Automating legal research inside the firm's own walls

Lawyer Assistant is an AI research system that searches a law firm's own documents, verifies citations, and produces evidence-backed answers — so legal professionals spend their time on judgment, not page-turning.

Visit Website View Source

The problem

Hours lost to searching: legal professionals spend significant time locating precedents, clauses, and facts across large collections of case files and contracts.
Verification burden: every answer must be traceable to a source — an unattributed claim is worthless in legal work.
Confidentiality: client data cannot be sent to cloud AI tools without risking privilege.

The existing workflow

A question arrives — from a partner, a client matter, or a review task.
An associate searches folders and databases by hand, then reads dozens of pages to find relevant passages.
Facts get extracted into notes, cross-checked, and written up with citations — slow, inconsistent, and easy to get wrong under deadline pressure.

The automated workflow

The lawyer asks a question in plain English.
The system retrieves the most relevant passages from the firm's own corpus using hybrid semantic and keyword search.
Every citation is verified against the source document before it reaches the answer.
An evidence-backed answer comes back with inline citations the lawyer can check — the lawyer, not the system, makes the call.

How it's built

Retrieval: BGE-M3 embeddings (ChromaDB) fused with BM25 keyword search, re-ranked for relevance.
Inference: local language models served via Ollama — quantized GGUF weights, no account, no cloud path, no telemetry.
Deployment: a desktop application for Windows, macOS, and Linux that runs entirely on hardware the firm owns.

Evaluation — measured before trusted

The system was tested against its own benchmark suite before being presented as dependable: a 200-question runner, a golden evaluation set, and CUAD-based contract tasks measuring retrieval quality and citation correctness.

Citation accuracy

96%

Search recall

94%

Data privacy

100%

Figures from the project's own evaluation suite — see backend/benchmark in the GitHub repository. Data privacy is architectural: offline mode has no cloud path.

The result

Research that used to consume an afternoon of manual reading becomes a cited answer a lawyer verifies in minutes — on hardware the firm owns, with privilege intact. The lawyer stays the decision-maker; the automation does the legwork.

Have a research workflow like this?

If your team spends hours searching and reviewing documents, tell me about it — I'll give you an honest assessment of whether it can be automated.

Describe Your Workflow