@tekmidian/pai 0.9.9 → 0.9.11

package/dist/daemon-mcp/index.mjs

@@ -4860,6 +4860,52 @@ async function startShim() {
 		min_occurrences: number().int().min(1).optional().describe("Minimum total chunk occurrences across all projects. Default: 3."),
 		limit: number().int().min(1).max(100).optional().describe("Maximum number of tunnels to return. Default: 20.")
 	}, async (args) => proxyTool("memory_tunnels", args));
+	server.tool("memory_feedback", [
+		"Apply relevance feedback to memory chunks to improve future search ranking.",
+		"",
+		"After reading search results, rate their relevance to adjust the EMA-based",
+		"feedback weights. Higher-rated chunks will score better in future searches",
+		"via the multiplier: final_score *= (0.5 + relevance_score).",
+		"",
+		"Also updates feedback weights on KG entities mentioned in those chunks.",
+		"",
+		"Rating scale: 1 = not relevant, 3 = somewhat relevant, 5 = highly relevant.",
+		"",
+		"Use chunk IDs from memory_search compact format results."
+	].join("\n"), {
+		chunk_ids: array(string()).describe("Array of chunk IDs to apply feedback to (from memory_search results)."),
+		rating: number().min(1).max(5).describe("Relevance rating from 1 (not relevant) to 5 (highly relevant)."),
+		tenant_id: string().optional().describe("Tenant ID for entity feedback scoping. Default: 'default'.")
+	}, async (args) => proxyTool("memory_feedback", args));
+	server.tool("memory_kg_search", [
+		"Graph-completion search: combines vector search with knowledge-graph neighborhood expansion.",
+		"",
+		"Algorithm:",
+		"  Phase 1: Wide vector search (seed chunks) using the query embedding",
+		"  Phase 2: Extract entity mentions from seed chunks (matched against kg_entities)",
+		"  Phase 3: BFS neighborhood expansion in kg_triples (1-2 hops from matched entities)",
+		"  Phase 4: Re-rank all collected triples against the query embedding",
+		"",
+		"Returns ranked KG triples with relevance scores, surfacing graph-derived context",
+		"that pure vector search would miss — relationships, facts, and entity connections.",
+		"",
+		"Requires Postgres backend with a populated kg_triples table.",
+		"Use kg_add to populate the knowledge graph, or rely on automatic extraction.",
+		"",
+		"Parameters:",
+		"  query              — Free-text query (converted to embedding for Phase 1)",
+		"  project_id         — Restrict seed search to a specific project (optional)",
+		"  wide_k             — Number of seed chunks from Phase 1 (default: 50)",
+		"  neighborhood_depth — BFS hop depth for KG expansion (default: 1, max: 2)",
+		"  top_k              — Maximum triples to return after re-ranking (default: 20)"
+	].join("\n"), {
+		query: string().describe("Free-text search query — used to generate embedding for Phase 1 vector search."),
+		project_id: number().int().optional().describe("Restrict seed vector search to a specific project ID. Omit to search all projects."),
+		wide_k: number().int().min(1).max(200).optional().describe("Number of seed chunks to fetch in Phase 1 vector search. Default: 50."),
+		neighborhood_depth: number().int().min(1).max(2).optional().describe("BFS hop depth for KG neighborhood expansion. Default: 1. Max: 2."),
+		top_k: number().int().min(1).max(100).optional().describe("Maximum KG triples to return after re-ranking. Default: 20."),
+		tenant_id: string().optional().describe("Tenant ID for entity lookup scoping. Default: 'default'.")
+	}, async (args) => proxyTool("memory_kg_search", args));
 	const transport = new StdioServerTransport();
 	await server.connect(transport);
 }