@vellumai/assistant 0.6.6 → 0.7.1

package/src/memory/v2/prompts/sweep.ts

@@ -0,0 +1,56 @@
+/**
+ * Memory v2 — sweep prompt template.
+ *
+ * Body taken from §9 of the design doc (`memoized-spinning-wadler.md`). The
+ * template uses two placeholders that the sweep job substitutes at runtime:
+ *
+ *   - `{{ASSISTANT_NAME}}` — the assistant's display name (from IDENTITY.md
+ *     when available, else "the assistant" so the prompt still parses).
+ *   - `{{USER_NAME}}` — the guardian's display name (from the guardian
+ *     persona when available, else "the user").
+ *
+ * Kept here (under `prompts/`) rather than inlined in `sweep-job.ts` so the
+ * prompt body is reviewable on its own and the job module stays focused on
+ * orchestration. The same convention applies to the consolidation prompt
+ * landing in PR 20.
+ */
+
+/** Sentinel substituted with the assistant's display name at runtime. */
+const ASSISTANT_NAME_PLACEHOLDER = "{{ASSISTANT_NAME}}";
+
+/** Sentinel substituted with the guardian's display name at runtime. */
+const USER_NAME_PLACEHOLDER = "{{USER_NAME}}";
+
+/**
+ * Sweep prompt — body from design doc §9. The model is asked to surface
+ * additional facts/preferences/plans/etc. that should be remembered but
+ * aren't already in the existing buffer. It MUST emit a JSON-shaped tool
+ * call with an `entries` array; the runtime parses the response via the
+ * tool definition declared in `sweep-job.ts`.
+ *
+ * The `existingBuffer` text is appended at the call site (rather than
+ * templated here) so we don't inadvertently expand `{{` inside user buffer
+ * content. Recent messages are likewise appended outside the template.
+ */
+const SWEEP_PROMPT = `You are a background helper for ${ASSISTANT_NAME_PLACEHOLDER}. Read these recent messages between ${ASSISTANT_NAME_PLACEHOLDER} and ${USER_NAME_PLACEHOLDER}. The assistant has already called \`remember()\` for the entries shown in \`existingBuffer\`.
+
+Identify additional facts, preferences, plans, corrections, names, dates, decisions, or notable felt moments that should be remembered but aren't already in \`existingBuffer\`. Emit a list of \`remember()\` entries (each one line, in the assistant's first-person voice). Don't duplicate. Prefer to over-remember rather than miss things.
+
+Return only the \`entries\` array.`;
+
+/**
+ * Resolve `SWEEP_PROMPT` with assistant + user names substituted. Falls back
+ * to neutral defaults so the prompt still produces well-formed English when
+ * either name is unavailable on this workspace.
+ */
+export function renderSweepPrompt(opts: {
+  assistantName: string | null;
+  userName: string | null;
+}): string {
+  const assistant = opts.assistantName?.trim() || "the assistant";
+  const user = opts.userName?.trim() || "the user";
+  return SWEEP_PROMPT.replaceAll(
+    ASSISTANT_NAME_PLACEHOLDER,
+    assistant,
+  ).replaceAll(USER_NAME_PLACEHOLDER, user);
+}

package/src/memory/v2/qdrant.ts

@@ -0,0 +1,342 @@
+// ---------------------------------------------------------------------------
+// Memory v2 — Qdrant collection for concept pages
+// ---------------------------------------------------------------------------
+//
+// Owns a dedicated Qdrant collection, keyed by concept-page slug, that holds
+// dense + sparse embeddings of every page under `memory/concepts/`. The
+// collection is separate from the v1 `memory` collection so v2 retrieval can
+// roll out (and roll back) without disturbing the v1 graph + PKB hot path.
+//
+// Mirrors the dense + sparse named-vectors layout used by `VellumQdrantClient`
+// in `qdrant-client.ts` (the v1 PKB collection setup pattern). Connection
+// settings — URL, vector size, on-disk storage — flow through the same env →
+// config precedence as v1 via `resolveQdrantUrl` and `config.memory.qdrant.*`,
+// so users get a consistent Qdrant target without separate v2 knobs.
+//
+// Per-channel queries: `hybridQueryConceptPages` runs separate dense and
+// sparse queries (no Qdrant-side RRF). Callers do their own weighted-sum
+// fusion using `dense_weight` / `sparse_weight` from `config.memory.v2`,
+// which RRF fusion would discard.
+
+import { QdrantClient as QdrantRestClient } from "@qdrant/js-client-rest";
+import { v5 as uuidv5 } from "uuid";
+
+import { getConfig } from "../../config/loader.js";
+import { getLogger } from "../../util/logger.js";
+import type { SparseEmbedding } from "../embedding-types.js";
+import { resolveQdrantUrl } from "../qdrant-client.js";
+
+const log = getLogger("memory-v2-qdrant");
+
+/** Name of the dedicated Qdrant collection holding concept-page embeddings. */
+export const MEMORY_V2_COLLECTION = "memory_v2_concept_pages";
+
+/**
+ * Stable UUIDv5 namespace used to derive a deterministic Qdrant point ID from
+ * a slug. The namespace is an arbitrary fixed UUID; what matters is that the
+ * same slug always maps to the same point ID so upserts replace in place
+ * instead of accumulating duplicates.
+ */
+const SLUG_NAMESPACE = "8b9c5d4f-0e1a-4f3b-9c2d-7e8f1a2b3c4d";
+
+export interface ConceptPagePayload {
+  slug: string;
+  updated_at: number;
+}
+
+/** Per-channel score for a single concept-page hit returned by hybrid query. */
+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.
+   */
+  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.
+   */
+  sparseScore?: number;
+}
+
+let _client: QdrantRestClient | null = null;
+let _collectionReady = false;
+
+/** Lazily create a Qdrant REST client bound to the resolved URL. */
+function getClient(): QdrantRestClient {
+  if (_client) return _client;
+  const config = getConfig();
+  _client = new QdrantRestClient({
+    url: resolveQdrantUrl(config),
+    checkCompatibility: false,
+  });
+  return _client;
+}
+
+/**
+ * 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.
+ */
+export async function ensureConceptPageCollection(): Promise<void> {
+  if (_collectionReady) return;
+
+  const client = getClient();
+  const config = getConfig();
+  const vectorSize = config.memory.qdrant.vectorSize;
+  const onDisk = config.memory.qdrant.onDisk;
+
+  try {
+    const exists = await client.collectionExists(MEMORY_V2_COLLECTION);
+    if (exists.exists) {
+      _collectionReady = true;
+      return;
+    }
+  } catch (err) {
+    // Treat "not found"-shaped errors as "needs creation" and fall through.
+    if (!isCollectionMissing(err)) throw err;
+  }
+
+  log.info(
+    { collection: MEMORY_V2_COLLECTION, vectorSize },
+    "Creating Qdrant collection for memory v2 concept pages",
+  );
+
+  try {
+    await client.createCollection(MEMORY_V2_COLLECTION, {
+      vectors: {
+        dense: {
+          size: vectorSize,
+          distance: "Cosine",
+          on_disk: onDisk,
+        },
+      },
+      sparse_vectors: {
+        sparse: {}, // Qdrant auto-infers sparse vector params
+      },
+      hnsw_config: {
+        on_disk: onDisk,
+        m: 16,
+        ef_construct: 100,
+      },
+      on_disk_payload: onDisk,
+    });
+  } catch (err) {
+    // 409 = a concurrent caller created the collection — that's fine.
+    if (
+      err instanceof Error &&
+      "status" in err &&
+      (err as { status: number }).status === 409
+    ) {
+      _collectionReady = true;
+      return;
+    }
+    throw err;
+  }
+
+  // Slug is the only payload field we filter on; index it once at create-time
+  // so upserts and slug-restricted queries don't pay a per-call indexing cost.
+  await client.createPayloadIndex(MEMORY_V2_COLLECTION, {
+    field_name: "slug",
+    field_schema: "keyword",
+  });
+
+  _collectionReady = true;
+}
+
+/**
+ * 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.
+ */
+export async function upsertConceptPageEmbedding(params: {
+  slug: string;
+  dense: number[];
+  sparse: SparseEmbedding;
+  updatedAt: number;
+}): Promise<void> {
+  await ensureConceptPageCollection();
+
+  const { slug, dense, sparse, updatedAt } = params;
+  const client = getClient();
+  const pointId = pointIdForSlug(slug);
+
+  const upsertOnce = () =>
+    client.upsert(MEMORY_V2_COLLECTION, {
+      wait: true,
+      points: [
+        {
+          id: pointId,
+          vector: { dense, sparse },
+          payload: { slug, updated_at: updatedAt },
+        },
+      ],
+    });
+
+  try {
+    await upsertOnce();
+  } catch (err) {
+    if (isCollectionMissing(err)) {
+      _collectionReady = false;
+      await ensureConceptPageCollection();
+      await upsertOnce();
+      return;
+    }
+    throw err;
+  }
+}
+
+/** Remove the embedding for a slug. Idempotent: no-op when the slug is absent. */
+export async function deleteConceptPageEmbedding(slug: string): Promise<void> {
+  await ensureConceptPageCollection();
+
+  const client = getClient();
+  const doDelete = () =>
+    client.delete(MEMORY_V2_COLLECTION, {
+      wait: true,
+      points: [pointIdForSlug(slug)],
+    });
+
+  try {
+    await doDelete();
+  } catch (err) {
+    if (isCollectionMissing(err)) {
+      _collectionReady = false;
+      await ensureConceptPageCollection();
+      await doDelete();
+      return;
+    }
+    throw err;
+  }
+}
+
+/**
+ * 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.
+ *
+ * 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.
+ *
+ * `restrictToSlugs`, when provided, filters the search server-side to only
+ * those slugs (Qdrant `slug IN [...]` filter). Used by `simBatch` when the
+ * candidate set is already known so we don't waste hits on unrelated pages.
+ * An empty list short-circuits to no results — the caller is asking for
+ * "nothing", not "everything".
+ */
+export async function hybridQueryConceptPages(
+  dense: number[],
+  sparse: SparseEmbedding,
+  limit: number,
+  restrictToSlugs?: readonly string[],
+): Promise<ConceptPageQueryResult[]> {
+  if (restrictToSlugs && restrictToSlugs.length === 0) {
+    // An empty restriction means "no candidates"; skip the round-trip.
+    return [];
+  }
+
+  await ensureConceptPageCollection();
+
+  const client = getClient();
+  const filter = restrictToSlugs
+    ? { must: [{ key: "slug", match: { any: [...restrictToSlugs] } }] }
+    : undefined;
+
+  const denseQuery = () =>
+    client.query(MEMORY_V2_COLLECTION, {
+      query: dense,
+      using: "dense",
+      limit,
+      with_payload: true,
+      filter,
+    });
+  const sparseQuery = () =>
+    client.query(MEMORY_V2_COLLECTION, {
+      query: sparse,
+      using: "sparse",
+      limit,
+      with_payload: true,
+      filter,
+    });
+
+  // Run both queries concurrently — they hit independent named vectors.
+  const runQueries = async () => Promise.all([denseQuery(), sparseQuery()]);
+
+  let denseResults;
+  let sparseResults;
+  try {
+    [denseResults, sparseResults] = await runQueries();
+  } catch (err) {
+    if (isCollectionMissing(err)) {
+      _collectionReady = false;
+      await ensureConceptPageCollection();
+      [denseResults, sparseResults] = await runQueries();
+    } else {
+      throw err;
+    }
+  }
+
+  // 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 });
+    }
+  }
+
+  return Array.from(merged.values());
+}
+
+/**
+ * Detect "collection not found" errors so callers can reset readiness and
+ * retry after an external deletion (e.g. workspace reset).
+ */
+function isCollectionMissing(err: unknown): boolean {
+  if (
+    err &&
+    typeof err === "object" &&
+    "status" in err &&
+    (err as { status: number }).status === 404
+  ) {
+    return true;
+  }
+  const msg = err instanceof Error ? err.message : String(err);
+  return (
+    msg.includes("Not found") ||
+    msg.includes("doesn't exist") ||
+    msg.includes("not found")
+  );
+}
+
+/**
+ * Derive the deterministic Qdrant point ID for a slug. Qdrant requires
+ * UUID/integer IDs; UUIDv5 keeps the mapping stable across processes so
+ * upserts replace in place.
+ */
+function pointIdForSlug(slug: string): string {
+  return uuidv5(slug, SLUG_NAMESPACE);
+}
+
+/** @internal Test-only: reset module-level singletons. */
+export function _resetMemoryV2QdrantForTests(): void {
+  _client = null;
+  _collectionReady = false;
+}

package/src/memory/v2/sim.ts

@@ -0,0 +1,206 @@
+// ---------------------------------------------------------------------------
+// Memory v2 — Hybrid dense + sparse similarity
+// ---------------------------------------------------------------------------
+//
+// Computes per-page similarity scores against a query text by fusing dense
+// (cosine) and sparse (BM25-style) channels via a normalized weighted sum.
+// This is the building block the per-turn activation formula (`A_o`) uses to
+// score candidate concept pages against the latest user message, the latest
+// assistant message, and NOW context.
+//
+// Why weighted-sum fusion (not RRF):
+//   The activation formula in §4 of the design doc multiplies similarity
+//   scores by config-tunable coefficients (`c_user`, `c_assistant`, `c_now`)
+//   and adds them together. RRF would discard the score magnitudes the
+//   coefficients operate on — it returns a rank-based pseudo-score that does
+//   not blend smoothly with `d · A(n, t)`. Hybrid sim therefore queries each
+//   channel separately and fuses with the configured `dense_weight` /
+//   `sparse_weight` (which the schema validates sum to 1.0).
+//
+// Sparse normalization:
+//   Dense cosine similarity is already in [0, 1]. Qdrant's sparse score is
+//   on a different, unbounded scale (it depends on query and document term
+//   weights), so we divide by the per-batch maximum sparse score to bring
+//   it into [0, 1] before fusing. This is the design doc's choice (§4) —
+//   batch-relative normalization is sufficient because the score is consumed
+//   only as a per-turn ordering signal, not compared across turns.
+
+import type { AssistantConfig } from "../../config/types.js";
+import {
+  embedWithBackend,
+  generateSparseEmbedding,
+} from "../embedding-backend.js";
+import { clampUnitInterval } from "../validation.js";
+import { hybridQueryConceptPages } from "./qdrant.js";
+import { hybridQuerySkills } from "./skill-qdrant.js";
+
+/**
+ * Clamp a value into the closed unit interval [0, 1]. Re-exported under the
+ * design-doc name so call sites that mirror the formula in §4 read cleanly.
+ */
+export const clamp01 = clampUnitInterval;
+
+/**
+ * Compute hybrid (dense + sparse) similarity scores between a query text and
+ * a fixed set of candidate concept-page slugs.
+ *
+ * Steps:
+ *   1. Embed the query text (dense via the configured embedding backend,
+ *      sparse via the in-process TF-IDF encoder).
+ *   2. Run server-side dense + sparse queries against the v2 concept-page
+ *      Qdrant collection, restricted to `candidateSlugs` so we don't waste
+ *      query bandwidth on unrelated pages.
+ *   3. Fuse: per slug, `score = clamp01(dense_weight · denseCosine +
+ *      sparse_weight · normalizedSparse)`. Sparse scores are normalized by
+ *      the per-batch maximum (so the largest is 1.0); slugs missing from a
+ *      channel contribute 0 from that channel.
+ *
+ * Returns a `Map<slug, score>` containing only the candidate slugs that hit
+ * in at least one channel. Slugs in `candidateSlugs` that miss both channels
+ * are absent from the map; callers should treat absence as score = 0 (the
+ * activation pipeline does this implicitly when reading back A_o).
+ *
+ * Edge cases:
+ *   - Empty `candidateSlugs` → returns an empty map without touching Qdrant
+ *     or the embedding backend.
+ *   - Empty query text or all-zero sparse vector → still queries (dense may
+ *     still hit), and the sparse contribution to fusion is zero.
+ */
+export async function simBatch(
+  text: string,
+  candidateSlugs: readonly string[],
+  config: AssistantConfig,
+): Promise<Map<string, number>> {
+  if (candidateSlugs.length === 0) {
+    return new Map();
+  }
+
+  // Sparse uses the shared TF-IDF encoder so the query and stored vectors
+  // share a vocabulary with PKB indexing.
+  const denseResult = await embedWithBackend(config, [text]);
+  const denseVector = denseResult.vectors[0];
+  const sparseVector = generateSparseEmbedding(text);
+
+  const hits = await hybridQueryConceptPages(
+    denseVector,
+    sparseVector,
+    candidateSlugs.length,
+    candidateSlugs,
+  );
+
+  if (hits.length === 0) {
+    return new Map();
+  }
+
+  const maxSparse = computeMaxSparse(hits);
+  const { dense_weight: denseWeight, sparse_weight: sparseWeight } =
+    config.memory.v2;
+
+  const scores = new Map<string, number>();
+  for (const hit of hits) {
+    scores.set(hit.slug, fuseHit(hit, maxSparse, denseWeight, sparseWeight));
+  }
+  return scores;
+}
+
+/**
+ * Compute hybrid (dense + sparse) similarity scores between a query text and
+ * a fixed set of candidate skill ids. Mirrors `simBatch` but targets the
+ * dedicated `memory_v2_skills` Qdrant collection via `hybridQuerySkills`.
+ *
+ * Differences from `simBatch`:
+ *   - Keys are skill `id` values (not concept-page slugs).
+ *   - Restricts the query to the caller's candidate ids server-side via
+ *     `hybridQuerySkills`'s `restrictToIds` parameter. Without this, when the
+ *     skills collection has more skills than `ids.length`, Qdrant would
+ *     return its global top-K and candidate ids absent from that top-K would
+ *     silently score 0 — corrupting the activation calculation.
+ *
+ * Returns a `Map<id, score>` of fused scores in [0, 1]. Ids that did not hit
+ * either channel are absent from the map.
+ *
+ * Edge cases:
+ *   - Empty `ids` → returns an empty map without touching Qdrant or the
+ *     embedding backend.
+ *   - Empty query text → still queries (dense may still hit), and the sparse
+ *     contribution is zero.
+ */
+export async function simSkillBatch(
+  text: string,
+  ids: readonly string[],
+  config: AssistantConfig,
+): Promise<Map<string, number>> {
+  if (ids.length === 0) {
+    return new Map();
+  }
+
+  const denseResult = await embedWithBackend(config, [text]);
+  const denseVector = denseResult.vectors[0];
+  const sparseVector = generateSparseEmbedding(text);
+
+  const hits = await hybridQuerySkills(
+    denseVector,
+    sparseVector,
+    ids.length,
+    ids,
+  );
+
+  if (hits.length === 0) {
+    return new Map();
+  }
+
+  // Defensive post-filter — `hybridQuerySkills` restricts server-side, so
+  // every hit should already be in `ids`, but keep this guard so a buggy
+  // payload (e.g. a missing/typoed id index) can't silently inject
+  // out-of-set ids into the score map.
+  const idSet = new Set(ids);
+  const filtered = hits.filter((h) => idSet.has(h.id));
+  if (filtered.length === 0) {
+    return new Map();
+  }
+
+  const maxSparse = computeMaxSparse(filtered);
+  const { dense_weight: denseWeight, sparse_weight: sparseWeight } =
+    config.memory.v2;
+
+  const scores = new Map<string, number>();
+  for (const hit of filtered) {
+    scores.set(hit.id, fuseHit(hit, maxSparse, denseWeight, sparseWeight));
+  }
+  return scores;
+}
+
+/**
+ * Per-batch sparse-score maximum used for normalization. Hits missing from
+ * the sparse channel contribute 0 (handled by the `undefined` guard).
+ */
+function computeMaxSparse(
+  hits: ReadonlyArray<{ sparseScore?: number }>,
+): number {
+  let max = 0;
+  for (const hit of hits) {
+    if (hit.sparseScore !== undefined && hit.sparseScore > max) {
+      max = hit.sparseScore;
+    }
+  }
+  return max;
+}
+
+/**
+ * Fuse a single hit's dense + sparse scores into a normalized [0, 1] score
+ * via `clamp01(dense_weight · dense + sparse_weight · sparse/maxSparse)`.
+ * Missing-channel scores contribute 0.
+ */
+function fuseHit(
+  hit: { denseScore?: number; sparseScore?: number },
+  maxSparse: number,
+  denseWeight: number,
+  sparseWeight: number,
+): number {
+  const dense = hit.denseScore ?? 0;
+  const sparseNormalized =
+    hit.sparseScore !== undefined && maxSparse > 0
+      ? hit.sparseScore / maxSparse
+      : 0;
+  return clamp01(denseWeight * dense + sparseWeight * sparseNormalized);
+}

package/src/memory/v2/skill-content.ts

@@ -0,0 +1,42 @@
+import { getConfig } from "../../config/loader.js";
+import type { SkillCapabilityInput } from "../../skills/skill-memory.js";
+
+/**
+ * Render the prose-style capability statement embedded into the
+ * `memory_v2_skills` Qdrant collection and rendered in
+ * `### Skills You Can Use`. Capped at 500 chars to match v1's behavior.
+ */
+export function buildSkillContent(input: SkillCapabilityInput): string {
+  let content = `The "${input.displayName}" skill (${input.id}) is available. ${input.description}.`;
+  if (input.activationHints && input.activationHints.length > 0) {
+    content += ` Use when: ${input.activationHints.join("; ")}.`;
+  }
+  if (input.avoidWhen && input.avoidWhen.length > 0) {
+    content += ` Avoid when: ${input.avoidWhen.join("; ")}.`;
+  }
+  if (content.length > 500) {
+    content = content.slice(0, 500);
+  }
+  return content;
+}
+
+/**
+ * mcp-setup is special-cased in v1 (`capability-seed.ts:102-112`):
+ * its description is augmented with the list of configured MCP server
+ * names so the model can pattern-match against them. Port verbatim.
+ */
+export function augmentMcpSetupDescription(
+  input: SkillCapabilityInput,
+): SkillCapabilityInput {
+  if (input.id !== "mcp-setup") return input;
+  const servers = getConfig().mcp?.servers;
+  if (!servers) return input;
+  const names = Object.keys(servers).filter(
+    (name) => servers[name]?.enabled !== false,
+  );
+  if (names.length === 0) return input;
+  return {
+    ...input,
+    description: `${input.description} Configured: ${names.join(", ")}`,
+  };
+}