A measured retrieval service for cases and reference documents

A complete build of the capstone. Two fictional firms' synthetic matters and precedent libraries in the ingest hand-off format, structure-aware chunks with contextual headers, an event-driven index whose every read is scoped by the session, hybrid search with an optional cross-encoder, a tool CLI and records graph for a claude -p agent, cached long context and a router, and evaluations with bootstrap intervals on a dev split and a held-out split, ending in a recommendation memo.

Back to the project

The service is one index with one rule: every read takes a Scope, and a Scope is only ever built from the session. Chunking, hybrid search, the agent's tools, long context and the router all sit on top of that rule, so tenant isolation and freshness are properties of the index rather than checks that each feature has to remember. The second decision that shapes everything is that the query set and its labels are fixed before anything is tuned, with a held-out 20% that is run once at the end. Every model call goes through claude -p on your subscription. Embeddings go through the OpenAI SDK, on your own OpenRouter key or a free local model through Ollama, with a deterministic fake when neither is configured. Every person, firm, address and title number is invented.

Back to the project

A measured retrieval service for cases and reference documents

A measured retrieval service for cases and reference documents

Sign in to keep reading the answer.

Sign in to keep reading the answer.