Legal Assistant

Citation-grounded Q&A across your firm’s Vault. Lives at /admin/legal/research. Same engine that powers the Practice agents and Bots.

Asking a question

Pick a session from the left rail or click New session. Type a question; the answer streams back with inline [#N] citations that link to evidence cards on the right panel.

Mic button (next to send) does voice input via the browser’s native SpeechRecognition API — Chrome, Safari, Edge supported; Firefox not.

Hybrid retrieval

question  ─►  embed       ─►  pgvector cosine  top-K
              │
              └►  websearch_to_tsquery  ─►  ts_rank_cd  top-K
                                            │
                          Reciprocal Rank Fusion (k=60)  ─►  top-8 chunks
                                            │
                          Matter-team ACL gate
                                            │
                          LLM synthesize with inline [#N] markers

Knowledge sources

Above the input is a chip row toggling Vault / Firm templates / Federal caselaw. Default is Vault only. See Knowledge sources.

Citations

Each answer ships with up to 8 citation cards: provenance badge (VAULT / TEMPLATE / CASELAW), document title, chunk index, excerpt, and an Open source ↗ link for off-Vault sources. Clicking [#N] in the answer scrolls the corresponding card into view.

Why?

Every assistant message has a small Why? badge. Click it to open the right-edge trace panel — total latency / cost, per-event timing, model, and the retrieval breakdown (chunks per source, fine-tune flag, etc.). See Why-Engine trace.

No fabrication contract

The system prompt enforces:

Inline [#N] citation required for every claim
“If the chunks don’t answer the question, say so plainly”
No invented case citations

When the vault returns no chunks the Assistant says so and stops — it doesn’t fall back to model knowledge.