Deep Research Mode

Multi-hop reasoning for cross-document analysis and comparative legal questions

Deep Research Mode is built for the questions that take an associate a full afternoon - the ones where you need to read multiple cases, cross-reference statutes, and synthesize a position. It runs an iterative research loop that drafts an answer, identifies what is missing, runs follow-up searches, and refines.

When To Use It

Use Deep Research for questions that need more than one pass:

• “Compare how the Second and Ninth Circuits have treated implied warranty disclaimers in software licenses”
• “Walk me through the elements of promissory estoppel in California and identify which our facts satisfy”
• “What is the current status of personal jurisdiction over foreign defendants after the recent Supreme Court decisions?”
• “Analyze this complaint and tell me which causes of action are strongest given the case law”

For simpler questions (“what does Section 230 say?”), standard mode is faster and cheaper.

How It Works

When Deep Research is on, the system:

1. Plan

   Reads your question and generates an initial research plan.

2. Round 1 retrieval

   Runs the first round of searches across your documents, case law, and statutes.

3. Draft

   Produces a preliminary answer from the retrieved sources.

4. Gap analysis

   Reviews the draft for missing authorities, weak reasoning, and counter-arguments.

5. Follow-up rounds

   Generates targeted queries for the gaps; repeats up to three rounds.

6. Synthesis

   Returns a final answer with every source cited and verification badges applied.

Higher Retrieval Limits

Deep Research pulls a larger pool of candidate sources before narrowing down - so the model can compare more authorities and pick the best ones rather than working with the first few that match.

Setting	Standard Mode	Deep Research
Candidate sources per query	Standard pool	Expanded pool
Follow-up queries	None	Up to 3 rounds
Cross-document synthesis	Single pass	Iterative
Best for	Direct questions	Comparative analysis

Gap Analysis Between Rounds

After each round, the system asks itself: “What is this draft missing? Are there counter-authorities? Is there a temporal issue? Are the cases still good law?” The follow-up queries it generates are aimed at those gaps, not just at re-searching the original question.

You can see the intermediate reasoning steps in the response - what was searched, what was found, and what was decided.

Hallucination Guard

Each round runs through claim-level verification before the answer is returned. If the model writes a sentence that is not actually supported by the retrieved sources, the system flags it and re-runs that section. See Answer Verification for how this works.

When To Turn It Off

• Quick lookups (“what is the citation for Marbury v. Madison?”)
• Drafting tasks that do not need research (“draft me an NDA”)
• Conversational follow-ups inside an existing thread

Deep Research is slower and burns more credits per question than standard mode. Reserve it for the questions where the depth genuinely pays off - comparative analysis, unsettled doctrine, fact-driven motion practice.

Deep Research is slower and more expensive per question. Use it where the depth pays off.

How To Enable

Click the Deep Research toggle in the chat composer. The toggle saves per matter, so if you turn it on for a litigation matter, it stays on while you work in that workspace.

Tips

Be specific about jurisdiction. “California state courts” produces a tighter answer than “California”. For federal questions, name the circuit.

Anchor to your facts. Paste the key facts of your matter into the question; Deep Research will tie the analysis directly to them instead of returning a textbook summary.

Run inside a matter. When the matter has documents uploaded (complaint, contracts, deposition transcripts), Deep Research pulls from them alongside the public corpus.

Deep Research prompt

Deep Research - comparative circuit analysis applied to your facts, with counter-authorities surfaced.

CopiedCopy prompt

Related

Agent Mode

Tool-using research with web search and URL fetching.

Answer Verification

How claims are checked against sources at every round.

Retrieval Strategies

The different ways the system finds relevant sources.

Citation Graph

Confirm cited cases are still good law before relying on them.