RAG & Retrieval

System prompt that forbids parametric knowledge, requires a cited source span for every claim, and defines explicit behaviour when the retrieved context does not cover the question.

Analyses a legal document type and produces a justified chunking strategy — boundaries, overlap, metadata, and edge cases — instead of arbitrary fixed-size splitting.

Generates an evaluation set from a given corpus, including questions that should fail, to measure retrieval and faithfulness rather than only happy-path answering.

A judge prompt that decomposes an answer into atomic claims and verifies each against the retrieved chunks, returning a faithfulness score plus the list of unsupported claims.

Turns a client's natural-language question into several retrieval queries in legal terminology, with synonyms and metadata filters.

Reranks candidate chunks by true relevance to the question and filters noise before the context reaches the generator, with a reason per decision.

Identifies conflicts among retrieved legal sources — versions, amendments, contradictory clauses — and classifies the conflict type instead of silently picking one.

Turns retrieved chunks into references a user can verify independently — document, clause, page — and explicitly flags what cannot be verified.

Extracts structured metadata fields from a legal document to enable retrieval filtering, with confidence levels separating what was found from what was inferred.

Separates retrieval, ranking, indexing, and generation failures so the fix targets the right layer instead of endlessly tweaking the prompt.

Balances lexical and semantic search for legal content, where clause numbers and defined terms need exact matching while conceptual questions need semantic similarity.

Identifies and replaces personal and privileged client information in legal text before it reaches an external model, preserving legal meaning and reversibility.

A system prompt separating legal information from legal advice, defining exactly when to refuse, what to say, and where to escalate.