filepath: string; displayPath: string; bodyLength: number }[] {
  const allFiles = db.prepare(`
    SELECT
      'qmd://' || d.collection || '/' || d.path as virtual_path,
      LENGTH(content.doc) as body_length,
      d.path,
      d.collection
    FROM documents d
    JOIN content ON content.hash = d.hash
    WHERE d.active = 1
  `).all() as { virtual_path: string; body_length: number; path: string; collection: string }[];

  const isMatch = picomatch(pattern);
  return allFiles
    .filter(f => isMatch(f.virtual_path) || isMatch(f.path))
    .map(f => ({
      filepath: f.virtual_path,  // Virtual path for precise lookup
      displayPath: f.path,        // Relative path for display
      bodyLength: f.body_length
    }));
}

// =============================================================================
// Context
// =============================================================================

/**
 * Get context for a file path using hierarchical inheritance.
 * Contexts are collection-scoped and inherit from parent directories.
 * For example, context at "/talks" applies to "/talks/2024/keynote.md".
 *
 * @param db Database instance (unused - kept for compatibility)
 * @param collectionName Collection name
 * @param path Relative path within the collection
 * @returns Context string or null if no context is defined
 */
export function getContextForPath(db: Database, collectionName: string, path: string): string | null {
  const config = collectionsLoadConfig();
  const coll = getCollection(collectionName);

  if (!coll) return null;

  // Collect ALL matching contexts (global + all path prefixes)
  const contexts: string[] = [];

  // Add global context if present
  if (config.global_context) {
    contexts.push(config.global_context);
  }

  // Add all matching path contexts (from most general to most specific)
  if (coll.context) {
    const normalizedPath = path.startsWith("/") ? path : `/${path}`;

    // Collect all matching prefixes
    const matchingContexts: { prefix: string; context: string }[] = [];
    for (const [prefix, context] of Object.entries(coll.context)) {
      const normalizedPrefix = prefix.startsWith("/") ? prefix : `/${prefix}`;
      if (normalizedPath.startsWith(normalizedPrefix)) {
        matchingContexts.push({ prefix: normalizedPrefix, context });
      }
    }

    // Sort by prefix length (shortest/most general first)
    matchingContexts.sort((a, b) => a.prefix.length - b.prefix.length);

    // Add all matching contexts
    for (const match of matchingContexts) {
      contexts.push(match.context);
    }
  }

  // Join all contexts with double newline
  return contexts.length > 0 ? contexts.join('\n\n') : null;
}

/**
 * Get context for a file path (virtual or filesystem).
 * Resolves the collection and relative path using the YAML collections config.
 */
export function getContextForFile(db: Database, filepath: string): string | null {
  // Handle undefined or null filepath
  if (!filepath) return null;

  // Get all collections from YAML config
  const collections = collectionsListCollections();
  const config = collectionsLoadConfig();

  // Parse virtual path format: qmd://collection/path
  let collectionName: string | null = null;
  let relativePath: string | null = null;

  const parsedVirtual = filepath.startsWith('qmd://') ? parseVirtualPath(filepath) : null;
  if (parsedVirtual) {
    collectionName = parsedVirtual.collectionName;
    relativePath = parsedVirtual.path;
  } else {
    // Filesystem path: find which collection this absolute path belongs to
    for (const coll of collections) {
      // Skip collections with missing paths
      if (!coll || !coll.path) continue;

      if (filepath.startsWith(coll.path + '/') || filepath === coll.path) {
        collectionName = coll.name;
        // Extract relative path
        relativePath = filepath.startsWith(coll.path + '/')
          ? filepath.slice(coll.path.length + 1)
          : '';
        break;
      }
    }

    if (!collectionName || relativePath === null) return null;
  }

  // Get the collection from config
  const coll = getCollection(collectionName);
  if (!coll) return null;

  // Verify this document exists in the database
  const doc = db.prepare(`
    SELECT d.path
    FROM documents d
    WHERE d.collection = ? AND d.path = ? AND d.active = 1
    LIMIT 1
  `).get(collectionName, relativePath) as { path: string } | null;

  if (!doc) return null;

  // Collect ALL matching contexts (global + all path prefixes)
  const contexts: string[] = [];

  // Add global context if present
  if (config.global_context) {
    contexts.push(config.global_context);
  }

  // Add all matching path contexts (from most general to most specific)
  if (coll.context) {
    const normalizedPath = relativePath.startsWith("/") ? relativePath : `/${relativePath}`;

    // Collect all matching prefixes
    const matchingContexts: { prefix: string; context: string }[] = [];
    for (const [prefix, context] of Object.entries(coll.context)) {
      const normalizedPrefix = prefix.startsWith("/") ? prefix : `/${prefix}`;
      if (normalizedPath.startsWith(normalizedPrefix)) {
        matchingContexts.push({ prefix: normalizedPrefix, context });
      }
    }

    // Sort by prefix length (shortest/most general first)
    matchingContexts.sort((a, b) => a.prefix.length - b.prefix.length);

    // Add all matching contexts
    for (const match of matchingContexts) {
      contexts.push(match.context);
    }
  }

  // Join all contexts with double newline
  return contexts.length > 0 ? contexts.join('\n\n') : null;
}

/**
 * Get collection by name from YAML config.
 * Returns collection metadata from ~/.config/qmd/index.yml
 */
export function getCollectionByName(db: Database, name: string): { name: string; pwd: string; glob_pattern: string } | null {
  const collection = getCollection(name);
  if (!collection) return null;

  return {
    name: collection.name,
    pwd: collection.path,
    glob_pattern: collection.pattern,
  };
}

/**
 * List all collections with document counts from database.
 * Merges YAML config with database statistics.
 */
export function listCollections(db: Database): { name: string; pwd: string; glob_pattern: string; doc_count: number; active_count: number; last_modified: string | null }[] {
  const collections = collectionsListCollections();

  // Get document counts from database for each collection
  const result = collections.map(coll => {
    const stats = db.prepare(`
      SELECT
        COUNT(d.id) as doc_count,
        SUM(CASE WHEN d.active = 1 THEN 1 ELSE 0 END) as active_count,
        MAX(d.modified_at) as last_modified
      FROM documents d
      WHERE d.collection = ?
    `).get(coll.name) as { doc_count: number; active_count: number; last_modified: string | null } | null;

    return {
      name: coll.name,
      pwd: coll.path,
      glob_pattern: coll.pattern,
      doc_count: stats?.doc_count || 0,
      active_count: stats?.active_count || 0,
      last_modified: stats?.last_modified || null,
    };
  });

  return result;
}

/**
 * Remove a collection and clean up its documents.
 * Uses collections.ts to remove from YAML config and cleans up database.
 */
export function removeCollection(db: Database, collectionName: string): { deletedDocs: number; cleanedHashes: number } {
  // Delete documents from database
  const docResult = db.prepare(`DELETE FROM documents WHERE collection = ?`).run(collectionName);

  // Clean up orphaned content hashes
  const cleanupResult = db.prepare(`
    DELETE FROM content
    WHERE hash NOT IN (SELECT DISTINCT hash FROM documents WHERE active = 1)
  `).run();

  // Remove from YAML config (returns true if found and removed)
  collectionsRemoveCollection(collectionName);

  return {
    deletedDocs: docResult.changes,
    cleanedHashes: cleanupResult.changes
  };
}

/**
 * Rename a collection.
 * Updates both YAML config and database documents table.
 */
export function renameCollection(db: Database, oldName: string, newName: string): void {
  // Update all documents with the new collection name in database
  db.prepare(`UPDATE documents SET collection = ? WHERE collection = ?`)
    .run(newName, oldName);

  // Rename in YAML config
  collectionsRenameCollection(oldName, newName);
}

// =============================================================================
// Context Management Operations
// =============================================================================

/**
 * Insert or update a context for a specific collection and path prefix.
 */
export function insertContext(db: Database, collectionId: number, pathPrefix: string, context: string): void {
  // Get collection name from ID
  const coll = db.prepare(`SELECT name FROM collections WHERE id = ?`).get(collectionId) as { name: string } | null;
  if (!coll) {
    throw new Error(`Collection with id ${collectionId} not found`);
  }

  // Use collections.ts to add context
  collectionsAddContext(coll.name, pathPrefix, context);
}

/**
 * Delete a context for a specific collection and path prefix.
 * Returns the number of contexts deleted.
 */
export function deleteContext(db: Database, collectionName: string, pathPrefix: string): number {
  // Use collections.ts to remove context
  const success = collectionsRemoveContext(collectionName, pathPrefix);
  return success ? 1 : 0;
}

/**
 * Delete all global contexts (contexts with empty path_prefix).
 * Returns the number of contexts deleted.
 */
export function deleteGlobalContexts(db: Database): number {
  let deletedCount = 0;

  // Remove global context
  setGlobalContext(undefined);
  deletedCount++;

  // Remove root context (empty string) from all collections
  const collections = collectionsListCollections();
  for (const coll of collections) {
    const success = collectionsRemoveContext(coll.name, '');
    if (success) {
      deletedCount++;
    }
  }

  return deletedCount;
}

/**
 * List all contexts, grouped by collection.
 * Returns contexts ordered by collection name, then by path prefix length (longest first).
 */
export function listPathContexts(db: Database): { collection_name: string; path_prefix: string; context: string }[] {
  const allContexts = collectionsListAllContexts();

  // Convert to expected format and sort
  return allContexts.map(ctx => ({
    collection_name: ctx.collection,
    path_prefix: ctx.path,
    context: ctx.context,
  })).sort((a, b) => {
    // Sort by collection name first
    if (a.collection_name !== b.collection_name) {
      return a.collection_name.localeCompare(b.collection_name);
    }
    // Then by path prefix length (longest first)
    if (a.path_prefix.length !== b.path_prefix.length) {
      return b.path_prefix.length - a.path_prefix.length;
    }
    // Then alphabetically
    return a.path_prefix.localeCompare(b.path_prefix);
  });
}

/**
 * Get all collections (name only - from YAML config).
 */
export function getAllCollections(db: Database): { name: string }[] {
  const collections = collectionsListCollections();
  return collections.map(c => ({ name: c.name }));
}

/**
 * Check which collections don't have any context defined.
 * Returns collections that have no context entries at all (not even root context).
 */
export function getCollectionsWithoutContext(db: Database): { name: string; pwd: string; doc_count: number }[] {
  // Get all collections from YAML config
  const yamlCollections = collectionsListCollections();

  // Filter to those without context
  const collectionsWithoutContext: { name: string; pwd: string; doc_count: number }[] = [];

  for (const coll of yamlCollections) {
    // Check if collection has any context
    if (!coll.context || Object.keys(coll.context).length === 0) {
      // Get doc count from database
      const stats = db.prepare(`
        SELECT COUNT(d.id) as doc_count
        FROM documents d
        WHERE d.collection = ? AND d.active = 1
      `).get(coll.name) as { doc_count: number } | null;

      collectionsWithoutContext.push({
        name: coll.name,
        pwd: coll.path,
        doc_count: stats?.doc_count || 0,
      });
    }
  }

  return collectionsWithoutContext.sort((a, b) => a.name.localeCompare(b.name));
}

/**
 * Get top-level directories in a collection that don't have context.
 * Useful for suggesting where context might be needed.
 */
export function getTopLevelPathsWithoutContext(db: Database, collectionName: string): string[] {
  // Get all paths in the collection from database
  const paths = db.prepare(`
    SELECT DISTINCT path FROM documents
    WHERE collection = ? AND active = 1
  `).all(collectionName) as { path: string }[];

  // Get existing contexts for this collection from YAML
  const yamlColl = getCollection(collectionName);
  if (!yamlColl) return [];

  const contextPrefixes = new Set<string>();
  if (yamlColl.context) {
    for (const prefix of Object.keys(yamlColl.context)) {
      contextPrefixes.add(prefix);
    }
  }

  // Extract top-level directories (first path component)
  const topLevelDirs = new Set<string>();
  for (const { path } of paths) {
    const parts = path.split('/').filter(Boolean);
    if (parts.length > 1) {
      const dir = parts[0];
      if (dir) topLevelDirs.add(dir);
    }
  }

  // Filter out directories that already have context (exact or parent)
  const missing: string[] = [];
  for (const dir of topLevelDirs) {
    let hasContext = false;

    // Check if this dir or any parent has context
    for (const prefix of contextPrefixes) {
      if (prefix === '' || prefix === dir || dir.startsWith(prefix + '/')) {
        hasContext = true;
        break;
      }
    }

    if (!hasContext) {
      missing.push(dir);
    }
  }

  return missing.sort();
}

// =============================================================================
// FTS Search
// =============================================================================

function sanitizeFTS5Term(term: string): string {
  return term.replace(/[^\p{L}\p{N}']/gu, '').toLowerCase();
}

/**
 * Parse lex query syntax into FTS5 query.
 *
 * Supports:
 * - Quoted phrases: "exact phrase" → "exact phrase" (exact match)
 * - Negation: -term or -"phrase" → uses FTS5 NOT operator
 * - OR: term1 OR term2 (case-insensitive)
 * - Plain terms: term → "term"* (prefix match)
 *
 * FTS5 NOT is a binary operator: `term1 NOT term2` means "match term1 but not term2".
 * So `-term` only works when there are also positive terms.
 *
 * Examples:
 *   performance -sports     → "performance"* NOT "sports"*
 *   "machine learning"      → "machine learning"
 *   auth OR authentication  → ("auth"* OR "authentication"*)
 */
function buildFTS5Query(query: string): string | null {
  const positive: string[] = [];
  const negative: string[] = [];
  const orGroups: string[][] = [[]]; // Track OR groupings
  let currentOrGroup = 0;

  let i = 0;
  const s = query.trim();

  while (i < s.length) {
    // Skip whitespace
    while (i < s.length && /\s/.test(s[i]!)) i++;
    if (i >= s.length) break;

    // Check for negation prefix
    const negated = s[i] === '-';
    if (negated) i++;

    // Check for quoted phrase
    if (s[i] === '"') {
      const start = i + 1;
      i++;
      while (i < s.length && s[i] !== '"') i++;
      const phrase = s.slice(start, i).trim();
      i++; // skip closing quote
      if (phrase.length > 0) {
        const sanitized = phrase.split(/\s+/).map(t => sanitizeFTS5Term(t)).filter(t => t).join(' ');
        if (sanitized) {
          const ftsPhrase = `"${sanitized}"`;  // Exact phrase, no prefix match
          if (negated) {
            negative.push(ftsPhrase);
          } else {
            positive.push(ftsPhrase);
            orGroups[currentOrGroup]!.push(ftsPhrase);
          }
        }
      }
    } else {
      // Plain term (until whitespace or quote)
      const start = i;
      while (i < s.length && !/[\s"]/.test(s[i]!)) i++;
      const term = s.slice(start, i);

      // Check for OR operator
      if (term.toUpperCase() === 'OR') {
        // Start new OR group
        currentOrGroup++;
        orGroups.push([]);
      } else if (term.toUpperCase() === 'AND' || term.toUpperCase() === 'NOT') {
        // AND is implicit, NOT should use - prefix
        continue;
      } else {
        const sanitized = sanitizeFTS5Term(term);
        if (sanitized) {
          const ftsTerm = `"${sanitized}"*`;  // Prefix match
          if (negated) {
            negative.push(ftsTerm);
          } else {
            positive.push(ftsTerm);
            orGroups[currentOrGroup]!.push(ftsTerm);
          }
        }
      }
    }
  }

  if (positive.length === 0 && negative.length === 0) return null;

  // If only negative terms, we can't search (FTS5 NOT is binary)
  if (positive.length === 0) {
    // Fall back to searching without negation
    return null;
  }

  // Build the positive part with OR groups
  let result: string;
  if (orGroups.length > 1 && orGroups.some(g => g.length > 0)) {
    // Has OR groups - build (a OR b) AND c structure
    const orParts = orGroups.filter(g => g.length > 0).map(g =>
      g.length === 1 ? g[0]! : `(${g.join(' OR ')})`
    );
    result = orParts.join(' AND ');
  } else {
    // Simple AND of all positive terms
    result = positive.join(' AND ');
  }

  // Add NOT clause for negative terms (FTS5: positive NOT negative)
  if (negative.length > 0) {
    // FTS5 NOT only works with single term on right side, chain them
    for (const neg of negative) {
      result = `${result} NOT ${neg}`;
    }
  }

  return result;
}

/**
 * Validate that a vec/hyde query doesn't use lex-only syntax.
 * Returns error message if invalid, null if valid.
 */
export function validateSemanticQuery(query: string): string | null {
  // Check for negation syntax
  if (/-\w/.test(query) || /-"/.test(query)) {
    return 'Negation (-term) is not supported in vec/hyde queries. Use lex for exclusions.';
  }
  // Check for quoted exact phrases (semantic search doesn't do exact matching)
  if (/"[^"]+"\s*$/.test(query.trim()) || /^"[^"]+"/.test(query.trim())) {
    // Single quoted phrase is the whole query - that's fine for hyde
    // But warn if it looks like they expect exact matching
  }
  // Check for OR operator (semantic search doesn't support boolean logic)
  if (/\bOR\b/i.test(query)) {
    return 'OR operator is not supported in vec/hyde queries. Use multiple lex queries or rephrase.';
  }
  return null;
}

export function searchFTS(db: Database, query: string, limit: number = 20, collectionName?: string): SearchResult[] {
  const ftsQuery = buildFTS5Query(query);
  if (!ftsQuery) return [];

  let sql = `
    SELECT
      'qmd://' || d.collection || '/' || d.path as filepath,
      d.collection || '/' || d.path as display_path,
      d.title,
      content.doc as body,
      d.hash,
      bm25(documents_fts, 10.0, 1.0) as bm25_score
    FROM documents_fts f
    JOIN documents d ON d.id = f.rowid
    JOIN content ON content.hash = d.hash
    WHERE documents_fts MATCH ? AND d.active = 1
  `;
  const params: (string | number)[] = [ftsQuery];

  if (collectionName) {
    sql += ` AND d.collection = ?`;
    params.push(String(collectionName));
  }

  // bm25 lower is better; sort ascending.
  sql += ` ORDER BY bm25_score ASC LIMIT ?`;
  params.push(limit);

  const rows = db.prepare(sql).all(...params) as { filepath: string; display_path: string; title: string; body: string; hash: string; bm25_score: number }[];
  return rows.map(row => {
    const collectionName = row.filepath.split('//')[1]?.split('/')[0] || "";
    // Convert bm25 (negative, lower is better) into a stable [0..1) score where higher is better.
    // FTS5 BM25 scores are negative (e.g., -10 is strong, -2 is weak).
    // |x| / (1 + |x|) maps: strong(-10)→0.91, medium(-2)→0.67, weak(-0.5)→0.33, none(0)→0.
    // Monotonic and query-independent — no per-query normalization needed.
    const score = Math.abs(row.bm25_score) / (1 + Math.abs(row.bm25_score));
    return {
      filepath: row.filepath,
      displayPath: row.display_path,
      title: row.title,
      hash: row.hash,
      docid: getDocid(row.hash),
      collectionName,
      modifiedAt: "",  // Not available in FTS query
      bodyLength: row.body.length,
      body: row.body,
      context: getContextForFile(db, row.filepath),
      score,
      source: "fts" as const,
    };
  });
}

// =============================================================================
// Vector Search
// =============================================================================

export async function searchVec(db: Database, query: string, model: string, limit: number = 20, collectionName?: string, session?: ILLMSession, precomputedEmbedding?: number[]): Promise<SearchResult[]> {
  const tableExists = db.prepare(`SELECT name FROM sqlite_master WHERE type='table' AND name='vectors_vec'`).get();
  if (!tableExists) return [];

  const embedding = precomputedEmbedding ?? await getEmbedding(query, model, true, session);
  if (!embedding) return [];

  // IMPORTANT: We use a two-step query approach here because sqlite-vec virtual tables
  // hang indefinitely when combined with JOINs in the same query. Do NOT try to
  // "optimize" this by combining into a single query with JOINs - it will break.
  // See: https://github.com/tobi/qmd/pull/23

  // Step 1: Get vector matches from sqlite-vec (no JOINs allowed)
  const vecResults = db.prepare(`
    SELECT hash_seq, distance
    FROM vectors_vec
    WHERE embedding MATCH ? AND k = ?
  `).all(new Float32Array(embedding), limit * 3) as { hash_seq: string; distance: number }[];

  if (vecResults.length === 0) return [];

  // Step 2: Get chunk info and document data
  const hashSeqs = vecResults.map(r => r.hash_seq);
  const distanceMap = new Map(vecResults.map(r => [r.hash_seq, r.distance]));

  // Build query for document lookup
  const placeholders = hashSeqs.map(() => '?').join(',');
  let docSql = `
    SELECT
      cv.hash || '_' || cv.seq as hash_seq,
      cv.hash,
      cv.pos,
      'qmd://' || d.collection || '/' || d.path as filepath,
      d.collection || '/' || d.path as display_path,
      d.title,
      content.doc as body
    FROM content_vectors cv
    JOIN documents d ON d.hash = cv.hash AND d.active = 1
    JOIN content ON content.hash = d.hash
    WHERE cv.hash || '_' || cv.seq IN (${placeholders})
  `;
  const params: string[] = [...hashSeqs];

  if (collectionName) {
    docSql += ` AND d.collection = ?`;
    params.push(collectionName);
  }

  const docRows = db.prepare(docSql).all(...params) as {
    hash_seq: string; hash: string; pos: number; filepath: string;
    display_path: string; title: string; body: string;
  }[];

  // Combine with distances and dedupe by filepath
  const seen = new Map<string, { row: typeof docRows[0]; bestDist: number }>();
  for (const row of docRows) {
    const distance = distanceMap.get(row.hash_seq) ?? 1;
    const existing = seen.get(row.filepath);
    if (!existing || distance < existing.bestDist) {
      seen.set(row.filepath, { row, bestDist: distance });
    }
  }

  return Array.from(seen.values())
    .sort((a, b) => a.bestDist - b.bestDist)
    .slice(0, limit)
    .map(({ row, bestDist }) => {
      const collectionName = row.filepath.split('//')[1]?.split('/')[0] || "";
      return {
        filepath: row.filepath,
        displayPath: row.display_path,
        title: row.title,
        hash: row.hash,
        docid: getDocid(row.hash),
        collectionName,
        modifiedAt: "",  // Not available in vec query
        bodyLength: row.body.length,
        body: row.body,
        context: getContextForFile(db, row.filepath),
        score: 1 - bestDist,  // Cosine similarity = 1 - cosine distance
        source: "vec" as const,
        chunkPos: row.pos,
      };
    });
}

// =============================================================================
// Embeddings
// =============================================================================

async function getEmbedding(text: string, model: string, isQuery: boolean, session?: ILLMSession): Promise<number[] | null> {
  // Format text using the appropriate prompt template
  const formattedText = isQuery ? formatQueryForEmbedding(text) : formatDocForEmbedding(text);
  const result = session
    ? await session.embed(formattedText, { model, isQuery })
    : await getDefaultLlamaCpp().embed(formattedText, { model, isQuery });
  return result?.embedding || null;
}

/**
 * Get all unique content hashes that need embeddings (from active documents).
 * Returns hash, document body, and a sample path for display purposes.
 */
export function getHashesForEmbedding(db: Database): { hash: string; body: string; path: string }[] {
  return db.prepare(`
    SELECT d.hash, c.doc as body, MIN(d.path) as path
    FROM documents d
    JOIN content c ON d.hash = c.hash
    LEFT JOIN content_vectors v ON d.hash = v.hash AND v.seq = 0
    WHERE d.active = 1 AND v.hash IS NULL
    GROUP BY d.hash
  `).all() as { hash: string; body: string; path: string }[];
}

/**
 * Clear all embeddings from the database (force re-index).
 * Deletes all rows from content_vectors and drops the vectors_vec table.
 */
export function clearAllEmbeddings(db: Database): void {
  db.exec(`DELETE FROM content_vectors`);
  db.exec(`DROP TABLE IF EXISTS vectors_vec`);
}

/**
 * Insert a single embedding into both content_vectors and vectors_vec tables.
 * The hash_seq key is formatted as "hash_seq" for the vectors_vec table.
 */
export function insertEmbedding(
  db: Database,
  hash: string,
  seq: number,
  pos: number,
  embedding: Float32Array,
  model: string,
  embeddedAt: string
): void {
  const hashSeq = `${hash}_${seq}`;
  const insertVecStmt = db.prepare(`INSERT OR REPLACE INTO vectors_vec (hash_seq, embedding) VALUES (?, ?)`);
  const insertContentVectorStmt = db.prepare(`INSERT OR REPLACE INTO content_vectors (hash, seq, pos, model, embedded_at) VALUES (?, ?, ?, ?, ?)`);

  insertVecStmt.run(hashSeq, embedding);
  insertContentVectorStmt.run(hash, seq, pos, model, embeddedAt);
}

// =============================================================================
// Query expansion
// =============================================================================

export async function expandQuery(query: string, model: string = DEFAULT_QUERY_MODEL, db: Database): Promise<ExpandedQuery[]> {
  // Check cache first — stored as JSON preserving types
  const cacheKey = getCacheKey("expandQuery", { query, model });
  const cached = getCachedResult(db, cacheKey);
  if (cached) {
    try {
      return JSON.parse(cached) as ExpandedQuery[];
    } catch {
      // Old cache format (pre-typed, newline-separated text) — re-expand
    }
  }

  const llm = getDefaultLlamaCpp();
  // Note: LlamaCpp uses hardcoded model, model parameter is ignored
  const results = await llm.expandQuery(query);

  // Map Queryable[] → ExpandedQuery[] (same shape, decoupled from llm.ts internals).
  // Filter out entries that duplicate the original query text.
  const expanded: ExpandedQuery[] = results
    .filter(r => r.text !== query)
    .map(r => ({ type: r.type, text: r.text }));

  if (expanded.length > 0) {
    setCachedResult(db, cacheKey, JSON.stringify(expanded));
  }

  return expanded;
}

// =============================================================================
// Reranking
// =============================================================================

export async function rerank(query: string, documents: { file: string; text: string }[], model: string = DEFAULT_RERANK_MODEL, db: Database): Promise<{ file: string; score: number }[]> {
  const cachedResults: Map<string, number> = new Map();
  const uncachedDocs: RerankDocument[] = [];

  // Check cache for each document
  // Cache key includes chunk text — different queries can select different chunks
  // from the same file, and the reranker score depends on which chunk was sent.
  for (const doc of documents) {
    const cacheKey = getCacheKey("rerank", { query, file: doc.file, model, chunk: doc.text });
    const cached = getCachedResult(db, cacheKey);
    if (cached !== null) {
      cachedResults.set(doc.file, parseFloat(cached));
    } else {
      uncachedDocs.push({ file: doc.file, text: doc.text });
    }
  }

  // Rerank uncached documents using LlamaCpp
  if (uncachedDocs.length > 0) {
    const llm = getDefaultLlamaCpp();
    const rerankResult = await llm.rerank(query, uncachedDocs, { model });

    // Cache results — use original doc.text for cache key (result.file lacks chunk text)
    const textByFile = new Map(documents.map(d => [d.file, d.text]));
    for (const result of rerankResult.results) {
      const cacheKey = getCacheKey("rerank", { query, file: result.file, model, chunk: textByFile.get(result.file) || "" });
      setCachedResult(db, cacheKey, result.score.toString());
      cachedResults.set(result.file, result.score);
    }
  }

  // Return all results sorted by score
  return documents
    .map(doc => ({ file: doc.file, score: cachedResults.get(doc.file) || 0 }))
    .sort((a, b) => b.score - a.score);
}

// =============================================================================
// Reciprocal Rank Fusion
// =============================================================================

export function reciprocalRankFusion(
  resultLists: RankedResult[][],
  weights: number[] = [],
  k: number = 60
): RankedResult[] {
  const scores = new Map<string, { result: RankedResult; rrfScore: number; topRank: number }>();

  for (let listIdx = 0; listIdx < resultLists.length; listIdx++) {
    const list = resultLists[listIdx];
    if (!list) continue;
    const weight = weights[listIdx] ?? 1.0;

    for (let rank = 0; rank < list.length; rank++) {
      const result = list[rank];
      if (!result) continue;
      const rrfContribution = weight / (k + rank + 1);
      const existing = scores.get(result.file);

      if (existing) {
        existing.rrfScore += rrfContribution;
        existing.topRank = Math.min(existing.topRank, rank);
      } else {
        scores.set(result.file, {
          result,
          rrfScore: rrfContribution,
          topRank: rank,
        });
      }
    }
  }

  // Top-rank bonus
  for (const entry of scores.values()) {
    if (entry.topRank === 0) {
      entry.rrfScore += 0.05;
    } else if (entry.topRank <= 2) {
      entry.rrfScore += 0.02;
    }
  }

  return Array.from(scores.values())
    .sort((a, b) => b.rrfScore - a.rrfScore)
    .map(e => ({ ...e.result, score: e.rrfScore }));
}

// =============================================================================
// Document retrieval
// =============================================================================

type DbDocRow = {
  virtual_path: string;
  display_path: string;
  title: string;
  hash: string;
  collection: string;
  path: string;
  modified_at: string;
  body_length: number;
  body?: string;
};

/**
 * Find a document by filename/path, docid (#hash), or with fuzzy matching.
 * Returns document metadata without body by default.
 *
 * Supports:
 * - Virtual paths: qmd://collection/path/to/file.md
 * - Absolute paths: /path/to/file.md
 * - Relative paths: path/to/file.md
 * - Short docid: #abc123 (first 6 chars of hash)
 */
export function findDocument(db: Database, filename: string, options: { includeBody?: boolean } = {}): DocumentResult | DocumentNotFound {
  let filepath = filename;
  const colonMatch = filepath.match(/:(\d+)$/);
  if (colonMatch) {
    filepath = filepath.slice(0, -colonMatch[0].length);
  }

  // Check if this is a docid lookup (#abc123, abc123, "#abc123", "abc123", etc.)
  if (isDocid(filepath)) {
    const docidMatch = findDocumentByDocid(db, filepath);
    if (docidMatch) {
      filepath = docidMatch.filepath;
    } else {
      return { error: "not_found", query: filename, similarFiles: [] };
    }
  }

  if (filepath.startsWith('~/')) {
    filepath = homedir() + filepath.slice(1);
  }

  const bodyCol = options.includeBody ? `, content.doc as body` : ``;

  // Build computed columns
  // Note: absoluteFilepath is computed from YAML collections after query
  const selectCols = `
    'qmd://' || d.collection || '/' || d.path as virtual_path,
    d.collection || '/' || d.path as display_path,
    d.title,
    d.hash,
    d.collection,
    d.modified_at,
    LENGTH(content.doc) as body_length
    ${bodyCol}
  `;

  // Try to match by virtual path first
  let doc = db.prepare(`
    SELECT ${selectCols}
    FROM documents d
    JOIN content ON content.hash = d.hash
    WHERE 'qmd://' || d.collection || '/' || d.path = ? AND d.active = 1
  `).get(filepath) as DbDocRow | null;

  // Try fuzzy match by virtual path
  if (!doc) {
    doc = db.prepare(`
      SELECT ${selectCols}
      FROM documents d
      JOIN content ON content.hash = d.hash
      WHERE 'qmd://' || d.collection || '/' || d.path LIKE ? AND d.active = 1
      LIMIT 1
    `).get(`%${filepath}`) as DbDocRow | null;
  }

  // Try to match by absolute path (requires looking up collection paths from YAML)
  if (!doc && !filepath.startsWith('qmd://')) {
    const collections = collectionsListCollections();
    for (const coll of collections) {
      let relativePath: string | null = null;

      // If filepath is absolute and starts with collection path, extract relative part
      if (filepath.startsWith(coll.path + '/')) {
        relativePath = filepath.slice(coll.path.length + 1);
      }
      // Otherwise treat filepath as relative to collection
      else if (!filepath.startsWith('/')) {
        relativePath = filepath;
      }

      if (relativePath) {
        doc = db.prepare(`
          SELECT ${selectCols}
          FROM documents d
          JOIN content ON content.hash = d.hash
          WHERE d.collection = ? AND d.path = ? AND d.active = 1
        `).get(coll.name, relativePath) as DbDocRow | null;
        if (doc) break;
      }
    }
  }

  if (!doc) {
    const similar = findSimilarFiles(db, filepath, 5, 5);
    return { error: "not_found", query: filename, similarFiles: similar };
  }

  // Get context using virtual path
  const virtualPath = doc.virtual_path || `qmd://${doc.collection}/${doc.display_path}`;
  const context = getContextForFile(db, virtualPath);

  return {
    filepath: virtualPath,
    displayPath: doc.display_path,
    title: doc.title,
    context,
    hash: doc.hash,
    docid: getDocid(doc.hash),
    collectionName: doc.collection,
    modifiedAt: doc.modified_at,
    bodyLength: doc.body_length,
    ...(options.includeBody && doc.body !== undefined && { body: doc.body }),
  };
}

/**
 * Get the body content for a document
 * Optionally slice by line range
 */
export function getDocumentBody(db: Database, doc: DocumentResult | { filepath: string }, fromLine?: number, maxLines?: number): string | null {
  const filepath = doc.filepath;

  // Try to resolve document by filepath (absolute or virtual)
  let row: { body: string } | null = null;

  // Try virtual path first
  if (filepath.startsWith('qmd://')) {
    row = db.prepare(`
      SELECT content.doc as body
      FROM documents d
      JOIN content ON content.hash = d.hash
      WHERE 'qmd://' || d.collection || '/' || d.path = ? AND d.active = 1
    `).get(filepath) as { body: string } | null;
  }

  // Try absolute path by looking up in YAML collections
  if (!row) {
    const collections = collectionsListCollections();
    for (const coll of collections) {
      if (filepath.startsWith(coll.path + '/')) {
        const relativePath = filepath.slice(coll.path.length + 1);
        row = db.prepare(`
          SELECT content.doc as body
          FROM documents d
          JOIN content ON content.hash = d.hash
          WHERE d.collection = ? AND d.path = ? AND d.active = 1
        `).get(coll.name, relativePath) as { body: string } | null;
        if (row) break;
      }
    }
  }

  if (!row) return null;

  let body = row.body;
  if (fromLine !== undefined || maxLines !== undefined) {
    const lines = body.split('\n');
    const start = (fromLine || 1) - 1;
    const end = maxLines !== undefined ? start + maxLines : lines.length;
    body = lines.slice(start, end).join('\n');
  }

  return body;
}

/**
 * Find multiple documents by glob pattern or comma-separated list
 * Returns documents without body by default (use getDocumentBody to load)
 */
export function findDocuments(
  db: Database,
  pattern: string,
  options: { includeBody?: boolean; maxBytes?: number } = {}
): { docs: MultiGetResult[]; errors: string[] } {
  const isCommaSeparated = pattern.includes(',') && !pattern.includes('*') && !pattern.includes('?');
  const errors: string[] = [];
  const maxBytes = options.maxBytes ?? DEFAULT_MULTI_GET_MAX_BYTES;

  const bodyCol = options.includeBody ? `, content.doc as body` : ``;
  const selectCols = `
    'qmd://' || d.collection || '/' || d.path as virtual_path,
    d.collection || '/' || d.path as display_path,
    d.title,
    d.hash,
    d.collection,
    d.modified_at,
    LENGTH(content.doc) as body_length
    ${bodyCol}
  `;

  let fileRows: DbDocRow[];

  if (isCommaSeparated) {
    const names = pattern.split(',').map(s => s.trim()).filter(Boolean);
    fileRows = [];
    for (const name of names) {
      let doc = db.prepare(`
        SELECT ${selectCols}
        FROM documents d
        JOIN content ON content.hash = d.hash
        WHERE 'qmd://' || d.collection || '/' || d.path = ? AND d.active = 1
      `).get(name) as DbDocRow | null;
      if (!doc) {
        doc = db.prepare(`
          SELECT ${selectCols}
          FROM documents d
          JOIN content ON content.hash = d.hash
          WHERE 'qmd://' || d.collection || '/' || d.path LIKE ? AND d.active = 1
          LIMIT 1
        `).get(`%${name}`) as DbDocRow | null;
      }
      if (doc) {
        fileRows.push(doc);
      } else {
        const similar = findSimilarFiles(db, name, 5, 3);
        let msg = `File not found: ${name}`;
        if (similar.length > 0) {
          msg += ` (did you mean: ${similar.join(', ')}?)`;
        }
        errors.push(msg);
      }
    }
  } else {
    // Glob pattern match
    const matched = matchFilesByGlob(db, pattern);
    if (matched.length === 0) {
      errors.push(`No files matched pattern: ${pattern}`);
      return { docs: [], errors };
    }
    const virtualPaths = matched.map(m => m.filepath);
    const placeholders = virtualPaths.map(() => '?').join(',');
    fileRows = db.prepare(`
      SELECT ${selectCols}
      FROM documents d
      JOIN content ON content.hash = d.hash
      WHERE 'qmd://' || d.collection || '/' || d.path IN (${placeholders}) AND d.active = 1
    `).all(...virtualPaths) as DbDocRow[];
  }

  const results: MultiGetResult[] = [];

  for (const row of fileRows) {
    // Get context using virtual path
    const virtualPath = row.virtual_path || `qmd://${row.collection}/${row.display_path}`;
    const context = getContextForFile(db, virtualPath);

    if (row.body_length > maxBytes) {
      results.push({
        doc: { filepath: virtualPath, displayPath: row.display_path },
        skipped: true,
        skipReason: `File too large (${Math.round(row.body_length / 1024)}KB > ${Math.round(maxBytes / 1024)}KB)`,
      });
      continue;
    }

    results.push({
      doc: {
        filepath: virtualPath,
        displayPath: row.display_path,
        title: row.title || row.display_path.split('/').pop() || row.display_path,
        context,
        hash: row.hash,
        docid: getDocid(row.hash),
        collectionName: row.collection,
        modifiedAt: row.modified_at,
        bodyLength: row.body_length,
        ...(options.includeBody && row.body !== undefined && { body: row.body }),
      },
      skipped: false,
    });
  }

  return { docs: results, errors };
}

// =============================================================================
// Status
// =============================================================================

export function getStatus(db: Database): IndexStatus {
  // Load collections from YAML
  const yamlCollections = collectionsListCollections();

  // Get document counts and last update times for each collection
  const collections = yamlCollections.map(col => {
    const stats = db.prepare(`
      SELECT
        COUNT(*) as active_count,
        MAX(modified_at) as last_doc_update
      FROM documents
      WHERE collection = ? AND active = 1
    `).get(col.name) as { active_count: number; last_doc_update: string | null };

    return {
      name: col.name,
      path: col.path,
      pattern: col.pattern,
      documents: stats.active_count,
      lastUpdated: stats.last_doc_update || new Date().toISOString(),
    };
  });

  // Sort by last update time (most recent first)
  collections.sort((a, b) => {
    if (!a.lastUpdated) return 1;
    if (!b.lastUpdated) return -1;
    return new Date(b.lastUpdated).getTime() - new Date(a.lastUpdated).getTime();
  });

  const totalDocs = (db.prepare(`SELECT COUNT(*) as c FROM documents WHERE active = 1`).get() as { c: number }).c;
  const needsEmbedding = getHashesNeedingEmbedding(db);
  const hasVectors = !!db.prepare(`SELECT name FROM sqlite_master WHERE type='table' AND name='vectors_vec'`).get();

  return {
    totalDocuments: totalDocs,
    needsEmbedding,
    hasVectorIndex: hasVectors,
    collections,
  };
}

// =============================================================================
// Snippet extraction
// =============================================================================

export type SnippetResult = {
  line: number;           // 1-indexed line number of best match
  snippet: string;        // The snippet text with diff-style header
  linesBefore: number;    // Lines in document before snippet
  linesAfter: number;     // Lines in document after snippet
  snippetLines: number;   // Number of lines in snippet
};

export function extractSnippet(body: string, query: string, maxLen = 500, chunkPos?: number, chunkLen?: number): SnippetResult {
  const totalLines = body.split('\n').length;
  let searchBody = body;
  let lineOffset = 0;

  if (chunkPos && chunkPos > 0) {
    // Search within the chunk region, with some padding for context
    // Use provided chunkLen or fall back to max chunk size (covers variable-length chunks)
    const searchLen = chunkLen || CHUNK_SIZE_CHARS;
    const contextStart = Math.max(0, chunkPos - 100);
    const contextEnd = Math.min(body.length, chunkPos + searchLen + 100);
    searchBody = body.slice(contextStart, contextEnd);
    if (contextStart > 0) {
      lineOffset = body.slice(0, contextStart).split('\n').length - 1;
    }
  }

  const lines = searchBody.split('\n');
  const queryTerms = query.toLowerCase().split(/\s+/).filter(t => t.length > 0);
  let bestLine = 0, bestScore = -1;

  for (let i = 0; i < lines.length; i++) {
    const lineLower = (lines[i] ?? "").toLowerCase();
    let score = 0;
    for (const term of queryTerms) {
      if (lineLower.includes(term)) score++;
    }
    if (score > bestScore) {
      bestScore = score;
      bestLine = i;
    }
  }

  const start = Math.max(0, bestLine - 1);
  const end = Math.min(lines.length, bestLine + 3);
  const snippetLines = lines.slice(start, end);
  let snippetText = snippetLines.join('\n');

  // If we focused on a chunk window and it produced an empty/whitespace-only snippet,
  // fall back to a full-document snippet so we always show something useful.
  if (chunkPos && chunkPos > 0 && snippetText.trim().length === 0) {
    return extractSnippet(body, query, maxLen, undefined);
  }

  if (snippetText.length > maxLen) snippetText = snippetText.substring(0, maxLen - 3) + "...";

  const absoluteStart = lineOffset + start + 1; // 1-indexed
  const snippetLineCount = snippetLines.length;
  const linesBefore = absoluteStart - 1;
  const linesAfter = totalLines - (absoluteStart + snippetLineCount - 1);

  // Format with diff-style header: @@ -start,count @@ (linesBefore before, linesAfter after)
  const header = `@@ -${absoluteStart},${snippetLineCount} @@ (${linesBefore} before, ${linesAfter} after)`;
  const snippet = `${header}\n${snippetText}`;

  return {
    line: lineOffset + bestLine + 1,
    snippet,
    linesBefore,
    linesAfter,
    snippetLines: snippetLineCount,
  };
}

// =============================================================================
// Shared helpers (used by both CLI and MCP)
// =============================================================================

/**
 * Add line numbers to text content.
 * Each line becomes: "{lineNum}: {content}"
 */
export function addLineNumbers(text: string, startLine: number = 1): string {
  const lines = text.split('\n');
  return lines.map((line, i) => `${startLine + i}: ${line}`).join('\n');
}

// =============================================================================
// Shared search orchestration
//
// hybridQuery() and vectorSearchQuery() are standalone functions (not Store
// methods) because they are orchestration over primitives — same rationale as
// reciprocalRankFusion(). They take a Store as first argument so both CLI
// and MCP can share the identical pipeline.
// =============================================================================

/**
 * Optional progress hooks for search orchestration.
 * CLI wires these to stderr for user feedback; MCP leaves them unset.
 */
export interface SearchHooks {
  /** BM25 probe found strong signal — expansion will be skipped */
  onStrongSignal?: (topScore: number) => void;
  /** Query expansion complete. Empty array = strong signal skip (no expansion). */
  onExpand?: (original: string, expanded: ExpandedQuery[]) => void;
  /** Reranking is about to start */
  onRerankStart?: (chunkCount: number) => void;
  /** Reranking finished */
  onRerankDone?: () => void;
}

export interface HybridQueryOptions {
  collection?: string;
  limit?: number;           // default 10
  minScore?: number;        // default 0
  candidateLimit?: number;  // default RERANK_CANDIDATE_LIMIT
  hooks?: SearchHooks;
}

export interface HybridQueryResult {
  file: string;             // internal filepath (qmd://collection/path)
  displayPath: string;
  title: string;
  body: string;             // full document body (for snippet extraction)
  bestChunk: string;        // best chunk text
  bestChunkPos: number;     // char offset of best chunk in body
  score: number;            // blended score (full precision)
  context: string | null;   // user-set context
  docid: string;            // content hash prefix (6 chars)
}

/**
 * Hybrid search: BM25 + vector + query expansion + RRF + chunked reranking.
 *
 * Pipeline:
 * 1. BM25 probe → skip expansion if strong signal
 * 2. expandQuery() → typed query variants (lex/vec/hyde)
 * 3. Type-routed search: original→vector, lex→FTS, vec/hyde→vector
 * 4. RRF fusion → slice to candidateLimit
 * 5. chunkDocument() + keyword-best-chunk selection
 * 6. rerank on chunks (NOT full bodies — O(tokens) trap)
 * 7. Position-aware score blending (RRF rank × reranker score)
 * 8. Dedup by file, filter by minScore, slice to limit
 */
export async function hybridQuery(
  store: Store,
  query: string,
  options?: HybridQueryOptions
): Promise<HybridQueryResult[]> {
  const limit = options?.limit ?? 10;
  const minScore = options?.minScore ?? 0;
  const candidateLimit = options?.candidateLimit ?? RERANK_CANDIDATE_LIMIT;
  const collection = options?.collection;
  const hooks = options?.hooks;

  const rankedLists: RankedResult[][] = [];
  const docidMap = new Map<string, string>(); // filepath -> docid
  const hasVectors = !!store.db.prepare(
    `SELECT name FROM sqlite_master WHERE type='table' AND name='vectors_vec'`
  ).get();

  // Step 1: BM25 probe — strong signal skips expensive LLM expansion
  // Pass collection directly into FTS query (filter at SQL level, not post-hoc)
  const initialFts = store.searchFTS(query, 20, collection);
  const topScore = initialFts[0]?.score ?? 0;
  const secondScore = initialFts[1]?.score ?? 0;
  const hasStrongSignal = initialFts.length > 0
    && topScore >= STRONG_SIGNAL_MIN_SCORE
    && (topScore - secondScore) >= STRONG_SIGNAL_MIN_GAP;

  if (hasStrongSignal) hooks?.onStrongSignal?.(topScore);

  // Step 2: Expand query (or skip if strong signal)
  const expanded = hasStrongSignal
    ? []
    : await store.expandQuery(query);

  hooks?.onExpand?.(query, expanded);

  // Seed with initial FTS results (avoid re-running original query FTS)
  if (initialFts.length > 0) {
    for (const r of initialFts) docidMap.set(r.filepath, r.docid);
    rankedLists.push(initialFts.map(r => ({
      file: r.filepath, displayPath: r.displayPath,
      title: r.title, body: r.body || "", score: r.score,
    })));
  }

  // Step 3: Route searches by query type
  //
  // Strategy: run all FTS queries immediately (they're sync/instant), then
  // batch-embed all vector queries in one embedBatch() call, then run
  // sqlite-vec lookups with pre-computed embeddings.

  // 3a: Run FTS for all lex expansions right away (no LLM needed)
  for (const q of expanded) {
    if (q.type === 'lex') {
      const ftsResults = store.searchFTS(q.text, 20, collection);
      if (ftsResults.length > 0) {
        for (const r of ftsResults) docidMap.set(r.filepath, r.docid);
        rankedLists.push(ftsResults.map(r => ({
          file: r.filepath, displayPath: r.displayPath,
          title: r.title, body: r.body || "", score: r.score,
        })));
      }
    }
  }

  // 3b: Collect all texts that need vector search (original query + vec/hyde expansions)
  if (hasVectors) {
    const vecQueries: { text: string; isOriginal: boolean }[] = [
      { text: query, isOriginal: true },
    ];
    for (const q of expanded) {
      if (q.type === 'vec' || q.type === 'hyde') {
        vecQueries.push({ text: q.text, isOriginal: false });
      }
    }

    // Batch embed all vector queries in a single call
    const llm = getDefaultLlamaCpp();
    const textsToEmbed = vecQueries.map(q => formatQueryForEmbedding(q.text));
    const embeddings = await llm.embedBatch(textsToEmbed);

    // Run sqlite-vec lookups with pre-computed embeddings
    for (let i = 0; i < vecQueries.length; i++) {
      const embedding = embeddings[i]?.embedding;
      if (!embedding) continue;

      const vecResults = await store.searchVec(
        vecQueries[i]!.text, DEFAULT_EMBED_MODEL, 20, collection,
        undefined, embedding
      );
      if (vecResults.length > 0) {
        for (const r of vecResults) docidMap.set(r.filepath, r.docid);
        rankedLists.push(vecResults.map(r => ({
          file: r.filepath, displayPath: r.displayPath,
          title: r.title, body: r.body || "", score: r.score,
        })));
      }
    }
  }

  // Step 4: RRF fusion — first 2 lists (original FTS + first vec) get 2x weight
  const weights = rankedLists.map((_, i) => i < 2 ? 2.0 : 1.0);
  const fused = reciprocalRankFusion(rankedLists, weights);
  const candidates = fused.slice(0, candidateLimit);

  if (candidates.length === 0) return [];

  // Step 5: Chunk documents, pick best chunk per doc for reranking.
  // Reranking full bodies is O(tokens) — the critical perf lesson that motivated this refactor.
  const queryTerms = query.toLowerCase().split(/\s+/).filter(t => t.length > 2);
  const chunksToRerank: { file: string; text: string }[] = [];
  const docChunkMap = new Map<string, { chunks: { text: string; pos: number }[]; bestIdx: number }>();

  for (const cand of candidates) {
    const chunks = chunkDocument(cand.body);
    if (chunks.length === 0) continue;

    // Pick chunk with most keyword overlap (fallback: first chunk)
    let bestIdx = 0;
    let bestScore = -1;
    for (let i = 0; i < chunks.length; i++) {
      const chunkLower = chunks[i]!.text.toLowerCase();
      const score = queryTerms.reduce((acc, term) => acc + (chunkLower.includes(term) ? 1 : 0), 0);
      if (score > bestScore) { bestScore = score; bestIdx = i; }
    }

    chunksToRerank.push({ file: cand.file, text: chunks[bestIdx]!.text });
    docChunkMap.set(cand.file, { chunks, bestIdx });
  }

  // Step 6: Rerank chunks (NOT full bodies)
  hooks?.onRerankStart?.(chunksToRerank.length);
  const reranked = await store.rerank(query, chunksToRerank);
  hooks?.onRerankDone?.();

  // Step 7: Blend RRF position score with reranker score
  // Position-aware weights: top retrieval results get more protection from reranker disagreement
  const candidateMap = new Map(candidates.map(c => [c.file, {
    displayPath: c.displayPath, title: c.title, body: c.body,
  }]));
  const rrfRankMap = new Map(candidates.map((c, i) => [c.file, i + 1]));

  const blended = reranked.map(r => {
    const rrfRank = rrfRankMap.get(r.file) || candidateLimit;
    let rrfWeight: number;
    if (rrfRank <= 3) rrfWeight = 0.75;
    else if (rrfRank <= 10) rrfWeight = 0.60;
    else rrfWeight = 0.40;
    const rrfScore = 1 / rrfRank;
    const blendedScore = rrfWeight * rrfScore + (1 - rrfWeight) * r.score;

    const candidate = candidateMap.get(r.file);
    const chunkInfo = docChunkMap.get(r.file);
    const bestIdx = chunkInfo?.bestIdx ?? 0;
    const bestChunk = chunkInfo?.chunks[bestIdx]?.text || candidate?.body || "";
    const bestChunkPos = chunkInfo?.chunks[bestIdx]?.pos || 0;

    return {
      file: r.file,
      displayPath: candidate?.displayPath || "",
      title: candidate?.title || "",
      body: candidate?.body || "",
      bestChunk,
      bestChunkPos,
      score: blendedScore,
      context: store.getContextForFile(r.file),
      docid: docidMap.get(r.file) || "",
    };
  }).sort((a, b) => b.score - a.score);

  // Step 8: Dedup by file (safety net — prevents duplicate output)
  const seenFiles = new Set<string>();
  return blended
    .filter(r => {
      if (seenFiles.has(r.file)) return false;
      seenFiles.add(r.file);
      return true;
    })
    .filter(r => r.score >= minScore)
    .slice(0, limit);
}

export interface VectorSearchOptions {
  collection?: string;
  limit?: number;           // default 10
  minScore?: number;        // default 0.3
  hooks?: Pick<SearchHooks, 'onExpand'>;
}

export interface VectorSearchResult {
  file: string;
  displayPath: string;
  title: string;
  body: string;
  score: number;
  context: string | null;
  docid: string;
}

/**
 * Vector-only semantic search with query expansion.
 *
 * Pipeline:
 * 1. expandQuery() → typed variants, filter to vec/hyde only (lex irrelevant here)
 * 2. searchVec() for original + vec/hyde variants (sequential — node-llama-cpp embed limitation)
 * 3. Dedup by filepath (keep max score)
 * 4. Sort by score descending, filter by minScore, slice to limit
 */
export async function vectorSearchQuery(
  store: Store,
  query: string,
  options?: VectorSearchOptions
): Promise<VectorSearchResult[]> {
  const limit = options?.limit ?? 10;
  const minScore = options?.minScore ?? 0.3;
  const collection = options?.collection;

  const hasVectors = !!store.db.prepare(
    `SELECT name FROM sqlite_master WHERE type='table' AND name='vectors_vec'`
  ).get();
  if (!hasVectors) return [];

  // Expand query — filter to vec/hyde only (lex queries target FTS, not vector)
  const allExpanded = await store.expandQuery(query);
  const vecExpanded = allExpanded.filter(q => q.type !== 'lex');
  options?.hooks?.onExpand?.(query, vecExpanded);

  // Run original + vec/hyde expanded through vector, sequentially — concurrent embed() hangs
  const queryTexts = [query, ...vecExpanded.map(q => q.text)];
  const allResults = new Map<string, VectorSearchResult>();
  for (const q of queryTexts) {
    const vecResults = await store.searchVec(q, DEFAULT_EMBED_MODEL, limit, collection);
    for (const r of vecResults) {
      const existing = allResults.get(r.filepath);
      if (!existing || r.score > existing.score) {
        allResults.set(r.filepath, {
          file: r.filepath,
          displayPath: r.displayPath,
          title: r.title,
          body: r.body || "",
          score: r.score,
          context: store.getContextForFile(r.filepath),
          docid: r.docid,
        });
      }
    }
  }

  return Array.from(allResults.values())
    .sort((a, b) => b.score - a.score)
    .filter(r => r.score >= minScore)
    .slice(0, limit);
}

// =============================================================================
// Structured search — pre-expanded queries from LLM
// =============================================================================

/**
 * A single sub-search in a structured search request.
 * Matches the format used in QMD training data.
 */
export interface StructuredSubSearch {
  /** Search type: 'lex' for BM25 keywords, 'vec' for semantic, 'hyde' for hypothetical document */
  type: 'lex' | 'vec' | 'hyde';
  /** The search query text */
  query: string;
}

export interface StructuredSearchOptions {
  collection?: string;      // Single collection filter
  collections?: string[];   // Multiple collections filter (OR)
  limit?: number;           // default 10
  minScore?: number;        // default 0
  candidateLimit?: number;  // default RERANK_CANDIDATE_LIMIT
  /** Future: domain intent hint for routing/boosting */
  intent?: string;
  hooks?: SearchHooks;
}

/**
 * Structured search: execute pre-expanded queries without LLM query expansion.
 *
 * Designed for LLM callers (MCP/HTTP) that generate their own query expansions.
 * Skips the internal expandQuery() step — goes directly to:
 *
 * Pipeline:
 * 1. Route searches: lex→FTS, vec/hyde→vector (batch embed)
 * 2. RRF fusion across all result lists
 * 3. Chunk documents + keyword-best-chunk selection
 * 4. Rerank on chunks
 * 5. Position-aware score blending
 * 6. Dedup, filter, slice
 *
 * This is the recommended endpoint for capable LLMs — they can generate
 * better query variations than our small local model, especially for
 * domain-specific or nuanced queries.
 */
export async function structuredSearch(
  store: Store,
  searches: StructuredSubSearch[],
  options?: StructuredSearchOptions
): Promise<HybridQueryResult[]> {
  const limit = options?.limit ?? 10;
  const minScore = options?.minScore ?? 0;
  const candidateLimit = options?.candidateLimit ?? RERANK_CANDIDATE_LIMIT;
  const hooks = options?.hooks;

  // Normalize collection filter to array (undefined = all collections)
  const collections: string[] | undefined = options?.collections
    ?? (options?.collection ? [options.collection] : undefined);

  if (searches.length === 0) return [];

  // Validate semantic queries don't use lex-only syntax
  for (const search of searches) {
    if (search.type === 'vec' || search.type === 'hyde') {
      const error = validateSemanticQuery(search.query);
      if (error) {
        throw new Error(`Invalid ${search.type} query: ${error}`);
      }
    }
  }

  const rankedLists: RankedResult[][] = [];
  const docidMap = new Map<string, string>(); // filepath -> docid
  const hasVectors = !!store.db.prepare(
    `SELECT name FROM sqlite_master WHERE type='table' AND name='vectors_vec'`
  ).get();

  // Helper to run search across collections (or all if undefined)
  const collectionList = collections ?? [undefined]; // undefined = all collections

  // Step 1: Run FTS for all lex searches (sync, instant)
  for (const search of searches) {
    if (search.type === 'lex') {
      for (const coll of collectionList) {
        const ftsResults = store.searchFTS(search.query, 20, coll);
        if (ftsResults.length > 0) {
          for (const r of ftsResults) docidMap.set(r.filepath, r.docid);
          rankedLists.push(ftsResults.map(r => ({
            file: r.filepath, displayPath: r.displayPath,
            title: r.title, body: r.body || "", score: r.score,
          })));
        }
      }
    }
  }

  // Step 2: Batch embed and run vector searches for vec/hyde
  if (hasVectors) {
    const vecSearches = searches.filter(s => s.type === 'vec' || s.type === 'hyde');
    if (vecSearches.length > 0) {
      const llm = getDefaultLlamaCpp();
      const textsToEmbed = vecSearches.map(s => formatQueryForEmbedding(s.query));
      const embeddings = await llm.embedBatch(textsToEmbed);

      for (let i = 0; i < vecSearches.length; i++) {
        const embedding = embeddings[i]?.embedding;
        if (!embedding) continue;

        for (const coll of collectionList) {
          const vecResults = await store.searchVec(
            vecSearches[i]!.query, DEFAULT_EMBED_MODEL, 20, coll,
            undefined, embedding
          );
          if (vecResults.length > 0) {
            for (const r of vecResults) docidMap.set(r.filepath, r.docid);
            rankedLists.push(vecResults.map(r => ({
              file: r.filepath, displayPath: r.displayPath,
              title: r.title, body: r.body || "", score: r.score,
            })));
          }
        }
      }
    }
  }

  if (rankedLists.length === 0) return [];

  // Step 3: RRF fusion — first list gets 2x weight (assume caller ordered by importance)
  const weights = rankedLists.map((_, i) => i === 0 ? 2.0 : 1.0);
  const fused = reciprocalRankFusion(rankedLists, weights);
  const candidates = fused.slice(0, candidateLimit);

  if (candidates.length === 0) return [];

  hooks?.onExpand?.("", []); // Signal no expansion (pre-expanded)

  // Step 4: Chunk documents, pick best chunk per doc for reranking
  // Use first lex query as the "query" for keyword matching, or first vec if no lex
  const primaryQuery = searches.find(s => s.type === 'lex')?.query
    || searches.find(s => s.type === 'vec')?.query
    || searches[0]?.query || "";
  const queryTerms = primaryQuery.toLowerCase().split(/\s+/).filter(t => t.length > 2);
  const chunksToRerank: { file: string; text: string }[] = [];
  const docChunkMap = new Map<string, { chunks: { text: string; pos: number }[]; bestIdx: number }>();

  for (const cand of candidates) {
    const chunks = chunkDocument(cand.body);
    if (chunks.length === 0) continue;

    // Pick chunk with most keyword overlap
    let bestIdx = 0;
    let bestScore = -1;
    for (let i = 0; i < chunks.length; i++) {
      const chunkLower = chunks[i]!.text.toLowerCase();
      const score = queryTerms.reduce((acc, term) => acc + (chunkLower.includes(term) ? 1 : 0), 0);
      if (score > bestScore) { bestScore = score; bestIdx = i; }
    }

    chunksToRerank.push({ file: cand.file, text: chunks[bestIdx]!.text });
    docChunkMap.set(cand.file, { chunks, bestIdx });
  }

  // Step 5: Rerank chunks
  hooks?.onRerankStart?.(chunksToRerank.length);
  const reranked = await store.rerank(primaryQuery, chunksToRerank);
  hooks?.onRerankDone?.();

  // Step 6: Blend RRF position score with reranker score
  const candidateMap = new Map(candidates.map(c => [c.file, {
    displayPath: c.displayPath, title: c.title, body: c.body,
  }]));
  const rrfRankMap = new Map(candidates.map((c, i) => [c.file, i + 1]));

  const blended = reranked.map(r => {
    const rrfRank = rrfRankMap.get(r.file) || candidateLimit;
    let rrfWeight: number;
    if (rrfRank <= 3) rrfWeight = 0.75;
    else if (rrfRank <= 10) rrfWeight = 0.60;
    else rrfWeight = 0.40;
    const rrfScore = 1 / rrfRank;
    const blendedScore = rrfWeight * rrfScore + (1 - rrfWeight) * r.score;

    const candidate = candidateMap.get(r.file);
    const chunkInfo = docChunkMap.get(r.file);
    const bestIdx = chunkInfo?.bestIdx ?? 0;
    const bestChunk = chunkInfo?.chunks[bestIdx]?.text || candidate?.body || "";
    const bestChunkPos = chunkInfo?.chunks[bestIdx]?.pos || 0;

    return {
      file: r.file,
      displayPath: candidate?.displayPath || "",
      title: candidate?.title || "",
      body: candidate?.body || "",
      bestChunk,
      bestChunkPos,
      score: blendedScore,
      context: store.getContextForFile(r.file),
      docid: docidMap.get(r.file) || "",
    };
  }).sort((a, b) => b.score - a.score);

  // Step 7: Dedup by file
  const seenFiles = new Set<string>();
  return blended
    .filter(r => {
      if (seenFiles.has(r.file)) return false;
      seenFiles.add(r.file);
      return true;
    })
    .filter(r => r.score >= minScore)
    .slice(0, limit);
}