A measured retrieval service for cases and reference documents

Every agent at a conveyancing firm, from the one drafting enquiries to the one reviewing the report on title, needs to find the right clause, register entry or reply and cite it. This project builds that shared retrieval layer once, properly, and leaves you with an eval harness you can rerun whenever an embedding model, a reranker or a price changes. It uses every lesson: embeddings, chunking and indexing (module 1), measured by the query set from lesson 4, then hybrid search, agentic retrieval, a records graph and the long-context comparison (module 2).

Every model call runs on your Claude subscription through Claude Code headless (claude -p). Embeddings come from a free local model through Ollama or, optionally, your own OpenRouter key; the reference answer shows both. Commercial rerankers are optional stretch steps. Expect 8–12 hours.