/**
 * VectorIndex
 *
 * Builds and queries a LanceDB vector index over the call graph functions.
 * Each function is represented by a document combining its signature, docstring,
 * file path, language, and topological metadata (fanIn/fanOut, hub, entry point).
 *
 * Storage: <outputDir>/vector-index/  (LanceDB database folder)
 * Table name: "functions"
 *
 * Usage:
 *   // Build (after openlore analyze --embed)
 *   await VectorIndex.build(outputDir, nodes, signatures, hubIds, entryPointIds, embedSvc);
 *
 *   // Search
 *   const results = await VectorIndex.search(outputDir, "authenticate user with JWT", embedSvc);
 */
import type { CallEdge, FunctionNode } from './call-graph.js';
import type { FileSignatureMap } from './signature-extractor.js';
import type { Embedder } from './embedding-service.js';
import { type MatchEvidence, type SearchableFields } from './retrieval-evidence.js';
export type { MatchEvidence, MatchField, RetrievalTier } from './retrieval-evidence.js';
export { TOKENIZER_VERSION, tokenize } from './bm25-tokenizer.js';
export interface FunctionRecord {
    id: string;
    name: string;
    filePath: string;
    className: string;
    language: string;
    signature: string;
    docstring: string;
    fanIn: number;
    fanOut: number;
    isHub: boolean;
    isEntryPoint: boolean;
    /** Concatenated text used for embedding */
    text: string;
    /** Embedding vector */
    vector: number[];
}
export interface SearchResult {
    record: Omit<FunctionRecord, 'vector'>;
    /**
     * Relevance score.  For hybrid search (default): RRF score, higher = more relevant.
     * For dense-only search: cosine distance from LanceDB, lower = more similar.
     */
    score: number;
    /** Optional for legacy test doubles; every production search path emits it. */
    scoreKind?: 'rrf' | 'bm25' | 'cosine_distance';
    /** Optional for legacy test doubles; every production search path emits it. */
    matchEvidence?: MatchEvidence;
    /** Vocabulary terms that contributed a non-zero score to this result. */
    expansionTerms?: string[];
}
export interface KeywordMissDiagnostics {
    missedTokens: string[];
    nearTokens: Array<{
        queryToken: string;
        indexedTokens: string[];
    }>;
}
export interface VectorIndexMeta {
    hasEmbeddings: boolean;
    dim: number;
    model: string | null;
    builtAt: string;
    /** Changes only after a complete rebuild; incremental mutations preserve it. */
    fullBuildAt?: string;
    schemaVersion: number;
    /**
     * Version of the BM25 tokenizer that produced this index's corpus. A mismatch
     * against the running `TOKENIZER_VERSION` means an incremental patch would mix
     * token sets, so `updateFiles` defers (`deferred: 'tokenizer-changed'`) and a
     * full rebuild re-stamps. A legacy meta without this field is treated as v1.
     */
    tokenizerVersion?: number;
    /** Corpus identity of the verified repository-vocabulary sidecar, when present. */
    vocabularyContentStamp?: string;
    /** Present only when an incremental update could neither add nor restore rows. */
    degraded?: {
        reason: 'incremental-update-restore-failed';
        recordedAt: string;
    };
}
declare function dbPathFor(outputDir: string): string;
export interface Bm25Corpus {
    docs: Array<{
        id: string;
        tfMap: Map<string, number>;
        length: number;
    }>;
    /** term → number of documents containing it */
    df: Map<string, number>;
    avgLength: number;
    N: number;
}
export declare function buildBm25Corpus(records: Array<{
    id: string;
    text: string;
}>): Bm25Corpus;
/**
 * The per-corpus lookup structures a keyword query needs, built once per corpus OBJECT
 * and carried across an incremental patch rather than rebuilt.
 *
 * `byTerm` maps a term to the ids of the documents containing it; `indexById` maps a
 * document id back to its position in `corpus.docs`. Postings are keyed by ID rather
 * than by position precisely so a patch can maintain them in O(edit): a patch drops
 * removed documents from the middle of `docs`, which renumbers every position after
 * them, but leaves every surviving document's id alone.
 *
 * Neither structure is part of {@link Bm25Corpus} and neither reaches the persisted
 * sidecar — both are derivable from `docs`.
 * (change: optimize-serving-hot-path-caches)
 */
interface Bm25Postings {
    byTerm: Map<string, Set<string>>;
    indexById: Map<string, number>;
    /** False when two documents share an id, which would make `indexById` lossy. */
    idsAreUnique: boolean;
}
export declare function _bm25WorkCountersForTesting(): {
    postingsBuilds: number;
    docsScored: number;
};
export declare function _resetBm25WorkCountersForTesting(): void;
/**
 * The doc indices that can score above zero for any of `tokenSets`, ascending.
 *
 * Exactly the set the caller's `score > 0 || expansionScore > 0` filter would keep:
 * `bm25Score` contributes only for a term with `df > 0` AND `tf > 0`, and both the idf
 * and the tf-norm factors are strictly positive whenever those hold — so "appears in
 * some posting list" and "scores above zero" are the same predicate. Scoring the
 * candidates instead of every document is what keeps a keyword query from walking the
 * whole corpus.
 *
 * If two documents ever share an id the position map is lossy, so this returns every
 * index rather than silently dropping one — a slow answer, never a wrong one.
 * (change: optimize-serving-hot-path-caches)
 */
export declare function bm25CandidateDocs(corpus: Bm25Corpus, ...tokenSets: ReadonlyArray<readonly string[]>): number[];
export declare function bm25Score(corpus: Bm25Corpus, queryTokens: string[], docIdx: number): number;
/** Attribute the ranker's exact aggregate term contributions across one bounded candidate's fields. */
export declare function bm25MatchEvidence(corpus: Bm25Corpus, queryTokens: string[], docIdx: number, fields: SearchableFields, tier?: 1 | 2): MatchEvidence;
declare const _cacheStats: {
    tableHits: number;
    tableMisses: number;
    bm25Hits: number;
    bm25Misses: number;
};
/** Clear every process-lifetime cache for one on-disk vector index. */
export declare function invalidateVectorIndexCaches(outputDir: string): void;
/** Test-only: expose the canonical cache identity used for an index path. */
export declare const _vectorIndexCacheIdentityForTesting: typeof dbPathFor;
/** Test-only: clear in-memory BM25 + LanceDB caches to force cold path. */
export declare function _resetVectorIndexCachesForTesting(): void;
/** Test-only proof that a cold request populated, and a warm request reused, each search cache. */
export declare function _vectorIndexCacheStatsForTesting(): Readonly<typeof _cacheStats>;
interface MutableTable {
    delete(predicate: string): Promise<unknown>;
    add(rows: Record<string, unknown>[]): Promise<unknown>;
}
declare function replaceRowsWithRestore(table: MutableTable, predicate: string | null, replacementRows: Record<string, unknown>[], previousRows: Record<string, unknown>[], onRestoreFailure: () => Promise<void>): Promise<void>;
/** Test-only: exercise the transactional delete/add helper without LanceDB. */
export declare const _replaceRowsWithRestoreForTesting: typeof replaceRowsWithRestore;
/**
 * Absorb one incremental update into an existing corpus, returning the new corpus and rows.
 *
 * Pure — no cache, no disk — so it can be checked directly against {@link buildBm25Corpus}, which
 * is the only assurance that matters here (see `bm25-incremental-patch.test.ts`).
 */
declare function patchBm25Corpus(previous: Bm25Corpus, previousRows: Record<string, unknown>[], changedFilePaths: Set<string>, newRows: Record<string, unknown>[], previousPostings?: Bm25Postings): {
    corpus: Bm25Corpus;
    rows: Record<string, unknown>[];
    postings?: Bm25Postings;
};
/** Test-only: drive {@link patchBm25Corpus} directly, to diff it against a full rebuild. */
export declare const _patchBm25CorpusForTesting: typeof patchBm25Corpus;
/**
 * Test-only: patch a corpus the way {@link patchBm25Cache} does — carrying the memoized
 * postings index across the new corpus object instead of leaving it to be rebuilt.
 * The cache-side half of the pure {@link patchBm25Corpus}, isolated from LanceDB.
 */
export declare function _patchBm25CorpusCarryingPostingsForTesting(previous: Bm25Corpus, previousRows: Record<string, unknown>[], changedFilePaths: Set<string>, newRows: Record<string, unknown>[]): Bm25Corpus;
export declare function searchableFieldsForFunctionRow(row: Record<string, unknown>): SearchableFields;
export declare class VectorIndex {
    /** User-facing disclosure for an index whose incremental rollback also failed. */
    static degradationNotice(outputDir: string): string | null;
    /**
     * Build (or rebuild) the vector index from call graph nodes + signatures.
     *
     * When `incremental` is true and an existing index is found, only functions
     * whose text has changed since the last build are re-embedded.  Unchanged
     * functions reuse their cached vectors.  Pass `incremental: false` (or omit
     * when no index exists) to do a full rebuild.
     *
     * Returns a summary of how many functions were embedded vs reused.
     *
     * When `embedSvc` is null, builds a **keyword-only (BM25)** index: the corpus
     * rows are written without a `vector` column and the meta sidecar records
     * `hasEmbeddings: false`. Search then serves BM25 results and never attempts
     * ANN. Re-building a previously-embedded index with `embedSvc=null` downgrades
     * it to BM25-only (overwrite + meta update), and vice-versa upgrades it.
     */
    static build(outputDir: string, nodes: FunctionNode[], signatures: FileSignatureMap[], hubIds: Set<string>, entryPointIds: Set<string>, embedSvc: Embedder | null, 
    /** Optional map of filePath → source content for skeleton-based body indexing */
    fileContents?: Map<string, string>, 
    /** When true, reuse cached vectors for unchanged functions */
    incremental?: boolean, vocabularyExpansion?: boolean, callEdges?: readonly Pick<CallEdge, 'callerId' | 'calleeId'>[]): Promise<{
        embedded: number;
        reused: number;
        total: number;
        hasEmbeddings: boolean;
        productionFunctions: number;
        testFunctions: number;
        signatureOnlySymbols: number;
    }>;
    private static buildUnlocked;
    /**
     * Watch-mode incremental update (Spec 13.1). Replace only the rows for the
     * changed files with freshly-built records — a row-level delete+add instead of
     * the full-corpus read+overwrite that build() performs. The cold build() path
     * is untouched, protecting the `analyze --embed` contract (G7).
     *
     *  - Embedded index: reuse existing vectors for rows whose embed-text is
     *    unchanged (queried for the changed files only, not the whole corpus),
     *    embed just the new/changed texts, then delete the changed files' old rows
     *    and add the rebuilt ones. The LanceDB table handle in _tableCache stays
     *    valid across row ops, so search() does not pay a reconnect.
     *  - BM25-only index: delete+add the changed files' documents and patch the
     *    cached BM25 corpus in place rather than dropping the whole corpus cache.
     */
    static updateFiles(outputDir: string, nodes: FunctionNode[], changedFilePaths: Set<string>, signatures: FileSignatureMap[], hubIds: Set<string>, entryPointIds: Set<string>, embedSvc: Embedder | null | undefined, fileContents?: Map<string, string>): Promise<{
        embedded: number;
        reused: number;
        total: number;
        hasEmbeddings: boolean;
        deferred?: 'model-changed' | 'tokenizer-changed';
    }>;
    private static updateFilesUnlocked;
    /**
     * Hybrid search over the index: dense (ANN) + sparse (BM25) merged via RRF.
     *
     * Dense recall fetches top `limit*5` candidates from the vector index.
     * Sparse recall scores the full corpus with BM25 (cached per session).
     * Reciprocal Rank Fusion (RRF) combines both rankings into a single list.
     *
     * Set `hybrid: false` to use dense-only search (original behaviour).
     * Returns up to `limit` results sorted by relevance (highest first).
     */
    static search(outputDir: string, query: string, embedSvc: Embedder | null | undefined, opts?: {
        limit?: number;
        language?: string;
        minFanIn?: number;
        /** Enable hybrid dense+sparse retrieval via RRF (default: true when embedSvc available) */
        hybrid?: boolean;
        /** Internal diagnostic: return the ordinary bounded candidate window before result cutoff. */
        traceCandidates?: boolean;
        /** Disable repository-vocabulary query expansion without rebuilding. */
        vocabularyExpansion?: boolean;
        /** Reports the retrieval mode that actually produced this result set. */
        onRetrievalMode?: (mode: 'keyword' | 'keyword+vocabulary' | 'semantic') => void;
    }): Promise<SearchResult[]>;
    /**
     * BM25-only search: used when no embedding service is available.
     * Scores the full corpus with BM25 and returns the top `limit` results.
     */
    private static _bm25Only;
    /**
     * Explain an empty keyword result using the corpus already loaded by search().
     * The lookup is bounded and deterministic: at most 12 distinct query tokens,
     * 3 near identifier tokens per miss, and no model or secondary index.
     */
    static keywordMissDiagnostics(outputDir: string, query: string): Promise<KeywordMissDiagnostics>;
    /**
     * Returns true if a vector index has been built for this output directory.
     */
    static exists(outputDir: string): boolean;
}
//# sourceMappingURL=vector-index.d.ts.map