Research

Research assistant

Ask legal questions with live citations, multi-jurisdiction support, and real-time activity tracking

The Research assistant is the engine at the centre of 8rney. It runs parallel searches across live legal databases, extracts authorities, and synthesises a cited answer — while showing you exactly what it is doing at every step.

How it works#

When you submit a research question, the system:

  1. Classifies the question by jurisdiction and legal domain
  2. Plans a multi-step query across the relevant sources
  3. Fetches and ranks authorities in parallel
  4. Synthesises a cited answer with confidence levels
  5. Generates follow-up questions for further exploration

You see a live activity feed throughout — "Searching Indian Kanoon…", "Extracting case details…", "Ranking authorities…" — so you know what is happening and can judge the quality of the result before using it.

Research pipeline modes#

The mode picker in the composer controls how deeply the AI reasons before it answers. Select the mode that fits the urgency and complexity of your question.

Fast

Low reasoning effort. Quick orientation answers, checking a basic rule, or confirming a known position. Returns in seconds.

Balanced

Medium reasoning. The right default for most day-to-day queries where you want a solid cited answer without the wait.

Deep

High reasoning effort. Best for complex multi-issue questions, conflicts of authority, or any issue where you need a thorough single-model pass.

Expert

Maximum reasoning in one pass. Use for the hardest questions — major deals, contested litigation points, and issues where a missed authority carries real risk.

Note

When using the Claude Opus provider, a fifth Max mode is available. Max triggers direct Bedrock maximum-thinking on Opus 4.6 — the deepest reasoning tier available on the platform. Reserve it for your most demanding research tasks.

Mode–provider matrix#

ModeGPT-5.5 (Azure)Claude Sonnet 4.6Claude Opus 4.6
FastLow reasoningLow adaptive thinkingLow adaptive thinking
BalancedMedium reasoningMedium adaptive thinkingMedium adaptive thinking
DeepHigh reasoningHigh adaptive thinking + summariesHigh adaptive thinking
Expertxhigh reasoningMax adaptive thinkingExtra-high adaptive thinking
MaxDirect Bedrock max thinking

Writing a strong research question#

1

State the issue plainly

Lead with the legal question, not the backstory. "What are the conditions for specific performance under Indian contract law?" is better than a paragraph of context.

2

Set the jurisdiction

Use the jurisdiction selector in the composer to specify Indian, Canadian, or another legal system. This routes the query to the right sources and citation format.

3

Attach relevant documents

Type @ in the composer to open the file mention menu. Select any uploaded document or Drive file to include it as context. The AI will read and cite from your own files alongside external databases.

4

Specify the output format

Ask for a memo outline, risk list, clause suggestions, or a direct answer so the response lands in a format you can act on immediately.

Sources#

JurisdictionSources
IndiaIndian Kanoon (case law), India Code (legislation), Indian courts
CanadaCanLII (case law and legislation)
GeneralAI reasoning from training data with uncertainty flags

Citation quality#

The Research assistant uses a verified-authorities system — a curated database of known authoritative cases per jurisdiction and area of law. When it finds a matching authority, it cites it with court, date, and citation string. When it cannot find a specific case, it says so rather than fabricating a citation.

Note

Anti-hallucination guardrails are on by default. The AI is instructed to flag uncertain citations with "I cannot verify this specific citation" rather than inventing details.

Follow-up suggestions#

At the end of every research response, the assistant suggests up to four follow-up questions. These are generated from the context of the current answer — not generic suggestions. Click any to continue the research thread.

Follow-up suggestions are filtered to stay within legal topics. If the query or answer contains no legal signal, the assistant returns a set of legal-domain starters instead of off-topic suggestions.

Stream recovery#

If the model returns no answer text on a given turn, the platform automatically attempts a non-streaming recovery pass before surfacing an error. If the recovery pass also fails, you will see the retrieved sources the system found — so the work is not lost and you can retry with a narrower question.

Outputs you can reuse#

Research summary

A cited answer you can share internally as a first-pass research note.

Memo scaffold

A structured outline ready for a lawyer to expand into advice or a brief.

Risk list

Issue-spotting bullets for contract review or matter planning.

Drafting handoff

Language suggestions that move directly into the Document editor.