@wiati/dfe-agent 0.1.0

package/dist/query/embedder.js

@@ -0,0 +1,69 @@
+/**
+ * embedder.ts — wrapper de @xenova/transformers com LRU cache em memoria.
+ *
+ * @see PLAN_SPRINT14.md Task E.1
+ *
+ * Decisao D6: modelo canonico e `paraphrase-multilingual-MiniLM-L12-v2`
+ * (mesmo do DFe-Agent Py via sentence-transformers). Drift entre Py e Node
+ * deve ser < 0.01 (gate D.7 validado por tests/integration/test_embedding_parity).
+ *
+ * LRU 128: cobre uso interativo; descarta entradas mais antigas.
+ */
+import { createHash } from "node:crypto";
+import { env, pipeline } from "@xenova/transformers";
+const MODEL_NAME = "Xenova/paraphrase-multilingual-MiniLM-L12-v2";
+const EMBEDDING_DIM = 384;
+const LRU_SIZE = 128;
+export const EMBEDDING_MODEL_NAME = MODEL_NAME;
+export const EMBEDDING_DIMENSION = EMBEDDING_DIM;
+const cache = new Map();
+function hash(text) {
+    return createHash("sha256")
+        .update(text.trim().toLowerCase())
+        .digest("hex");
+}
+let pipelineSingleton = null;
+async function getExtractor() {
+    if (pipelineSingleton)
+        return pipelineSingleton;
+    env.allowLocalModels = false;
+    env.useFS = false;
+    pipelineSingleton = await pipeline("feature-extraction", MODEL_NAME);
+    return pipelineSingleton;
+}
+/**
+ * Codifica `text` em vetor 384-d (cosine-normalized).
+ *
+ * Cache LRU: 2a chamada com mesmo texto retorna o mesmo Float32Array sem
+ * nova inferencia.
+ *
+ * @throws se o modelo nao conseguir carregar (problema conhecido: Node
+ *         v22.21.1 com sharp/better-sqlite3 nativos; ver Sprint 13).
+ */
+export async function encode(text) {
+    const key = hash(text);
+    const cached = cache.get(key);
+    if (cached) {
+        // Move-to-end (LRU)
+        cache.delete(key);
+        cache.set(key, cached);
+        return cached;
+    }
+    const extractor = await getExtractor();
+    const out = await extractor(text, { pooling: "mean", normalize: true });
+    const vec = new Float32Array(out.data);
+    cache.set(key, vec);
+    if (cache.size > LRU_SIZE) {
+        const firstKey = cache.keys().next().value;
+        if (firstKey !== undefined)
+            cache.delete(firstKey);
+    }
+    return vec;
+}
+export function clearCache() {
+    cache.clear();
+}
+export function cacheSize() {
+    return cache.size;
+}
+//# sourceMappingURL=embedder.js.map

package/dist/query/embedder.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"embedder.js","sourceRoot":"","sources":["../../src/query/embedder.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AACzC,OAAO,EAAE,GAAG,EAAE,QAAQ,EAAE,MAAM,sBAAsB,CAAC;AAErD,MAAM,UAAU,GAAG,8CAA8C,CAAC;AAClE,MAAM,aAAa,GAAG,GAAG,CAAC;AAC1B,MAAM,QAAQ,GAAG,GAAG,CAAC;AAErB,MAAM,CAAC,MAAM,oBAAoB,GAAG,UAAU,CAAC;AAC/C,MAAM,CAAC,MAAM,mBAAmB,GAAG,aAAa,CAAC;AAOjD,MAAM,KAAK,GAA8B,IAAI,GAAG,EAAE,CAAC;AAEnD,SAAS,IAAI,CAAC,IAAY;IACxB,OAAO,UAAU,CAAC,QAAQ,CAAC;SACxB,MAAM,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,WAAW,EAAE,CAAC;SACjC,MAAM,CAAC,KAAK,CAAC,CAAC;AACnB,CAAC;AAED,IAAI,iBAAiB,GAAe,IAAI,CAAC;AAEzC,KAAK,UAAU,YAAY;IACzB,IAAI,iBAAiB;QAAE,OAAO,iBAAiB,CAAC;IAChD,GAAG,CAAC,gBAAgB,GAAG,KAAK,CAAC;IAC7B,GAAG,CAAC,KAAK,GAAG,KAAK,CAAC;IAClB,iBAAiB,GAAG,MAAM,QAAQ,CAAC,oBAAoB,EAAE,UAAU,CAAC,CAAC;IACrE,OAAO,iBAAiB,CAAC;AAC3B,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,CAAC,KAAK,UAAU,MAAM,CAAC,IAAY;IACvC,MAAM,GAAG,GAAG,IAAI,CAAC,IAAI,CAAC,CAAC;IACvB,MAAM,MAAM,GAAG,KAAK,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;IAC9B,IAAI,MAAM,EAAE,CAAC;QACX,oBAAoB;QACpB,KAAK,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;QAClB,KAAK,CAAC,GAAG,CAAC,GAAG,EAAE,MAAM,CAAC,CAAC;QACvB,OAAO,MAAM,CAAC;IAChB,CAAC;IAED,MAAM,SAAS,GAAG,MAAM,YAAY,EAAE,CAAC;IACvC,MAAM,GAAG,GAAG,MAAM,SAAS,CAAC,IAAI,EAAE,EAAE,OAAO,EAAE,MAAM,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;IACxE,MAAM,GAAG,GAAG,IAAI,YAAY,CAAC,GAAG,CAAC,IAAyB,CAAC,CAAC;IAE5D,KAAK,CAAC,GAAG,CAAC,GAAG,EAAE,GAAG,CAAC,CAAC;IACpB,IAAI,KAAK,CAAC,IAAI,GAAG,QAAQ,EAAE,CAAC;QAC1B,MAAM,QAAQ,GAAG,KAAK,CAAC,IAAI,EAAE,CAAC,IAAI,EAAE,CAAC,KAA2B,CAAC;QACjE,IAAI,QAAQ,KAAK,SAAS;YAAE,KAAK,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;IACrD,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC;AAED,MAAM,UAAU,UAAU;IACxB,KAAK,CAAC,KAAK,EAAE,CAAC;AAChB,CAAC;AAED,MAAM,UAAU,SAAS;IACvB,OAAO,KAAK,CAAC,IAAI,CAAC;AACpB,CAAC"}

package/dist/query/ftsSearch.d.ts

@@ -0,0 +1,28 @@
+/**
+ * ftsSearch.ts — busca lexical via FTS5 (BM25).
+ *
+ * @see PLAN_SPRINT14.md Task E.2
+ *
+ * Estrategia:
+ - saneamento da query (escape chars que quebram MATCH)
+ - MATCH na tabela fts_chunks (BM25 nativo)
+ - ORDER BY bm25(fts_chunks) LIMIT k
+ *
+ * BM25 score: menor = mais relevante (FTS5 inverte o sinal).
+ */
+import type { Database as BetterSqlite3Database } from "better-sqlite3";
+export interface FtsHit {
+    chunk_id: number;
+    doc_id: number;
+    score: number;
+}
+/**
+ * Busca top-K chunks por BM25. Retorna hits ordenados do mais relevante
+ * (menor score BM25) ao menos relevante.
+ *
+ * @param handle Better-sqlite3 handle (read-only)
+ * @param query Query em linguagem natural
+ * @param k Numero de hits
+ */
+export declare function ftsSearch(handle: BetterSqlite3Database, query: string, k: number): FtsHit[];
+//# sourceMappingURL=ftsSearch.d.ts.map

package/dist/query/ftsSearch.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"ftsSearch.d.ts","sourceRoot":"","sources":["../../src/query/ftsSearch.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,OAAO,KAAK,EAAE,QAAQ,IAAI,qBAAqB,EAAE,MAAM,gBAAgB,CAAC;AAExE,MAAM,WAAW,MAAM;IACrB,QAAQ,EAAE,MAAM,CAAC;IACjB,MAAM,EAAE,MAAM,CAAC;IACf,KAAK,EAAE,MAAM,CAAC;CACf;AAaD;;;;;;;GAOG;AACH,wBAAgB,SAAS,CACvB,MAAM,EAAE,qBAAqB,EAC7B,KAAK,EAAE,MAAM,EACb,CAAC,EAAE,MAAM,GACR,MAAM,EAAE,CAmBV"}

package/dist/query/ftsSearch.js

@@ -0,0 +1,48 @@
+/**
+ * ftsSearch.ts — busca lexical via FTS5 (BM25).
+ *
+ * @see PLAN_SPRINT14.md Task E.2
+ *
+ * Estrategia:
+ - saneamento da query (escape chars que quebram MATCH)
+ - MATCH na tabela fts_chunks (BM25 nativo)
+ - ORDER BY bm25(fts_chunks) LIMIT k
+ *
+ * BM25 score: menor = mais relevante (FTS5 inverte o sinal).
+ */
+/**
+ * Remove caracteres que quebram o MATCH do FTS5. Mantemos alfanumericos,
+ * espacos e aspas (para phrase queries).
+ */
+function sanitizeQuery(q) {
+    return q
+        .replace(/[^\w\s"]/g, " ")
+        .replace(/\s+/g, " ")
+        .trim();
+}
+/**
+ * Busca top-K chunks por BM25. Retorna hits ordenados do mais relevante
+ * (menor score BM25) ao menos relevante.
+ *
+ * @param handle Better-sqlite3 handle (read-only)
+ * @param query Query em linguagem natural
+ * @param k Numero de hits
+ */
+export function ftsSearch(handle, query, k) {
+    const ftsQuery = sanitizeQuery(query);
+    if (!ftsQuery)
+        return [];
+    const rows = handle
+        .prepare(`SELECT chunk_id, doc_id, bm25(fts_chunks) AS score
+         FROM fts_chunks
+        WHERE fts_chunks MATCH ?
+        ORDER BY score
+        LIMIT ?`)
+        .all(ftsQuery, k);
+    return rows.map((r) => ({
+        chunk_id: r.chunk_id,
+        doc_id: r.doc_id,
+        score: r.score,
+    }));
+}
+//# sourceMappingURL=ftsSearch.js.map

package/dist/query/ftsSearch.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"ftsSearch.js","sourceRoot":"","sources":["../../src/query/ftsSearch.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAUH;;;GAGG;AACH,SAAS,aAAa,CAAC,CAAS;IAC9B,OAAO,CAAC;SACL,OAAO,CAAC,WAAW,EAAE,GAAG,CAAC;SACzB,OAAO,CAAC,MAAM,EAAE,GAAG,CAAC;SACpB,IAAI,EAAE,CAAC;AACZ,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,SAAS,CACvB,MAA6B,EAC7B,KAAa,EACb,CAAS;IAET,MAAM,QAAQ,GAAG,aAAa,CAAC,KAAK,CAAC,CAAC;IACtC,IAAI,CAAC,QAAQ;QAAE,OAAO,EAAE,CAAC;IAEzB,MAAM,IAAI,GAAG,MAAM;SAChB,OAAO,CACN;;;;gBAIU,CACX;SACA,GAAG,CAAC,QAAQ,EAAE,CAAC,CAA+D,CAAC;IAElF,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;QACtB,QAAQ,EAAE,CAAC,CAAC,QAAQ;QACpB,MAAM,EAAE,CAAC,CAAC,MAAM;QAChB,KAAK,EAAE,CAAC,CAAC,KAAK;KACf,CAAC,CAAC,CAAC;AACN,CAAC"}

package/dist/query/hybrid.d.ts

@@ -0,0 +1,31 @@
+/**
+ * hybrid.ts — Reciprocal Rank Fusion (RRF) entre vector search e FTS5.
+ *
+ * @see PLAN_SPRINT14.md Task E.3
+ *
+ * Algoritmo RRF canonico (Cormack et al., 2009):
+ *   rrf_score(d) = sum_modes (1 / (k0 + rank_m(d)))
+ * Onde k0 = 60 e rank_m(d) e' o rank do doc `d` no modo `m` (1-indexed).
+ *
+ * Documentos que aparecem em ambos os modos (vector + fts) recebem score
+ * somado, naturalmente subindo no ranking final.
+ */
+import type { VectorHit } from "./vectorSearch.js";
+import type { FtsHit } from "./ftsSearch.js";
+export interface HybridHit {
+    chunk_id: number;
+    doc_id: number;
+    score: number;
+    sources: Array<"vector" | "fts">;
+}
+/**
+ * Fonde resultados de vector + FTS via RRF.
+ *
+ * @param vectorHits Top-K do vector search (ordenado por distance)
+ * @param ftsHits Top-K do FTS5 (ordenado por bm25)
+ * @param k Tamanho final do ranking
+ * @returns Array de hits ordenados por score decrescente
+ */
+export declare function rrf(vectorHits: VectorHit[], ftsHits: FtsHit[], k: number, k0?: number): HybridHit[];
+export declare const RRF_K0_CANONICAL = 60;
+//# sourceMappingURL=hybrid.d.ts.map

package/dist/query/hybrid.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"hybrid.d.ts","sourceRoot":"","sources":["../../src/query/hybrid.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,mBAAmB,CAAC;AACnD,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,gBAAgB,CAAC;AAE7C,MAAM,WAAW,SAAS;IACxB,QAAQ,EAAE,MAAM,CAAC;IACjB,MAAM,EAAE,MAAM,CAAC;IACf,KAAK,EAAE,MAAM,CAAC;IACd,OAAO,EAAE,KAAK,CAAC,QAAQ,GAAG,KAAK,CAAC,CAAC;CAClC;AAID;;;;;;;GAOG;AACH,wBAAgB,GAAG,CACjB,UAAU,EAAE,SAAS,EAAE,EACvB,OAAO,EAAE,MAAM,EAAE,EACjB,CAAC,EAAE,MAAM,EACT,EAAE,GAAE,MAAe,GAClB,SAAS,EAAE,CA4Cb;AAED,eAAO,MAAM,gBAAgB,KAAS,CAAC"}

package/dist/query/hybrid.js

@@ -0,0 +1,67 @@
+/**
+ * hybrid.ts — Reciprocal Rank Fusion (RRF) entre vector search e FTS5.
+ *
+ * @see PLAN_SPRINT14.md Task E.3
+ *
+ * Algoritmo RRF canonico (Cormack et al., 2009):
+ *   rrf_score(d) = sum_modes (1 / (k0 + rank_m(d)))
+ * Onde k0 = 60 e rank_m(d) e' o rank do doc `d` no modo `m` (1-indexed).
+ *
+ * Documentos que aparecem em ambos os modos (vector + fts) recebem score
+ * somado, naturalmente subindo no ranking final.
+ */
+const RRF_K0 = 60;
+/**
+ * Fonde resultados de vector + FTS via RRF.
+ *
+ * @param vectorHits Top-K do vector search (ordenado por distance)
+ * @param ftsHits Top-K do FTS5 (ordenado por bm25)
+ * @param k Tamanho final do ranking
+ * @returns Array de hits ordenados por score decrescente
+ */
+export function rrf(vectorHits, ftsHits, k, k0 = RRF_K0) {
+    const scoreMap = new Map();
+    vectorHits.forEach((hit, i) => {
+        const existing = scoreMap.get(hit.chunk_id);
+        const rankScore = 1 / (k0 + i + 1);
+        if (existing) {
+            existing.score += rankScore;
+            existing.sources.add("vector");
+        }
+        else {
+            scoreMap.set(hit.chunk_id, {
+                chunk_id: hit.chunk_id,
+                doc_id: hit.doc_id,
+                score: rankScore,
+                sources: new Set(["vector"]),
+            });
+        }
+    });
+    ftsHits.forEach((hit, i) => {
+        const existing = scoreMap.get(hit.chunk_id);
+        const rankScore = 1 / (k0 + i + 1);
+        if (existing) {
+            existing.score += rankScore;
+            existing.sources.add("fts");
+        }
+        else {
+            scoreMap.set(hit.chunk_id, {
+                chunk_id: hit.chunk_id,
+                doc_id: hit.doc_id,
+                score: rankScore,
+                sources: new Set(["fts"]),
+            });
+        }
+    });
+    return Array.from(scoreMap.values())
+        .sort((a, b) => b.score - a.score)
+        .slice(0, k)
+        .map((entry) => ({
+        chunk_id: entry.chunk_id,
+        doc_id: entry.doc_id,
+        score: entry.score,
+        sources: Array.from(entry.sources),
+    }));
+}
+export const RRF_K0_CANONICAL = RRF_K0;
+//# sourceMappingURL=hybrid.js.map

package/dist/query/hybrid.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"hybrid.js","sourceRoot":"","sources":["../../src/query/hybrid.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAYH,MAAM,MAAM,GAAG,EAAE,CAAC;AAElB;;;;;;;GAOG;AACH,MAAM,UAAU,GAAG,CACjB,UAAuB,EACvB,OAAiB,EACjB,CAAS,EACT,KAAa,MAAM;IAEnB,MAAM,QAAQ,GAAG,IAAI,GAAG,EAA+F,CAAC;IAExH,UAAU,CAAC,OAAO,CAAC,CAAC,GAAG,EAAE,CAAC,EAAE,EAAE;QAC5B,MAAM,QAAQ,GAAG,QAAQ,CAAC,GAAG,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC;QAC5C,MAAM,SAAS,GAAG,CAAC,GAAG,CAAC,EAAE,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;QACnC,IAAI,QAAQ,EAAE,CAAC;YACb,QAAQ,CAAC,KAAK,IAAI,SAAS,CAAC;YAC5B,QAAQ,CAAC,OAAO,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC;QACjC,CAAC;aAAM,CAAC;YACN,QAAQ,CAAC,GAAG,CAAC,GAAG,CAAC,QAAQ,EAAE;gBACzB,QAAQ,EAAE,GAAG,CAAC,QAAQ;gBACtB,MAAM,EAAE,GAAG,CAAC,MAAM;gBAClB,KAAK,EAAE,SAAS;gBAChB,OAAO,EAAE,IAAI,GAAG,CAAmB,CAAC,QAAQ,CAAC,CAAC;aAC/C,CAAC,CAAC;QACL,CAAC;IACH,CAAC,CAAC,CAAC;IAEH,OAAO,CAAC,OAAO,CAAC,CAAC,GAAG,EAAE,CAAC,EAAE,EAAE;QACzB,MAAM,QAAQ,GAAG,QAAQ,CAAC,GAAG,CAAC,GAAG,CAAC,QAAQ,CAAC,CAAC;QAC5C,MAAM,SAAS,GAAG,CAAC,GAAG,CAAC,EAAE,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC;QACnC,IAAI,QAAQ,EAAE,CAAC;YACb,QAAQ,CAAC,KAAK,IAAI,SAAS,CAAC;YAC5B,QAAQ,CAAC,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC;QAC9B,CAAC;aAAM,CAAC;YACN,QAAQ,CAAC,GAAG,CAAC,GAAG,CAAC,QAAQ,EAAE;gBACzB,QAAQ,EAAE,GAAG,CAAC,QAAQ;gBACtB,MAAM,EAAE,GAAG,CAAC,MAAM;gBAClB,KAAK,EAAE,SAAS;gBAChB,OAAO,EAAE,IAAI,GAAG,CAAmB,CAAC,KAAK,CAAC,CAAC;aAC5C,CAAC,CAAC;QACL,CAAC;IACH,CAAC,CAAC,CAAC;IAEH,OAAO,KAAK,CAAC,IAAI,CAAC,QAAQ,CAAC,MAAM,EAAE,CAAC;SACjC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC;SACjC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC;SACX,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC;QACf,QAAQ,EAAE,KAAK,CAAC,QAAQ;QACxB,MAAM,EAAE,KAAK,CAAC,MAAM;QACpB,KAAK,EAAE,KAAK,CAAC,KAAK;QAClB,OAAO,EAAE,KAAK,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC;KACnC,CAAC,CAAC,CAAC;AACR,CAAC;AAED,MAAM,CAAC,MAAM,gBAAgB,GAAG,MAAM,CAAC"}

package/dist/query/index.d.ts

@@ -0,0 +1,45 @@
+/**
+ * query/index.ts — query engine Node (port de src/query/query_engine.py).
+ *
+ * @see PLAN_SPRINT14.md Fase E (5 tasks)
+ * @see src/query/query_engine.py — equivalente Py
+ *
+ * Modos opt-in (mesmo do CLI Py):
+ *   semantic    vector search com sqlite-vec (default)
+ *   fts         FTS5 BM25 lexical
+ *   hybrid      RRF k=60 entre vector + FTS5
+ *
+ * Constante NO_EVIDENCE_MESSAGE canonica — mesma string em
+ * src/query/context_builder.py do Py (gate dfe-rules.md #4).
+ *   NUNCA duplique este literal.
+ */
+import { EMBEDDING_MODEL_NAME } from "./embedder.js";
+import { hasSufficientEvidence } from "./contextBuilder.js";
+import { NO_EVIDENCE_MESSAGE, SUPPORTED_MODES, DEFAULT_TOP_K, MIN_RELEVANCE_SCORE, type SearchMode } from "./constants.js";
+export { NO_EVIDENCE_MESSAGE, SUPPORTED_MODES, DEFAULT_TOP_K };
+export type { SearchMode };
+export interface Source {
+    url: string;
+    title: string;
+    score: number;
+}
+export interface SearchResult {
+    answer: string;
+    sources: Source[];
+}
+export interface SearchOptions {
+    mode?: SearchMode;
+    topK?: number;
+    baseDirOverride?: string;
+}
+/**
+ * Resolve path da base RAG. Default: $DFE_AGENT_BASE_DIR ou ~/.dfe-agent.
+ */
+export declare function resolveBaseDir(override?: string): string;
+/**
+ * Busca principal. Retorna {answer, sources[]} no contrato canonico.
+ */
+export declare function search(question: string, opts?: SearchOptions): Promise<SearchResult>;
+export { EMBEDDING_MODEL_NAME, MIN_RELEVANCE_SCORE };
+export { hasSufficientEvidence };
+//# sourceMappingURL=index.d.ts.map

package/dist/query/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/query/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAMH,OAAO,EAAU,oBAAoB,EAAE,MAAM,eAAe,CAAC;AAK7D,OAAO,EAEL,qBAAqB,EAEtB,MAAM,qBAAqB,CAAC;AAC7B,OAAO,EACL,mBAAmB,EACnB,eAAe,EACf,aAAa,EACb,mBAAmB,EAGnB,KAAK,UAAU,EAChB,MAAM,gBAAgB,CAAC;AAExB,OAAO,EAAE,mBAAmB,EAAE,eAAe,EAAE,aAAa,EAAE,CAAC;AAC/D,YAAY,EAAE,UAAU,EAAE,CAAC;AAE3B,MAAM,WAAW,MAAM;IACrB,GAAG,EAAE,MAAM,CAAC;IACZ,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;CACf;AAED,MAAM,WAAW,YAAY;IAC3B,MAAM,EAAE,MAAM,CAAC;IACf,OAAO,EAAE,MAAM,EAAE,CAAC;CACnB;AAED,MAAM,WAAW,aAAa;IAC5B,IAAI,CAAC,EAAE,UAAU,CAAC;IAClB,IAAI,CAAC,EAAE,MAAM,CAAC;IACd,eAAe,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED;;GAEG;AACH,wBAAgB,cAAc,CAAC,QAAQ,CAAC,EAAE,MAAM,GAAG,MAAM,CAIxD;AAmED;;GAEG;AACH,wBAAsB,MAAM,CAC1B,QAAQ,EAAE,MAAM,EAChB,IAAI,GAAE,aAAkB,GACvB,OAAO,CAAC,YAAY,CAAC,CAmEvB;AAED,OAAO,EAAE,oBAAoB,EAAE,mBAAmB,EAAE,CAAC;AACrD,OAAO,EAAE,qBAAqB,EAAE,CAAC"}

package/dist/query/index.js

@@ -0,0 +1,153 @@
+/**
+ * query/index.ts — query engine Node (port de src/query/query_engine.py).
+ *
+ * @see PLAN_SPRINT14.md Fase E (5 tasks)
+ * @see src/query/query_engine.py — equivalente Py
+ *
+ * Modos opt-in (mesmo do CLI Py):
+ *   semantic    vector search com sqlite-vec (default)
+ *   fts         FTS5 BM25 lexical
+ *   hybrid      RRF k=60 entre vector + FTS5
+ *
+ * Constante NO_EVIDENCE_MESSAGE canonica — mesma string em
+ * src/query/context_builder.py do Py (gate dfe-rules.md #4).
+ *   NUNCA duplique este literal.
+ */
+import Database from "better-sqlite3";
+import { homedir, tmpdir } from "node:os";
+import { resolve } from "node:path";
+import { encode, EMBEDDING_MODEL_NAME } from "./embedder.js";
+import { vectorSearch } from "./vectorSearch.js";
+import { ftsSearch } from "./ftsSearch.js";
+import { rrf } from "./hybrid.js";
+import { QueryCache } from "./cache.js";
+import { buildContext, hasSufficientEvidence, } from "./contextBuilder.js";
+import { NO_EVIDENCE_MESSAGE, SUPPORTED_MODES, DEFAULT_TOP_K, MIN_RELEVANCE_SCORE, RECENCY_WEIGHT, RECENCY_HALF_LIFE_DAYS, } from "./constants.js";
+export { NO_EVIDENCE_MESSAGE, SUPPORTED_MODES, DEFAULT_TOP_K };
+/**
+ * Resolve path da base RAG. Default: $DFE_AGENT_BASE_DIR ou ~/.dfe-agent.
+ */
+export function resolveBaseDir(override) {
+    return override
+        ?? process.env.DFE_AGENT_BASE_DIR
+        ?? resolve(process.env.HOME ?? homedir() ?? tmpdir(), ".dfe-agent");
+}
+/**
+ * Recency boost (gate paridade com Py `src/query/query_engine.py:34-36`).
+ *
+ * Decaimento exponencial: doc recente (now) -> 1.0; doc com idade = half_life
+ * -> 0.5; doc 2x half_life -> 0.25; etc. Doc sem published_at -> 0.5 (neutro).
+ */
+function recencyScore(publishedAt, now = Date.now()) {
+    if (!publishedAt)
+        return 0.5;
+    const ts = Date.parse(publishedAt);
+    if (Number.isNaN(ts))
+        return 0.5;
+    const ageDays = (now - ts) / 86_400_000;
+    return Math.pow(0.5, ageDays / RECENCY_HALF_LIFE_DAYS);
+}
+/**
+ * Combina score (semantico/fts/rrf) com recencia via peso RECENCY_WEIGHT.
+ * Espelha Py `_apply_recency_and_topk` em query_engine.py:288-291.
+ */
+function applyRecency(score, publishedAt) {
+    const rec = recencyScore(publishedAt);
+    return (1 - RECENCY_WEIGHT) * score + RECENCY_WEIGHT * rec;
+}
+function hydrateChunks(handle, hits) {
+    if (hits.length === 0)
+        return [];
+    const ids = hits.map((h) => h.chunk_id);
+    const placeholders = ids.map(() => "?").join(",");
+    const rows = handle
+        .prepare(`SELECT c.id AS chunk_id, c.doc_id, c.text, d.url, d.title, d.published_at
+         FROM chunks c
+         JOIN documents d ON c.doc_id = d.id
+        WHERE c.id IN (${placeholders})`)
+        .all(...ids);
+    const byId = new Map(rows.map((r) => [r.chunk_id, r]));
+    return hits
+        .map((h) => {
+        const r = byId.get(h.chunk_id);
+        if (!r)
+            return null;
+        return {
+            chunk_id: r.chunk_id,
+            doc_id: r.doc_id,
+            text: r.text,
+            url: r.url,
+            title: r.title,
+            published_at: r.published_at,
+            score: h.score,
+        };
+    })
+        .filter((c) => c !== null);
+}
+/**
+ * Busca principal. Retorna {answer, sources[]} no contrato canonico.
+ */
+export async function search(question, opts = {}) {
+    const mode = (opts.mode ?? "semantic");
+    const topK = opts.topK ?? DEFAULT_TOP_K;
+    if (!SUPPORTED_MODES.includes(mode)) {
+        throw new Error(`unsupported mode: ${mode}. Supported: ${SUPPORTED_MODES.join(", ")}`);
+    }
+    if (!question.trim()) {
+        return { answer: NO_EVIDENCE_MESSAGE, sources: [] };
+    }
+    const dbPath = resolve(resolveBaseDir(opts.baseDirOverride), "dfe.db");
+    const handle = new Database(dbPath, { readonly: true });
+    try {
+        // 1. Cache check
+        const cache = new QueryCache(handle);
+        let queryVec = null;
+        if (mode !== "fts") {
+            queryVec = cache.get(mode, question);
+            if (!queryVec) {
+                queryVec = await encode(question);
+                cache.set(mode, question, queryVec);
+            }
+        }
+        // 2. Buscar chunks (score pre-recency)
+        let hits = [];
+        if (mode === "semantic") {
+            const vecHits = vectorSearch(handle, queryVec, topK);
+            // Converter distance (menor = melhor) em score (maior = melhor)
+            hits = vecHits.map((h) => ({ ...h, score: 1 - h.distance }));
+        }
+        else if (mode === "fts") {
+            // FTS5 retorna bm25 (menor = mais relevante); normalizamos para score
+            // crescente via 1/(1+bm25_normalized). Para paridade com Py, mantemos
+            // bm25 relativo (bm25 do chunk / -bm25_max).
+            const ftsHits = ftsSearch(handle, question, topK);
+            // FTS5 bm25() retorna score NEGATIVO (menor = melhor); invertemos.
+            // Exemplo: bm25 = -2.5 (bom) -> score = 2.5; bm25 = -0.5 (ruim) -> score = 0.5.
+            hits = ftsHits.map((h) => ({ ...h, score: -h.score }));
+        }
+        else {
+            const vecHits = vectorSearch(handle, queryVec, topK);
+            const ftsHits = ftsSearch(handle, question, topK);
+            const hybridHits = rrf(vecHits, ftsHits, topK);
+            hits = hybridHits.map((h) => ({ ...h, score: h.score }));
+        }
+        // 3. Hidratar (precisa de published_at para recency)
+        const hydrated = hydrateChunks(handle, hits);
+        // 4. Aplicar recency boost (gate paridade Py query_engine.py:288-291)
+        for (const c of hydrated) {
+            c.score = applyRecency(c.score, c.published_at ?? null);
+        }
+        // 5. Re-ordenar por score final (recency-adjusted)
+        hydrated.sort((a, b) => b.score - a.score);
+        // 6. Cortar para top-K final
+        const topKFinal = hydrated.slice(0, topK);
+        // 7. Montar contexto
+        return buildContext(topKFinal);
+    }
+    finally {
+        handle.close();
+    }
+}
+export { EMBEDDING_MODEL_NAME, MIN_RELEVANCE_SCORE };
+export { hasSufficientEvidence };
+//# sourceMappingURL=index.js.map

package/dist/query/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/query/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,OAAO,QAAQ,MAAM,gBAAgB,CAAC;AACtC,OAAO,EAAE,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AAC1C,OAAO,EAAE,OAAO,EAAE,MAAM,WAAW,CAAC;AAEpC,OAAO,EAAE,MAAM,EAAE,oBAAoB,EAAE,MAAM,eAAe,CAAC;AAC7D,OAAO,EAAE,YAAY,EAAkB,MAAM,mBAAmB,CAAC;AACjE,OAAO,EAAE,SAAS,EAAe,MAAM,gBAAgB,CAAC;AACxD,OAAO,EAAE,GAAG,EAAE,MAAM,aAAa,CAAC;AAClC,OAAO,EAAE,UAAU,EAAE,MAAM,YAAY,CAAC;AACxC,OAAO,EACL,YAAY,EACZ,qBAAqB,GAEtB,MAAM,qBAAqB,CAAC;AAC7B,OAAO,EACL,mBAAmB,EACnB,eAAe,EACf,aAAa,EACb,mBAAmB,EACnB,cAAc,EACd,sBAAsB,GAEvB,MAAM,gBAAgB,CAAC;AAExB,OAAO,EAAE,mBAAmB,EAAE,eAAe,EAAE,aAAa,EAAE,CAAC;AAoB/D;;GAEG;AACH,MAAM,UAAU,cAAc,CAAC,QAAiB;IAC9C,OAAO,QAAQ;WACV,OAAO,CAAC,GAAG,CAAC,kBAAkB;WAC9B,OAAO,CAAC,OAAO,CAAC,GAAG,CAAC,IAAI,IAAI,OAAO,EAAE,IAAI,MAAM,EAAE,EAAE,YAAY,CAAC,CAAC;AACxE,CAAC;AAWD;;;;;GAKG;AACH,SAAS,YAAY,CAAC,WAA0B,EAAE,MAAc,IAAI,CAAC,GAAG,EAAE;IACxE,IAAI,CAAC,WAAW;QAAE,OAAO,GAAG,CAAC;IAC7B,MAAM,EAAE,GAAG,IAAI,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC;IACnC,IAAI,MAAM,CAAC,KAAK,CAAC,EAAE,CAAC;QAAE,OAAO,GAAG,CAAC;IACjC,MAAM,OAAO,GAAG,CAAC,GAAG,GAAG,EAAE,CAAC,GAAG,UAAU,CAAC;IACxC,OAAO,IAAI,CAAC,GAAG,CAAC,GAAG,EAAE,OAAO,GAAG,sBAAsB,CAAC,CAAC;AACzD,CAAC;AAED;;;GAGG;AACH,SAAS,YAAY,CAAC,KAAa,EAAE,WAA0B;IAC7D,MAAM,GAAG,GAAG,YAAY,CAAC,WAAW,CAAC,CAAC;IACtC,OAAO,CAAC,CAAC,GAAG,cAAc,CAAC,GAAG,KAAK,GAAG,cAAc,GAAG,GAAG,CAAC;AAC7D,CAAC;AAED,SAAS,aAAa,CACpB,MAAyB,EACzB,IAAgE;IAEhE,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,CAAC;IACjC,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC;IACxC,MAAM,YAAY,GAAG,GAAG,CAAC,GAAG,CAAC,GAAG,EAAE,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;IAClD,MAAM,IAAI,GAAG,MAAM;SAChB,OAAO,CACN;;;yBAGmB,YAAY,GAAG,CACnC;SACA,GAAG,CAAC,GAAG,GAAG,CAAe,CAAC;IAC7B,MAAM,IAAI,GAAG,IAAI,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;IACvD,OAAO,IAAI;SACR,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE;QACT,MAAM,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC;QAC/B,IAAI,CAAC,CAAC;YAAE,OAAO,IAAI,CAAC;QACpB,OAAO;YACL,QAAQ,EAAE,CAAC,CAAC,QAAQ;YACpB,MAAM,EAAE,CAAC,CAAC,MAAM;YAChB,IAAI,EAAE,CAAC,CAAC,IAAI;YACZ,GAAG,EAAE,CAAC,CAAC,GAAG;YACV,KAAK,EAAE,CAAC,CAAC,KAAK;YACd,YAAY,EAAE,CAAC,CAAC,YAAY;YAC5B,KAAK,EAAE,CAAC,CAAC,KAAK;SACE,CAAC;IACrB,CAAC,CAAC;SACD,MAAM,CAAC,CAAC,CAAC,EAAsB,EAAE,CAAC,CAAC,KAAK,IAAI,CAAC,CAAC;AACnD,CAAC;AAED;;GAEG;AACH,MAAM,CAAC,KAAK,UAAU,MAAM,CAC1B,QAAgB,EAChB,OAAsB,EAAE;IAExB,MAAM,IAAI,GAAe,CAAC,IAAI,CAAC,IAAI,IAAI,UAAU,CAAe,CAAC;IACjE,MAAM,IAAI,GAAG,IAAI,CAAC,IAAI,IAAI,aAAa,CAAC;IAExC,IAAI,CAAC,eAAe,CAAC,QAAQ,CAAC,IAAI,CAAC,EAAE,CAAC;QACpC,MAAM,IAAI,KAAK,CAAC,qBAAqB,IAAI,gBAAgB,eAAe,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACzF,CAAC;IACD,IAAI,CAAC,QAAQ,CAAC,IAAI,EAAE,EAAE,CAAC;QACrB,OAAO,EAAE,MAAM,EAAE,mBAAmB,EAAE,OAAO,EAAE,EAAE,EAAE,CAAC;IACtD,CAAC;IAED,MAAM,MAAM,GAAG,OAAO,CAAC,cAAc,CAAC,IAAI,CAAC,eAAe,CAAC,EAAE,QAAQ,CAAC,CAAC;IACvE,MAAM,MAAM,GAAG,IAAI,QAAQ,CAAC,MAAM,EAAE,EAAE,QAAQ,EAAE,IAAI,EAAE,CAAC,CAAC;IAExD,IAAI,CAAC;QACH,iBAAiB;QACjB,MAAM,KAAK,GAAG,IAAI,UAAU,CAAC,MAAM,CAAC,CAAC;QACrC,IAAI,QAAQ,GAAwB,IAAI,CAAC;QACzC,IAAI,IAAI,KAAK,KAAK,EAAE,CAAC;YACnB,QAAQ,GAAG,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,QAAQ,CAAC,CAAC;YACrC,IAAI,CAAC,QAAQ,EAAE,CAAC;gBACd,QAAQ,GAAG,MAAM,MAAM,CAAC,QAAQ,CAAC,CAAC;gBAClC,KAAK,CAAC,GAAG,CAAC,IAAI,EAAE,QAAQ,EAAE,QAAQ,CAAC,CAAC;YACtC,CAAC;QACH,CAAC;QAED,uCAAuC;QACvC,IAAI,IAAI,GAA+D,EAAE,CAAC;QAE1E,IAAI,IAAI,KAAK,UAAU,EAAE,CAAC;YACxB,MAAM,OAAO,GAAgB,YAAY,CAAC,MAAM,EAAE,QAAS,EAAE,IAAI,CAAC,CAAC;YACnE,gEAAgE;YAChE,IAAI,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,KAAK,EAAE,CAAC,GAAG,CAAC,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC;QAC/D,CAAC;aAAM,IAAI,IAAI,KAAK,KAAK,EAAE,CAAC;YAC1B,sEAAsE;YACtE,sEAAsE;YACtE,6CAA6C;YAC7C,MAAM,OAAO,GAAa,SAAS,CAAC,MAAM,EAAE,QAAQ,EAAE,IAAI,CAAC,CAAC;YAC5D,mEAAmE;YACnE,gFAAgF;YAChF,IAAI,GAAG,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC;QACzD,CAAC;aAAM,CAAC;YACN,MAAM,OAAO,GAAG,YAAY,CAAC,MAAM,EAAE,QAAS,EAAE,IAAI,CAAC,CAAC;YACtD,MAAM,OAAO,GAAG,SAAS,CAAC,MAAM,EAAE,QAAQ,EAAE,IAAI,CAAC,CAAC;YAClD,MAAM,UAAU,GAAG,GAAG,CAAC,OAAO,EAAE,OAAO,EAAE,IAAI,CAAC,CAAC;YAC/C,IAAI,GAAG,UAAU,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,GAAG,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC,CAAC;QAC3D,CAAC;QAED,qDAAqD;QACrD,MAAM,QAAQ,GAAG,aAAa,CAAC,MAAM,EAAE,IAAI,CAAC,CAAC;QAE7C,sEAAsE;QACtE,KAAK,MAAM,CAAC,IAAI,QAAQ,EAAE,CAAC;YACzB,CAAC,CAAC,KAAK,GAAG,YAAY,CAAC,CAAC,CAAC,KAAK,EAAE,CAAC,CAAC,YAAY,IAAI,IAAI,CAAC,CAAC;QAC1D,CAAC;QAED,mDAAmD;QACnD,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,KAAK,GAAG,CAAC,CAAC,KAAK,CAAC,CAAC;QAE3C,6BAA6B;QAC7B,MAAM,SAAS,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAC,EAAE,IAAI,CAAC,CAAC;QAE1C,qBAAqB;QACrB,OAAO,YAAY,CAAC,SAAS,CAAC,CAAC;IACjC,CAAC;YAAS,CAAC;QACT,MAAM,CAAC,KAAK,EAAE,CAAC;IACjB,CAAC;AACH,CAAC;AAED,OAAO,EAAE,oBAAoB,EAAE,mBAAmB,EAAE,CAAC;AACrD,OAAO,EAAE,qBAAqB,EAAE,CAAC"}

package/dist/query/vectorSearch.d.ts

@@ -0,0 +1,32 @@
+/**
+ * vectorSearch.ts — busca semantica via sqlite-vec (vec_chunks).
+ *
+ * @see PLAN_SPRINT14.md Task E.1
+ *
+ * Estrategia:
+ - embedding da query (384-d, cosine-normalized)
+ - MATCH na tabela virtual vec_chunks (cosine distance)
+ - ORDER BY distance LIMIT k
+ - dedup por doc_id (mantem o chunk mais proximo por documento)
+ - boost temporal opcional (nao implementado aqui; ver E.5 contextBuilder)
+ */
+import type { Database as BetterSqlite3Database } from "better-sqlite3";
+export interface VectorHit {
+    chunk_id: number;
+    doc_id: number;
+    distance: number;
+}
+/**
+ * Carrega extensao sqlite-vec no handle. Idempotente.
+ */
+export declare function loadVecExtension(handle: BetterSqlite3Database): void;
+/**
+ * Busca top-K chunks via cosine distance. Retorna hits ordenados do mais
+ * similar (menor distance) ao menos similar.
+ *
+ * @param handle Better-sqlite3 handle aberto em modo readonly
+ * @param query_vec Vetor 384-d da query (cosine-normalized)
+ * @param k Numero de hits antes do dedup
+ */
+export declare function vectorSearch(handle: BetterSqlite3Database, query_vec: Float32Array, k: number): VectorHit[];
+//# sourceMappingURL=vectorSearch.d.ts.map

package/dist/query/vectorSearch.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"vectorSearch.d.ts","sourceRoot":"","sources":["../../src/query/vectorSearch.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,OAAO,KAAK,EAAE,QAAQ,IAAI,qBAAqB,EAAE,MAAM,gBAAgB,CAAC;AAGxE,MAAM,WAAW,SAAS;IACxB,QAAQ,EAAE,MAAM,CAAC;IACjB,MAAM,EAAE,MAAM,CAAC;IACf,QAAQ,EAAE,MAAM,CAAC;CAClB;AAED;;GAEG;AACH,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,qBAAqB,GAAG,IAAI,CAEpE;AAED;;;;;;;GAOG;AACH,wBAAgB,YAAY,CAC1B,MAAM,EAAE,qBAAqB,EAC7B,SAAS,EAAE,YAAY,EACvB,CAAC,EAAE,MAAM,GACR,SAAS,EAAE,CA4Bb"}

package/dist/query/vectorSearch.js

@@ -0,0 +1,54 @@
+/**
+ * vectorSearch.ts — busca semantica via sqlite-vec (vec_chunks).
+ *
+ * @see PLAN_SPRINT14.md Task E.1
+ *
+ * Estrategia:
+ - embedding da query (384-d, cosine-normalized)
+ - MATCH na tabela virtual vec_chunks (cosine distance)
+ - ORDER BY distance LIMIT k
+ - dedup por doc_id (mantem o chunk mais proximo por documento)
+ - boost temporal opcional (nao implementado aqui; ver E.5 contextBuilder)
+ */
+import * as sqliteVec from "sqlite-vec";
+/**
+ * Carrega extensao sqlite-vec no handle. Idempotente.
+ */
+export function loadVecExtension(handle) {
+    sqliteVec.load(handle);
+}
+/**
+ * Busca top-K chunks via cosine distance. Retorna hits ordenados do mais
+ * similar (menor distance) ao menos similar.
+ *
+ * @param handle Better-sqlite3 handle aberto em modo readonly
+ * @param query_vec Vetor 384-d da query (cosine-normalized)
+ * @param k Numero de hits antes do dedup
+ */
+export function vectorSearch(handle, query_vec, k) {
+    loadVecExtension(handle);
+    // sqlite-vec usa serializacao nativa; precisamos passar Buffer
+    const vecBuf = Buffer.from(query_vec.buffer);
+    const rows = handle
+        .prepare(`SELECT chunk_id, doc_id, distance
+         FROM vec_chunks
+        WHERE embedding MATCH ?
+        ORDER BY distance
+        LIMIT ?`)
+        .all(vecBuf, k);
+    // Dedup por doc_id (mantem o primeiro = menor distance)
+    const seen = new Set();
+    const deduped = [];
+    for (const r of rows) {
+        if (seen.has(r.doc_id))
+            continue;
+        seen.add(r.doc_id);
+        deduped.push({
+            chunk_id: r.chunk_id,
+            doc_id: r.doc_id,
+            distance: r.distance,
+        });
+    }
+    return deduped;
+}
+//# sourceMappingURL=vectorSearch.js.map

package/dist/query/vectorSearch.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"vectorSearch.js","sourceRoot":"","sources":["../../src/query/vectorSearch.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAGH,OAAO,KAAK,SAAS,MAAM,YAAY,CAAC;AAQxC;;GAEG;AACH,MAAM,UAAU,gBAAgB,CAAC,MAA6B;IAC5D,SAAS,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;AACzB,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,YAAY,CAC1B,MAA6B,EAC7B,SAAuB,EACvB,CAAS;IAET,gBAAgB,CAAC,MAAM,CAAC,CAAC;IAEzB,+DAA+D;IAC/D,MAAM,MAAM,GAAG,MAAM,CAAC,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,CAAC;IAC7C,MAAM,IAAI,GAAG,MAAM;SAChB,OAAO,CACN;;;;gBAIU,CACX;SACA,GAAG,CAAC,MAAM,EAAE,CAAC,CAAkE,CAAC;IAEnF,wDAAwD;IACxD,MAAM,IAAI,GAAG,IAAI,GAAG,EAAU,CAAC;IAC/B,MAAM,OAAO,GAAgB,EAAE,CAAC;IAChC,KAAK,MAAM,CAAC,IAAI,IAAI,EAAE,CAAC;QACrB,IAAI,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC;YAAE,SAAS;QACjC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC;QACnB,OAAO,CAAC,IAAI,CAAC;YACX,QAAQ,EAAE,CAAC,CAAC,QAAQ;YACpB,MAAM,EAAE,CAAC,CAAC,MAAM;YAChB,QAAQ,EAAE,CAAC,CAAC,QAAQ;SACrB,CAAC,CAAC;IACL,CAAC;IACD,OAAO,OAAO,CAAC;AACjB,CAAC"}