US-023

progress.txt

@@ -519,3 +519,17 @@
   - `notes.update` allows partial field updates (title or content can be omitted). Always re-fetch the full note from SQLite after the update write to get the correct combined text for embedding
   - `vectordb`'s `table.delete(where)` accepts a raw SQL WHERE clause string. UUID v4 IDs are safe to interpolate directly (only `[0-9a-f-]` characters)
 ---
+
+## 2026-02-24 - US-023
+- Implemented @KnowledgeAgent semantic search across all projects
+- Added `searchNotes()` to `src/main/db/vectordb.ts`: embeds query via `embedText()`, performs `table.search(vector).limit(k).execute()` on LanceDB notes table, returns `SearchResult[]` with id, projectId, content, _distance
+- Added `buildKnowledgeTools()` to `src/main/ai/orchestrator.ts`: defines `vector_search_all` tool that calls `searchNotes()`, joins SQLite for note title and project name, returns formatted results with `From: [Project Name] — [Note Title]` citation headers
+- Rewrote `knowledgeAgent()` from simple LLM stub to full tool-calling agent loop (mirrors projectAgent/generalAgent pattern: TOOL_CALLING_PROVIDERS check, bindTools, 5-iteration MAX_ITERATIONS loop with ToolMessage accumulation, fallback for providers without tool support)
+- Updated `makeKnowledgeAgentPrompt()` with `withTools` parameter, tool documentation for `vector_search_all`, and citation format instructions
+- Files changed: `src/main/db/vectordb.ts`, `src/main/ai/orchestrator.ts`
+- **Learnings for future iterations:**
+  - `openTable<T>('name')` in vectordb makes `search()` expect `T` as input type — omit the generic when using `search()` with a raw vector array
+  - The agent loop pattern (check TOOL_CALLING_PROVIDERS → buildTools → bindTools → iterate with ToolMessage) is now used consistently across all three agents (project, knowledge, general) — follow this pattern for any new agent
+  - LanceDB `table.search(vector).limit(k).execute()` returns objects with all stored fields plus `_distance` (L2 distance, lower = more similar)
+  - The `SearchResult` type is exported from `vectordb.ts` for reuse in the orchestrator — keep vector DB types in the DB module, not the AI module
+---