Discovery & Research for Legal

Discovery and Research workflows ask the system to find the right documents and synthesize what they say, with the citations being more important than the synthesis. Citorum's hybrid retrieval pipeline combines dense vector similarity over the embeddings of every document chunk, full-text search via Postgres' built-in indexing, and metadata filtering against tags, dates, jurisdictions, document types, and access-control attributes. Results are scored, deduplicated, and rank-fused into a context window.

The model produces a research summary; the audit log captures every document that contributed, every citation rendered, and the faithfulness score on every claim. The reviewing analyst — paralegal, compliance officer, claims coder, briefer — clicks through to source spans rather than relying on the summary verbatim. The pattern works because the citations are first-class output, not a footnote.

Legal teams operate under a regime that demands provenance for every claim. Citorum runs inside the firm or in-house environment and indexes the matter files, contracts, depositions, regulatory bulletins, and case-law sources the team already maintains. Retrieval is hybrid — vector and full-text search combined with metadata filtering on matter, jurisdiction, privilege class, and access-control attributes — so a paralegal sees only the documents they're cleared to see.

Every answer carries citations to the specific document spans the model drew from, plus a faithfulness label. Verified — Cite Source for answers grounded in the corpus, Review Recommended for ambiguous responses that an attorney should sanity-check, and Do Not Rely — Consult Expert when the corpus didn't actually support the answer or sources contradicted it. The audit log captures the full lineage so an opposing counsel's discovery request, an internal compliance review, or a malpractice inquiry can reconstruct exactly what was retrieved, what was sent to the model, what came back, and who asked.