@vellumai/assistant 0.7.2 → 0.8.0

package/src/memory/v2/qdrant.ts

@@ -48,20 +48,43 @@ export interface ConceptPagePayload {
 export interface ConceptPageQueryResult {
   slug: string;
   /**
-   * Dense cosine similarity, when the slug appeared in the dense top-`limit`.
-   * `undefined` if the slug only appeared in the sparse channel.
+   * Dense cosine similarity against the page body, when the slug appeared in
+   * the body dense top-`limit`. `undefined` if the slug only appeared in the
+   * sparse channel — or in a summary-side channel.
    */
   denseScore?: number;
   /**
-   * Sparse score, when the slug appeared in the sparse top-`limit`.
-   * `undefined` if the slug only appeared in the dense channel. Lives on a
-   * different scale than `denseScore` — callers must normalize before fusing.
+   * Sparse score against the page body, when the slug appeared in the body
+   * sparse top-`limit`. `undefined` if the slug only appeared in the dense
+   * channel. Lives on a different scale than `denseScore` — callers must
+   * normalize before fusing.
    */
   sparseScore?: number;
+  /**
+   * Dense cosine similarity against the page's frontmatter `summary`, when
+   * the page has a summary embedded and the slug appeared in the summary
+   * dense top-`limit`. `undefined` for pages without a summary embedding —
+   * those fall back to body-only scoring.
+   */
+  summaryDenseScore?: number;
+  /**
+   * Sparse score against the page's frontmatter `summary`, paired with
+   * `summaryDenseScore`. `undefined` for pages without a summary embedding.
+   */
+  summarySparseScore?: number;
 }
 
 let _client: QdrantRestClient | null = null;
 let _collectionReady = false;
+let _collectionReadyPromise: Promise<{ migrated: boolean }> | null = null;
+
+/**
+ * Named vectors the v2 concept-page collection must expose. Existing
+ * collections that lack any of these get destructively recreated by
+ * `ensureConceptPageCollectionOnce` — see the `migrated` return flag.
+ */
+const REQUIRED_DENSE_VECTORS = ["dense", "summary_dense"] as const;
+const REQUIRED_SPARSE_VECTORS = ["sparse", "summary_sparse"] as const;
 
 /** Lazily create a Qdrant REST client bound to the resolved URL. */
 function getClient(): QdrantRestClient {
@@ -75,27 +98,67 @@ function getClient(): QdrantRestClient {
 }
 
 /**
- * Create the v2 concept-page collection if it does not already exist.
- * Idempotent: a no-op when the collection is already present.
- *
- * Vector layout mirrors `VellumQdrantClient.ensureCollection` — named dense
- * (cosine, configurable size + on-disk) and sparse vectors. The vector size
- * and on-disk flag inherit from `config.memory.qdrant` so v2 stays aligned
- * with the user's existing embedding backend without separate knobs.
+ * Create the v2 concept-page collection if it does not already exist, or
+ * destructively recreate it when the existing schema is missing any of the
+ * required named vectors (see `REQUIRED_DENSE_VECTORS` /
+ * `REQUIRED_SPARSE_VECTORS`). The latter case is signalled to callers via
+ * `{ migrated: true }` so they can enqueue a backfill — pre-#29823
+ * collections lack `summary_dense` / `summary_sparse` and every query
+ * referencing those named vectors fails with HTTP 400 until the collection
+ * is rebuilt. Mirrors `VellumQdrantClient.ensureCollection` for v1.
  */
-export async function ensureConceptPageCollection(): Promise<void> {
-  if (_collectionReady) return;
+export async function ensureConceptPageCollection(): Promise<{
+  migrated: boolean;
+}> {
+  if (_collectionReady) return { migrated: false };
+  if (_collectionReadyPromise) return _collectionReadyPromise;
+
+  _collectionReadyPromise = ensureConceptPageCollectionOnce().finally(() => {
+    _collectionReadyPromise = null;
+  });
+  return _collectionReadyPromise;
+}
 
+async function ensureConceptPageCollectionOnce(): Promise<{
+  migrated: boolean;
+}> {
   const client = getClient();
   const config = getConfig();
   const vectorSize = config.memory.qdrant.vectorSize;
   const onDisk = config.memory.qdrant.onDisk;
 
+  let migrated = false;
+
   try {
     const exists = await client.collectionExists(MEMORY_V2_COLLECTION);
     if (exists.exists) {
-      _collectionReady = true;
-      return;
+      // Assume compatible on probe failure rather than risk a destructive
+      // recreate — mirrors v1's posture in `VellumQdrantClient.ensureCollection`.
+      let info: Awaited<ReturnType<typeof client.getCollection>>;
+      try {
+        info = await client.getCollection(MEMORY_V2_COLLECTION);
+      } catch (err) {
+        log.warn(
+          { err, collection: MEMORY_V2_COLLECTION },
+          "Failed to probe v2 collection schema; assuming compatible",
+        );
+        _collectionReady = true;
+        return { migrated: false };
+      }
+
+      const missing = missingNamedVectors(info);
+      if (missing.length === 0) {
+        _collectionReady = true;
+        return { migrated: false };
+      }
+
+      log.warn(
+        { collection: MEMORY_V2_COLLECTION, missingNamedVectors: missing },
+        "Memory v2 concept-page collection schema drift detected — deleting and recreating; embeddings will be regenerated by background reembed",
+      );
+      await client.deleteCollection(MEMORY_V2_COLLECTION);
+      migrated = true;
+      // Fall through to creation below.
     }
   } catch (err) {
     // Treat "not found"-shaped errors as "needs creation" and fall through.
@@ -115,15 +178,28 @@ export async function ensureConceptPageCollection(): Promise<void> {
           distance: "Cosine",
           on_disk: onDisk,
         },
+        // Optional second dense vector covering the page's frontmatter
+        // `summary`. Pages without a summary store nothing under this name —
+        // Qdrant supports per-point named-vector subsets — so the named-vector
+        // index stays cheap until summaries are populated.
+        summary_dense: {
+          size: vectorSize,
+          distance: "Cosine",
+          on_disk: onDisk,
+        },
       },
       sparse_vectors: {
         sparse: {}, // Qdrant auto-infers sparse vector params
+        summary_sparse: {}, // BM25 sparse vector for the summary
       },
       hnsw_config: {
         on_disk: onDisk,
         m: 16,
         ef_construct: 100,
       },
+      optimizers_config: {
+        default_segment_number: 2,
+      },
       on_disk_payload: onDisk,
     });
   } catch (err) {
@@ -134,7 +210,7 @@ export async function ensureConceptPageCollection(): Promise<void> {
       (err as { status: number }).status === 409
     ) {
       _collectionReady = true;
-      return;
+      return { migrated };
     }
     throw err;
   }
@@ -147,32 +223,86 @@ export async function ensureConceptPageCollection(): Promise<void> {
   });
 
   _collectionReady = true;
+  return { migrated };
+}
+
+/**
+ * Return the names of required named vectors absent from the collection's
+ * current schema. An empty array means the collection is fully migrated.
+ *
+ * If the response shape is unparseable (e.g. Qdrant returns an unexpected
+ * structure) we treat it as "everything is missing" so the caller's drift
+ * branch fires — combined with the `getCollection` try/catch in the caller,
+ * a thrown probe falls back to "assume compatible" while a parsed-but-empty
+ * response triggers the safer recreate.
+ */
+function missingNamedVectors(
+  info: Awaited<ReturnType<QdrantRestClient["getCollection"]>>,
+): string[] {
+  const params = info.config?.params;
+  const dense = params?.vectors;
+  const sparse = (params as { sparse_vectors?: unknown } | undefined)
+    ?.sparse_vectors;
+  const denseNames =
+    dense && typeof dense === "object" && !("size" in dense)
+      ? new Set(Object.keys(dense))
+      : new Set<string>();
+  const sparseNames =
+    sparse && typeof sparse === "object"
+      ? new Set(Object.keys(sparse as Record<string, unknown>))
+      : new Set<string>();
+
+  const missing: string[] = [];
+  for (const name of REQUIRED_DENSE_VECTORS) {
+    if (!denseNames.has(name)) missing.push(name);
+  }
+  for (const name of REQUIRED_SPARSE_VECTORS) {
+    if (!sparseNames.has(name)) missing.push(name);
+  }
+  return missing;
 }
 
 /**
  * Upsert a concept page's dense + sparse embedding. The point ID is derived
  * deterministically from the slug so subsequent calls for the same slug
  * replace the prior point in place rather than accumulating duplicates.
+ *
+ * `summary` is optional — supplied when the page's frontmatter carries a
+ * `summary`, omitted otherwise. Pages without a summary store only the body
+ * vectors and fall back to body-only scoring at query time. The grouped
+ * shape enforces at the type level that summary dense and sparse are
+ * always written together.
  */
 export async function upsertConceptPageEmbedding(params: {
   slug: string;
   dense: number[];
   sparse: SparseEmbedding;
+  summary?: { dense: number[]; sparse: SparseEmbedding };
   updatedAt: number;
 }): Promise<void> {
   await ensureConceptPageCollection();
 
-  const { slug, dense, sparse, updatedAt } = params;
+  const { slug, dense, sparse, summary, updatedAt } = params;
   const client = getClient();
   const pointId = pointIdForSlug(slug);
 
+  // Qdrant lets us upsert any subset of named vectors per point. The summary
+  // entries appear only when the caller passed a `summary` block — pairing
+  // them at the type level keeps query-time fusion symmetric with the body
+  // channels.
+  const vector: Record<string, number[] | SparseEmbedding> = { dense, sparse };
+  if (summary) {
+    vector.summary_dense = summary.dense;
+    vector.summary_sparse = summary.sparse;
+  }
+
   const upsertOnce = () =>
     client.upsert(MEMORY_V2_COLLECTION, {
       wait: true,
       points: [
         {
           id: pointId,
-          vector: { dense, sparse },
+          vector,
           payload: { slug, updated_at: updatedAt },
         },
       ],
@@ -215,15 +345,134 @@ export async function deleteConceptPageEmbedding(slug: string): Promise<void> {
   }
 }
 
+/**
+ * Remove every point whose slug starts with the given prefix and whose
+ * remaining suffix is not in `activeSuffixes`. Used by the skill-seed flow to
+ * drop stale `skills/<id>` slugs after a skill is uninstalled or disabled,
+ * since skills now share the concept-page collection rather than living in a
+ * dedicated one.
+ *
+ * Idempotent: when the live `<prefix>*` slugs already match `activeSuffixes`,
+ * the function performs a single scroll and no deletes.
+ */
+export async function pruneSlugsWithPrefixExcept(
+  prefix: string,
+  activeSuffixes: readonly string[],
+): Promise<void> {
+  await ensureConceptPageCollection();
+
+  const client = getClient();
+  const activeSet = new Set(activeSuffixes);
+
+  const doPrune = async (): Promise<void> => {
+    const stalePointIds: Array<string | number> = [];
+    let offset: string | number | undefined = undefined;
+    const maxIterations = 10_000;
+    const batchSize = 256;
+    for (let i = 0; i < maxIterations; i++) {
+      const result = await client.scroll(MEMORY_V2_COLLECTION, {
+        limit: batchSize,
+        with_payload: true,
+        with_vector: false,
+        ...(offset !== undefined ? { offset } : {}),
+      });
+      for (const point of result.points) {
+        const slug = (point.payload as { slug?: unknown } | null)?.slug;
+        if (typeof slug !== "string") continue;
+        if (!slug.startsWith(prefix)) continue;
+        const suffix = slug.slice(prefix.length);
+        if (!activeSet.has(suffix)) {
+          stalePointIds.push(point.id);
+        }
+      }
+      const next = result.next_page_offset;
+      if (next == null) break;
+      offset = typeof next === "string" ? next : (next as number);
+    }
+
+    if (stalePointIds.length === 0) return;
+
+    await client.delete(MEMORY_V2_COLLECTION, {
+      wait: true,
+      points: stalePointIds,
+    });
+  };
+
+  try {
+    await doPrune();
+  } catch (err) {
+    if (isCollectionMissing(err)) {
+      _collectionReady = false;
+      await ensureConceptPageCollection();
+      await doPrune();
+      return;
+    }
+    throw err;
+  }
+}
+
+/**
+ * Approximate count of points in the v2 concept-page collection. Used by the
+ * daemon-startup rebuild hook to detect "collection exists but empty" — the
+ * crash-mid-rebuild recovery case where a prior boot dropped + recreated the
+ * collection but died before reembed completed. Returns `0` if the collection
+ * does not exist or the count call fails (treated as "needs reembed" by the
+ * caller).
+ */
+export async function countConceptPagePoints(): Promise<number> {
+  await ensureConceptPageCollection();
+  try {
+    const result = await getClient().count(MEMORY_V2_COLLECTION, {
+      exact: false,
+    });
+    return result.count;
+  } catch (err) {
+    log.warn(
+      { err, collection: MEMORY_V2_COLLECTION },
+      "Failed to count v2 concept-page collection — treating as empty",
+    );
+    return 0;
+  }
+}
+
+/**
+ * Best-effort delete of the legacy `memory_v2_skills` Qdrant collection. Skill
+ * embeddings now live alongside concept pages in `memory_v2_concept_pages`
+ * under the `skills/<id>` slug prefix, so the dedicated collection is dead
+ * weight on installs upgraded from the split-collection era. Fire-and-forget:
+ * on a fresh install (collection never existed) or a transient Qdrant
+ * unavailable, we log and move on.
+ */
+export async function dropLegacySkillsCollection(): Promise<void> {
+  try {
+    const client = getClient();
+    const exists = await client.collectionExists("memory_v2_skills");
+    if (!exists.exists) return;
+    await client.deleteCollection("memory_v2_skills");
+    log.info("Deleted legacy memory_v2_skills Qdrant collection");
+  } catch (err) {
+    log.warn(
+      { err },
+      "Failed to drop legacy memory_v2_skills collection — non-fatal",
+    );
+  }
+}
+
 /**
  * Run separate dense and sparse queries against the concept-page collection
  * and return per-channel scores per slug. Callers fuse these — typically via
  * a normalized weighted-sum — because RRF would discard the score magnitudes
  * the activation formula needs.
  *
+ * Four channels are queried concurrently: body dense, body sparse, summary
+ * dense, summary sparse. The summary channels only return hits for pages whose
+ * frontmatter carries a `summary` (and therefore stored `summary_dense` /
+ * `summary_sparse` named vectors at upsert time). Pages without a summary
+ * surface body-only scores; callers fall back to body-only fusion for those.
+ *
  * Each channel returns up to `limit` hits. A slug is included in the result
- * if it appears in either channel; the missing channel's score is left
- * `undefined` so callers can detect single-channel matches.
+ * if it appears in any channel; missing channel scores stay `undefined` so
+ * callers can distinguish "no match in this channel" from "match with score 0".
  *
  * `restrictToSlugs`, when provided, filters the search server-side to only
  * those slugs (Qdrant `slug IN [...]` filter). Used by `simBatch` when the
@@ -257,42 +506,51 @@ export async function hybridQueryConceptPages(
   // Qdrant 1.13.x sparse-index crash that we've reproduced in the wild.
   const skipSparse = options?.skipSparse ?? false;
 
-  const denseQuery = () =>
+  const queryDense = (using: string) =>
     client.query(MEMORY_V2_COLLECTION, {
       query: dense,
-      using: "dense",
+      using,
       limit,
       with_payload: true,
       filter,
     });
-  const sparseQuery = () =>
+  const querySparse = (using: string) =>
     client.query(MEMORY_V2_COLLECTION, {
       query: sparse,
-      using: "sparse",
+      using,
       limit,
       with_payload: true,
       filter,
     });
 
-  // Run both queries concurrently — they hit independent named vectors.
-  // When sparse is gated off we still resolve a Promise so the destructuring
-  // below stays uniform; the empty `points: []` matches the shape of a
-  // no-hit Qdrant response.
+  // Run all four channels concurrently — they hit independent named vectors.
+  // When sparse is gated off the sparse channels still resolve a Promise so
+  // the destructuring below stays uniform; the empty `points: []` matches
+  // the shape of a no-hit Qdrant response.
   const emptyResult = {
     points: [] as Array<{ payload?: unknown; score?: number }>,
   };
   const runQueries = async () =>
-    Promise.all([denseQuery(), skipSparse ? emptyResult : sparseQuery()]);
+    Promise.all([
+      queryDense("dense"),
+      skipSparse ? emptyResult : querySparse("sparse"),
+      queryDense("summary_dense"),
+      skipSparse ? emptyResult : querySparse("summary_sparse"),
+    ]);
 
   let denseResults;
   let sparseResults;
+  let summaryDenseResults;
+  let summarySparseResults;
   try {
-    [denseResults, sparseResults] = await runQueries();
+    [denseResults, sparseResults, summaryDenseResults, summarySparseResults] =
+      await runQueries();
   } catch (err) {
     if (isCollectionMissing(err)) {
       _collectionReady = false;
       await ensureConceptPageCollection();
-      [denseResults, sparseResults] = await runQueries();
+      [denseResults, sparseResults, summaryDenseResults, summarySparseResults] =
+        await runQueries();
     } else {
       throw err;
     }
@@ -301,21 +559,22 @@ export async function hybridQueryConceptPages(
   // Merge by slug. Missing-side scores stay undefined so the fuser can tell
   // "no match in this channel" apart from "match with score 0".
   const merged = new Map<string, ConceptPageQueryResult>();
-  for (const point of denseResults.points ?? []) {
-    const slug = (point.payload as { slug?: unknown } | null)?.slug;
-    if (typeof slug !== "string") continue;
-    merged.set(slug, { slug, denseScore: point.score ?? 0 });
-  }
-  for (const point of sparseResults.points ?? []) {
-    const slug = (point.payload as { slug?: unknown } | null)?.slug;
-    if (typeof slug !== "string") continue;
-    const existing = merged.get(slug);
-    if (existing) {
-      existing.sparseScore = point.score ?? 0;
-    } else {
-      merged.set(slug, { slug, sparseScore: point.score ?? 0 });
+  const recordHit = (
+    points: Array<{ payload?: unknown; score?: number }> | undefined,
+    set: (entry: ConceptPageQueryResult, score: number) => void,
+  ): void => {
+    for (const point of points ?? []) {
+      const slug = (point.payload as { slug?: unknown } | null)?.slug;
+      if (typeof slug !== "string") continue;
+      const existing = merged.get(slug) ?? { slug };
+      set(existing, point.score ?? 0);
+      merged.set(slug, existing);
     }
-  }
+  };
+  recordHit(denseResults.points, (e, s) => (e.denseScore = s));
+  recordHit(sparseResults.points, (e, s) => (e.sparseScore = s));
+  recordHit(summaryDenseResults.points, (e, s) => (e.summaryDenseScore = s));
+  recordHit(summarySparseResults.points, (e, s) => (e.summarySparseScore = s));
 
   return Array.from(merged.values());
 }
@@ -437,4 +696,5 @@ function pointIdForSlug(slug: string): string {
 export function _resetMemoryV2QdrantForTests(): void {
   _client = null;
   _collectionReady = false;
+  _collectionReadyPromise = null;
 }

package/src/memory/v2/reranker.ts

@@ -0,0 +1,177 @@
+/** Memory v2 cross-encoder rerank — `(query, page-preview)` pairs scored by a local model. */
+
+import { createHash } from "node:crypto";
+
+import type { AssistantConfig } from "../../config/types.js";
+import { getLogger } from "../../util/logger.js";
+import { getWorkspaceDir } from "../../util/platform.js";
+import { getOrCreateRerankBackend } from "../rerank-local.js";
+import { readPage } from "./page-store.js";
+
+const log = getLogger("memory-v2-reranker");
+
+// ~512-token model context for bge-reranker-base; cap input to bound payload.
+const PASSAGE_CHAR_CAP = 240;
+
+interface CacheEntry {
+  scores: Map<string, number>;
+  ts: number;
+}
+
+const CACHE_TTL_MS = 2 * 60 * 1000;
+const CACHE_MAX_ENTRIES = 64;
+const cache = new Map<string, CacheEntry>();
+
+function cacheKey(query: string, slugs: readonly string[]): string {
+  const sorted = [...slugs].sort().join("\0");
+  return createHash("sha256").update(`${query}\0${sorted}`).digest("hex");
+}
+
+function evictExpired(now: number): void {
+  for (const [k, v] of cache) {
+    if (now - v.ts > CACHE_TTL_MS) cache.delete(k);
+  }
+  if (cache.size > CACHE_MAX_ENTRIES) {
+    const toDrop = cache.size - CACHE_MAX_ENTRIES;
+    let i = 0;
+    for (const k of cache.keys()) {
+      if (i++ >= toDrop) break;
+      cache.delete(k);
+    }
+  }
+}
+
+function buildPassage(slug: string, body: string): string {
+  const trimmed = body.replace(/^\s+/, "");
+  const blank = trimmed.search(/\n\s*\n/);
+  const para = blank === -1 ? trimmed : trimmed.slice(0, blank);
+  const stripped = para.replace(/^#+\s.*\n/, "").trim();
+  const compact = stripped.replace(/\s+/g, " ").slice(0, PASSAGE_CHAR_CAP);
+  return `${slug}\n${compact}`;
+}
+
+/**
+ * Run the cross-encoder over each candidate's first-paragraph preview for
+ * one or more queries against the same candidate set. Returns one
+ * `Map<slug, score>` per query, in the same order as the `queries` array.
+ *
+ * Multi-query batching: the user-channel and assistant-channel queries share
+ * a candidate set per turn, so scoring them in a single tokenizer +
+ * forward-pass call avoids the ONNX-invocation overhead of two serialised
+ * worker round-trips. Cache hits short-circuit per-query independently —
+ * a whitespace-only query yields an empty Map without hitting the backend.
+ *
+ * Failures (worker down, page read errors) yield empty Maps so callers can
+ * fall back to pure fused scores. Per-batch normalisation and boost math
+ * live in `computeOwnActivation`.
+ */
+export async function rerankCandidates(
+  queries: readonly string[],
+  candidates: readonly string[],
+  config: AssistantConfig,
+): Promise<Array<Map<string, number>>> {
+  if (queries.length === 0) return [];
+  if (candidates.length === 0) return queries.map(() => new Map());
+
+  const now = Date.now();
+  evictExpired(now);
+
+  const results: Array<Map<string, number> | null> = queries.map(() => null);
+  const uncachedIndices: number[] = [];
+  for (let i = 0; i < queries.length; i++) {
+    const q = queries[i];
+    if (q.trim().length === 0) {
+      results[i] = new Map();
+      continue;
+    }
+    const key = cacheKey(q, candidates);
+    const cached = cache.get(key);
+    if (cached) {
+      // Refresh insertion order so frequently-hit entries survive eviction.
+      cache.delete(key);
+      cache.set(key, { ...cached, ts: now });
+      results[i] = new Map(cached.scores);
+    } else {
+      uncachedIndices.push(i);
+    }
+  }
+
+  const finalize = (): Array<Map<string, number>> =>
+    results.map((r) => r ?? new Map());
+
+  if (uncachedIndices.length === 0) return finalize();
+
+  const workspaceDir = getWorkspaceDir();
+  const pages = await Promise.all(
+    candidates.map((slug) =>
+      readPage(workspaceDir, slug).catch((err) => {
+        log.debug({ err, slug }, "Reranker skipping page that failed to load");
+        return null;
+      }),
+    ),
+  );
+  const passages: string[] = [];
+  const slugsForPassages: string[] = [];
+  for (let i = 0; i < candidates.length; i++) {
+    const page = pages[i];
+    if (!page) continue;
+    passages.push(buildPassage(candidates[i], page.body));
+    slugsForPassages.push(candidates[i]);
+  }
+
+  if (passages.length === 0) {
+    for (const i of uncachedIndices) results[i] = new Map();
+    return finalize();
+  }
+
+  // One tokenizer + ONNX forward pass over every uncached query × passage
+  // pair. Pairs are laid out query-major: queries[uncached[0]] × passages,
+  // then queries[uncached[1]] × passages, etc.
+  const batchQueries: string[] = [];
+  const batchPassages: string[] = [];
+  for (const qi of uncachedIndices) {
+    const q = queries[qi];
+    for (const p of passages) {
+      batchQueries.push(q);
+      batchPassages.push(p);
+    }
+  }
+
+  const { model, dtype } = config.memory.v2.rerank;
+  let scores: number[];
+  try {
+    const backend = getOrCreateRerankBackend(model, dtype);
+    scores = await backend.score(batchQueries, batchPassages);
+  } catch (err) {
+    log.warn(
+      { err, model, n: batchPassages.length },
+      "Rerank backend failed; falling back to pure fused scores",
+    );
+    for (const i of uncachedIndices) results[i] = new Map();
+    return finalize();
+  }
+
+  for (let j = 0; j < uncachedIndices.length; j++) {
+    const qi = uncachedIndices[j];
+    const offset = j * passages.length;
+    const result = new Map<string, number>();
+    for (let i = 0; i < slugsForPassages.length; i++) {
+      const s = scores[offset + i];
+      if (typeof s !== "number" || Number.isNaN(s)) continue;
+      // sigmoid output should already be in [0, 1]; clamp defensively.
+      result.set(slugsForPassages[i], Math.max(0, Math.min(1, s)));
+    }
+    results[qi] = result;
+    cache.set(cacheKey(queries[qi], candidates), {
+      scores: new Map(result),
+      ts: now,
+    });
+  }
+
+  return finalize();
+}
+
+/** @internal Test-only: clear the LRU cache. */
+export function _resetRerankCacheForTests(): void {
+  cache.clear();
+}