Dicta indexes every published U.S. court opinion (the full CourtListener corpus, ~10.3M opinions) and ranks it the way the law ranks itself: by who cites whom. It answers natural-language legal questions with pinpoint citations, and it checks citations an AI or a person has already written.
Why It Matters
Language models invent case law. Lawyers have been sanctioned for filing briefs that cite cases that don't exist. Dicta's lead product is the guard against that failure: hand it a list of citations, and it tells you which are real, which match the case name they claim, and which are still good law.
Key Features
- >Citation Verification API: Batch-verify citations: exists or not, name matches or not, still good law or not. Name-match detection catches hallucinated citations that borrow a real reporter cite
- >Authority-Ranked Search: Tantivy BM25 over the full text, blended with PageRank over the 77M-edge citation graph, so landmark cases surface ahead of incidental keyword matches
- >Semantic Search: bge-base embeddings over opinion leads and chunked full text (FAISS / HNSW), fused with BM25 to answer questions keyword search can't
- >Statutes as Primary Authority: The U.S. Code, the CFR, and the statutes of 9 states (CA, NY, FL, PA, DE, OH, VA, TX, IL; 513K sections) in the same vector space, so answers lead with controlling statutory text
- >Treatment Signals: Direction-aware detection of "overruled by" / "abrogated by" so answers flag cases that may no longer be good law
- >Grounded Answers: Claude answers strictly from retrieved sources and cites every proposition
- >MCP Server:
verify_citations,resolve_citation,search_caselaw,get_case,check_treatment,search_statutes,lookup_statute,ask_caselaw, for Claude, ChatGPT, Cursor, or any MCP agent - >Metered API Gateway: API-key metering live at api.dicta.wire.band since June 24, 2026
How It Was Built
Legal Search v1 (June 10, 2026) indexed all 10.3M opinions on a CPU-only host. The deep-coverage pass, chunked full-text embeddings across the corpus, ran on rented GPUs in about four hours for roughly $4. The big indexes are memory-mapped, which took the serving heap from 41.7 GB to 0.8 GB and startup from hours to 11 seconds.
Tech Stack
Python / FastAPI search and RAG services, tantivy, FAISS / HNSW, bge-base-en-v1.5, PageRank citation authority, Claude for grounded answers, an MCP adapter, and a metering gateway. Built entirely on public-domain court opinions and statutes.