| name | route-by-query-shape |
| description | When the agent calls memory_search with a relationship-shaped query ("who did I talk to about X"), redirect to the knowledge_graph backend where it will actually find the answer. |
| tier | executable |
| bitterbot | {"always":false,"interceptors":[{"id":"route-by-query-shape:relationship","builtin":true,"activates_on":"memory_search with relationship-shaped query","intervention":"modify (route to knowledge_graph backend)"}]} |
route-by-query-shape
When you ask Bitterbot a relationship question — "who did I talk to about Q4?", "who worked with me on Foo?" — the vector-based memory_search often misses because the answer lives in entity relationships, not in chunk text. The graph backend gets these right.
This interceptor watches memory_search calls, detects relationship-shaped queries, extracts the salient entity, and re-routes the call to the graph backend before it executes.
What you'll see
For relationship questions, the agent will hit the knowledge graph rather than searching through unstructured text. Expect noticeably better hit rates on "who did X with Y" style questions.
Implementation
Built-in interceptor route-by-query-shape:relationship lives in src/agents/skills/builtin-interceptors/route-by-query-shape.ts. Fires up to 12 times per session.
