@vellumai/assistant 0.5.15 → 0.6.0

package/src/memory/graph/graph-search.ts

@@ -0,0 +1,266 @@
+// ---------------------------------------------------------------------------
+// Memory Graph — Qdrant vector search for graph nodes
+// ---------------------------------------------------------------------------
+
+import type { AssistantConfig } from "../../config/types.js";
+import { getLogger } from "../../util/logger.js";
+import { selectedBackendSupportsMultimodal } from "../embedding-backend.js";
+import type { EmbeddingInput } from "../embedding-types.js";
+import { embedAndUpsert } from "../job-utils.js";
+import { asString } from "../job-utils.js";
+import { enqueueMemoryJob, type MemoryJob } from "../jobs-store.js";
+import { isQdrantBreakerOpen } from "../qdrant-circuit-breaker.js";
+import { withQdrantBreaker } from "../qdrant-circuit-breaker.js";
+import {
+  getQdrantClient,
+  type QdrantSearchResult,
+  type QdrantSparseVector,
+} from "../qdrant-client.js";
+import { loadImageRefData } from "./image-ref-utils.js";
+import { getNode } from "./store.js";
+import type { MemoryNode } from "./types.js";
+
+const log = getLogger("graph-search");
+
+// ---------------------------------------------------------------------------
+// Search
+// ---------------------------------------------------------------------------
+
+export interface GraphSearchResult {
+  nodeId: string;
+  score: number;
+  text: string;
+}
+
+/**
+ * Semantic search across graph nodes in Qdrant. Returns scored node IDs
+ * that the caller can hydrate from the graph store.
+ *
+ * Filters to `target_type: "graph_node"` and optionally by scope.
+ */
+export async function searchGraphNodes(
+  queryVector: number[],
+  limit: number,
+  scopeIds?: string[],
+  sparseVector?: QdrantSparseVector,
+): Promise<GraphSearchResult[]> {
+  if (isQdrantBreakerOpen()) {
+    log.warn("Qdrant circuit breaker open, skipping graph search");
+    return [];
+  }
+
+  const client = getQdrantClient();
+
+  // Use hybrid search (dense + sparse with RRF fusion) when a non-empty
+  // sparse vector is available; otherwise fall back to dense-only search.
+  if (sparseVector && sparseVector.indices.length > 0) {
+    const filter = {
+      must: [
+        { key: "target_type", match: { value: "graph_node" } },
+        ...(scopeIds && scopeIds.length > 0
+          ? [{ key: "memory_scope_id", match: { any: scopeIds } }]
+          : []),
+      ],
+      must_not: [{ key: "_meta", match: { value: true } }],
+    };
+
+    const results: QdrantSearchResult[] = await withQdrantBreaker(() =>
+      client.hybridSearch({
+        denseVector: queryVector,
+        sparseVector,
+        filter,
+        limit,
+        prefetchLimit: limit * 3,
+      }),
+    );
+
+    return results.map((r) => ({
+      nodeId: r.payload.target_id,
+      score: r.score,
+      text: r.payload.text,
+    }));
+  }
+
+  // Dense-only fallback
+  const filter: Record<string, unknown> = {
+    must: [
+      {
+        key: "target_type",
+        match: { value: "graph_node" },
+      },
+    ],
+  };
+
+  if (scopeIds && scopeIds.length > 0) {
+    (filter.must as unknown[]).push({
+      key: "memory_scope_id",
+      match: { any: scopeIds },
+    });
+  }
+
+  const results: QdrantSearchResult[] = await withQdrantBreaker(async () => {
+    return client.search(queryVector, limit, filter);
+  });
+
+  return results.map((r) => ({
+    nodeId: r.payload.target_id,
+    score: r.score,
+    text: r.payload.text,
+  }));
+}
+
+// ---------------------------------------------------------------------------
+// Embedding job
+// ---------------------------------------------------------------------------
+
+/**
+ * Format a graph node's content for embedding. Prepends type metadata
+ * so the embedding captures structural information alongside content.
+ */
+function formatNodeForEmbedding(node: MemoryNode): string {
+  const parts = [`[${node.type}]`];
+  if (node.emotionalCharge.intensity > 0.3) {
+    const valenceLabel =
+      node.emotionalCharge.valence > 0.3
+        ? "positive"
+        : node.emotionalCharge.valence < -0.3
+          ? "negative"
+          : "neutral";
+    parts.push(`[${valenceLabel}]`);
+  }
+  parts.push(node.content);
+  return parts.join(" ");
+}
+
+/**
+ * Embed a graph node and upsert to Qdrant. Can be called directly
+ * (synchronous embedding during bootstrap) or via the job handler.
+ *
+ * When the node has image references and the Gemini embedding backend is
+ * available, embeds the image content directly for cross-modal retrieval
+ * (text queries match image memories in the same vector space). Falls back
+ * to text embedding with image description suffixes otherwise.
+ */
+export async function embedGraphNodeDirect(
+  node: MemoryNode,
+  config: AssistantConfig,
+): Promise<void> {
+  if (node.fidelity === "gone") return;
+
+  const text = formatNodeForEmbedding(node);
+  const extraPayload: Record<string, unknown> = {
+    created_at: node.created,
+    memory_scope_id: node.scopeId,
+    confidence: node.confidence,
+    importance: node.significance,
+    kind: node.type,
+  };
+
+  if (node.imageRefs && node.imageRefs.length > 0) {
+    const multimodalAvailable = await selectedBackendSupportsMultimodal(config);
+    if (multimodalAvailable) {
+      const imageData = await loadImageRefData(node.imageRefs[0]);
+      if (imageData) {
+        try {
+          const input: EmbeddingInput = {
+            type: "image",
+            data: imageData.data,
+            mimeType: imageData.mimeType,
+          };
+          await embedAndUpsert(config, "graph_node", node.id, input, {
+            ...extraPayload,
+            has_image: true,
+          });
+          return;
+        } catch (err) {
+          log.warn(
+            "Multimodal embed failed for node %s, falling back to text: %s",
+            node.id,
+            err instanceof Error ? err.message : String(err),
+          );
+        }
+      }
+    }
+
+    // Fallback: text embedding with image description suffix
+    const descSuffix = node.imageRefs.map((r) => r.description).join("; ");
+    const textWithImages = `${text}\n[images: ${descSuffix}]`;
+    await embedAndUpsert(
+      config,
+      "graph_node",
+      node.id,
+      textWithImages,
+      extraPayload,
+    );
+    return;
+  }
+
+  await embedAndUpsert(config, "graph_node", node.id, text, extraPayload);
+}
+
+/**
+ * Job handler: embed a graph node and upsert to Qdrant.
+ */
+export async function embedGraphNodeJob(
+  job: MemoryJob,
+  config: AssistantConfig,
+): Promise<void> {
+  const nodeId = asString(job.payload.nodeId);
+  if (!nodeId) return;
+
+  const node = getNode(nodeId);
+  if (!node) return;
+
+  await embedGraphNodeDirect(node, config);
+}
+
+/**
+ * Enqueue an embedding job for a graph node (async, for live conversations).
+ */
+export function enqueueGraphNodeEmbed(nodeId: string): void {
+  enqueueMemoryJob("embed_graph_node", { nodeId });
+}
+
+/**
+ * Job handler: embed a trigger's condition text and store the
+ * embedding on the trigger row.
+ */
+export async function embedGraphTriggerJob(
+  job: MemoryJob,
+  config: AssistantConfig,
+): Promise<void> {
+  const triggerId = asString(job.payload.triggerId);
+  if (!triggerId) return;
+
+  // Import here to avoid circular dependency
+  const { getDb } = await import("../db.js");
+  const { eq } = await import("drizzle-orm");
+  const { memoryGraphTriggers } = await import("../schema.js");
+  const { embedWithBackend } = await import("../embedding-backend.js");
+
+  const db = getDb();
+  const row = db
+    .select()
+    .from(memoryGraphTriggers)
+    .where(eq(memoryGraphTriggers.id, triggerId))
+    .get();
+
+  if (!row || !row.condition) return;
+
+  const result = await embedWithBackend(config, [row.condition]);
+  const vector = result.vectors[0];
+  if (!vector) return;
+
+  const buffer = Buffer.from(new Float32Array(vector).buffer);
+  db.update(memoryGraphTriggers)
+    .set({ conditionEmbedding: buffer })
+    .where(eq(memoryGraphTriggers.id, triggerId))
+    .run();
+}
+
+/**
+ * Enqueue a trigger embedding job.
+ */
+export function enqueueGraphTriggerEmbed(triggerId: string): void {
+  enqueueMemoryJob("graph_trigger_embed", { triggerId });
+}

package/src/memory/graph/image-ref-utils.ts

@@ -0,0 +1,29 @@
+import { eq } from "drizzle-orm";
+
+import { getDb } from "../db.js";
+import { extractMediaBlocks } from "../message-content.js";
+import { messages } from "../schema.js";
+import type { ImageRef } from "./types.js";
+
+/**
+ * Load image data from the messages table for an ImageRef.
+ * Returns null if the message or image block no longer exists
+ * (e.g., conversation was deleted).
+ */
+export async function loadImageRefData(
+  ref: ImageRef,
+): Promise<{ data: Buffer; mimeType: string } | null> {
+  const db = getDb();
+  const msg = db
+    .select({ content: messages.content })
+    .from(messages)
+    .where(eq(messages.id, ref.messageId))
+    .get();
+  if (!msg) return null;
+
+  const mediaBlocks = extractMediaBlocks(msg.content);
+  const block = mediaBlocks.find((b) => b.index === ref.blockIndex);
+  if (!block) return null;
+
+  return { data: block.data, mimeType: block.mimeType };
+}