Context Engine for
Modern Product Teams

Slack threads, code chunks, wiki sections, incidents, and custom database rows all land in one table — document, embedding, metadata, source, timestamps. Each connector defines what the data is, how to reach it, and how often to fetch. Adding a source is a pull request.

Raw text is keyword-searchable the moment it lands via a Postgres full-text index. For vector search, an LLM distills each thread into a normalized document — the question being asked, a summary, the resolution, and the systems and code references mentioned — and that is what gets embedded, not the raw transcript.

One embeddings table in Postgres. Anything that lands in it — from a Slack thread to a netlist — is immediately queryable through the same interface, with no per-source search stack to maintain.

Full-text catches exact tokens and pasted errors, embeddings catch paraphrase across vocabularies, IDF separates rare signal from filler, and age decay favors fresher answers. No single scorer is trusted on its own — each produces its own ranked list at query time.

Reciprocal rank fusion merges the incompatible lists — score(d) = Σ 1 / (60 + rank) — then duplicates are merged and per-file contributions capped. A diverse top twenty goes to a small reranker, the top ten survive, and winners are expanded with their neighboring context.

A planner picks the retrieval tools, an executor runs them, and a synthesizer writes the answer strictly from the evidence rows it was handed — every claim carrying a citation back to the thread, file, or ticket it came from.