KnowledgeBase

The problem

Legal research involves finding relevant material, relating facts across documents, checking quotations and keeping knowledge current as new sources arrive. The system supports these steps while preserving links back to the evidence.

The system

• Retrieval: keyword/vector hybrid search, HyDE query expansion and hierarchical retrieval from summaries and entities down to source material.
• Agent tools: MCP interfaces shared by clients including LibreChat, Claude.ai and ChatGPT. The LibreChat front end uses its own LangGraph-based agent framework.
• Evidence: extracted entities and facts retain supporting passages; quotation tools compare quoted text with the original source and reject mismatches.
• Delivery: AWS deployment, persistent background tasks, idempotent requests and result reuse. Automated checks cover rollback, concurrent operations and partial batch failures.

The implementation uses FastAPI, PostgreSQL/pgvector, Docker Compose and GitHub Actions. The pilot is iterated in response to lawyers’ feedback.

Retrieval design evaluation

In probability simulations with a fixed corpus size and five relevance densities, hierarchical retrieval required approximately 32–77% less simulated query token budget than flat retrieval at the same 90% mean recall.

These are results under the simulation’s reference routing and ranking assumptions, not measured production savings or a guarantee across all queries. Less accurate summaries or stronger flat ranking can reduce or reverse the advantage.

Cost and processing efficiency

• Server configuration: reduced server costs by approximately 67%.
• Incremental entity maintenance: replace repeated processing of all historical sources with the existing entity page plus new evidence. For appended sources of bounded size, cumulative source-material processing changes from O(n²) to O(n). This does not describe the cost of repeatedly reading and rewriting the growing entity page.
• Asynchronous model calls: use the Batch API for non-real-time work, with 50% lower input and output token unit prices for eligible calls. This is a unit-price reduction, not a halving of token consumption or the entire system bill.

The three improvements address different costs and are not added together into one savings percentage.