⚡ Bolt: [Graph Entity Similarity Search Optimization]

.jules/bolt.md

@@ -1,7 +1,3 @@
-## 2024-05-22 - Async File I/O in Synchronous Methods
-**Learning:** Even small JSON files (1000 items) can cause significant blocking (23ms) when written synchronously in a tight loop. `dataclasses.asdict` is relatively slow (5ms) but the main bottleneck was redundant disk writes. Offloading to `ThreadPoolExecutor` and removing redundant saves in `tick()` improved performance by 5x (23ms -> 4.7ms).
-**Action:** When optimizing "save on every change" patterns, check if intermediate saves (like in `tick`) are necessary and offload the actual I/O to a background thread to unblock the main execution path.
-
-## 2024-05-23 - Synchronous Audit Logging Bottleneck
-**Learning:** `ToolOrchestrator._audit_action` was performing synchronous file I/O (open/write/close) for every tool invocation. This introduced ~68ms latency per 1000 calls. Moving this to a background thread with `queue.Queue` reduced it to ~3ms (20x improvement).
-**Action:** For high-frequency logging or audit trails, always use an asynchronous writer or background thread to decouple I/O latency from the main execution path.
+## 2026-05-29 - Graph Similarity Search N+1 Query Optimization
+**Learning:** The entity similarity search performed N+1 queries by manually fetching related entities in a Python loop to calculate Jaccard similarity.
+**Action:** Replaced the loop with a single SQL query using CTEs to compute intersections and total counts, calculating similarity on the database side to avoid N+1 queries.