Searching the corpus and generating a grounded answer…
Answer
Related documentsGRAPH RAG
Connected to the sources above in the citation graph — open to explore their references.
Citations & References
Ask a legal or tax question in plain English. Every answer is grounded in the source documents, with clickable citations to the exact document, section, and page.
Reading the document and writing a faithful summary…
Summary
Pick any of the 100 corpus documents to get a concise, faithful summary of its key provisions.
Graph RAG — find which documents cite or reference a given Act, statute, or case. Something pure semantic search can't answer.
Try
Traversing the citation graph…
Documents citing “”
Enter an Act, statute, or case name to traverse the citation graph and see every corpus document that references it — grouped by type.
Every question is graded from a full evaluation run — run any one to confirm it live against the hand-authored ground truth.
Hybrid retrieval (vector + BM25, fused with Reciprocal Rank Fusion) plus Graph RAG over a Neo4j citation graph, with grounded generation via Gemini 2.5 Pro — every answer carries mandatory citations (document · section · page) and passes a citation-validation guard.
Evaluation Report
End-to-end on a hand-authored 100-query golden set (25 per category: Acts / Court Judgments / POV / Tax). Answer model: Gemini 2.5 Pro.
100% grounded / 0% hallucinated — every answer cites document · section · page, or correctly refuses
Overall performance on the golden set
Retrieval (Top-1 / Top-5)
Faithfulness
Correctness
Grounded / Halluc.
95% / 99%
0.96
91% / 98%
100% / 0%
Observations, limitations & improvements
Strengths. Strong retrieval with highly faithful, fully-cited answers — for a legal tool, not hallucinating is the critical property. Graph RAG answers relationship queries pure search can't.
Limitations. Tax retrieval is the soft spot (84% Top-1 — IRS pubs share heavy vocabulary; always within Top-5). The corpus spans 2025 + 2026 tax years, so year-agnostic questions can surface both. Embeddings are general-purpose (bge-base-en-v1.5), not legal-tuned.
Improvements. Tax-year disambiguation in the prompt (implemented); a legal-domain embedding model to lift Tax Top-1; broaden citation-extraction recall in the graph.