Skip to main content
When you ask a question, the system has to decide which passages from your documents and which cases from the corpus are most relevant. Different questions call for different strategies. You can let the system pick automatically or override it per request.

​
The Six Strategies

StrategyBest ForWhat It Does
Hybrid (default)General researchBalances meaning-based search with exact-keyword matching - good middle ground for most questions
Self-CritiqueAdaptive queriesDecides whether retrieval is even needed; checks its own answer and re-queries if confidence is low
CorrectiveHigh-stakes analysisVerifies that the retrieved passages are actually relevant before letting the model use them
HypotheticalAmbiguous queriesGenerates a draft answer first, then searches for sources that match the draft
Multi-QueryBroad researchRewrites your question into three variations and pools the results
Multi-HopComparative reasoningRuns iterative searches that build on each other, up to three rounds

​
At a glance

Hybrid

Default. Semantic + keyword. The safe everyday choice for most research questions.

Self-Critique

Adaptive. Decides whether retrieval is needed and re-runs if the first answer is weak.

Corrective

Defensive. Grades each retrieved passage for relevance and drops the noise before drafting.

Hypothetical

Drafts first, searches second. Best for vague or open-ended questions.

Multi-Query

Widest net. Rewrites your question into variations and pools the results.

Multi-Hop

Iterative. Builds searches on top of earlier rounds for comparative questions.

​
When To Use Each

​
Hybrid (default)

The everyday choice. Combines semantic understanding (so “vehicle” matches “automobile”) with exact-keyword matching (so a statute cite or party name gets found verbatim). Use this unless you have a specific reason to switch.

​
Self-Critique

Good when you do not know whether the question needs retrieval at all. The system might answer “what is res judicata?” from general knowledge but search the corpus for “has the Ninth Circuit applied res judicata to administrative decisions?” - Self-Critique makes that call automatically and re-runs if the first answer is weak.

​
Corrective

The most defensive option. Before passing retrieved passages to the model, the system grades each one for relevance and tosses out passages that are tangential. Use this when the cost of citing the wrong authority is high - regulatory filings, appellate briefs, opinion letters.

​
Hypothetical

When the question is vague, the system drafts a placeholder answer and uses that as the search query. This often pulls better sources than searching the literal question. Useful for open-ended drafting questions or when you do not know the right legal vocabulary yet.

​
Multi-Query

The widest net. The system rewrites your question into multiple variations and runs all of them, then deduplicates. Good when you want comprehensive coverage and are willing to wait a few extra seconds. Useful for client-facing memos that need to canvass the field.

​
Multi-Hop

For comparative or compositional questions - “compare X and Y”, “trace how the doctrine evolved”, “find cases where both A and B were at issue”. The system runs round one, sees what is missing, and runs round two and three to fill the gaps.

​
Choosing a Strategy

Three ways to set the strategy:
  1. Per request - Open the settings menu in the chat composer and pick a strategy for this question only
  2. Per matter - Set a default strategy for the matter (e.g., always Corrective for an appellate matter)
  3. Auto - Let the system pick based on the question type
For most users, Auto is fine for day-to-day work and you reach for specific strategies on high-stakes tasks.

​
Trade-Offs

Hybrid is the fastest baseline. Strategies that re-rank or re-search (Corrective, Multi-Query, Multi-Hop) take longer and use more compute in exchange for higher-precision sources. For day-to-day work the difference is rarely noticeable; for high-stakes filings, the extra time is worth it.

​
Tips

Default to Hybrid for routine work. Promote to Multi-Hop or Corrective only when the cost of a wrong cite is high - appellate briefs, opinion letters, regulatory filings.
Pair Multi-Hop with Deep Research. They reinforce each other on comparative questions like “how do the Fifth and Ninth Circuits differ on qualified immunity for excessive-force claims?”
Watch the source list. If a question returns weak sources, switch strategies and re-run rather than re-phrasing the question endlessly. A Hybrid miss is often a Corrective hit.

​
Force a strategy in your prompt

Corrective retrieval - high-precision sources for an appellate-grade question.

Deep Research Mode

Iterative analysis that uses multi-hop retrieval by default.

Answer Verification

Claim-level checking against retrieved sources.

Document Search

How retrieval works against your own uploaded documents.

Agent Mode

Let the agent pick the strategy autonomously per sub-question.