smartle

most rag setups assume all your knowledge lives in one kind of place — a document store, usually. real questions don’t respect that: answering something properly might mean pulling from a pdf, hitting an internal api, and querying a database, all for one question. smartle was my attempt at a system that routes across all of those instead of pretending everything’s a document.

a router agent looks at the incoming question and decides which specialised agent(s) should handle it — one agent works over document embeddings in chromadb, another calls structured apis, another queries databases directly, another handles general web-sourced knowledge. short-term memory lets follow-up questions in the same session build on what’s already been retrieved instead of starting cold every time.

the interesting failure mode was routing confidence — the router occasionally sent a question to the wrong specialist and got a confidently wrong answer instead of an honest “i don’t know.” tightening the routing logic to fail toward “ask the wrong agent to hand off” rather than “commit to a bad guess” ended up mattering more than any single agent’s retrieval quality.