Research Assistant Deep

Deep research and synthesis assistant. Use when you need to investigate a topic across multiple sources, compare findings, produce structured research reports, or do multi-step web research with citation tracking. NOT for simple single-query lookups (use web_search directly).

Research Assistant

Structured multi-source research with citation tracking and synthesis.

When to Use

• Multi-step investigation across 3+ sources
• Comparative analysis requiring side-by-side findings
• Reports needing citations and source attribution
• Topic exploration where depth matters more than speed

Workflow

1. Scope

Clarify the research question. Break broad topics into 2-5 sub-questions. Define the output format upfront (brief, detailed, table, report).

2. Search

Run 2-3 parallel web_search calls with varied query phrasing per sub-question. Use freshness and country filters when recency or region matters.

3. Fetch & Extract

For each promising result, use web_fetch with extractMode: "markdown" and maxChars: 8000. Skip paywalled or low-value pages quickly.

4. Synthesize

Cross-reference findings across sources. Flag contradictions explicitly. Assign confidence levels:

Level	Meaning
High	3+ independent sources agree
Medium	2 sources or 1 authoritative source
Low	Single unverified source

5. Cite & Deliver

Every factual claim gets a source line: [N] Source Title — URL

Output format:

text

## [Topic]

### Key Findings
- Finding 1 [1]
- Finding 2 [2,3]

### Contradictions
- Source A says X [1], Source B says Y [2]

### Sources
[1] Title — https://...
[2] Title — https://...

Tips

• Vary queries: rephrase, use synonyms, add site: filters
• Go deep on one source before moving on: don't skim 20 pages
• Track dead ends: note queries that returned nothing to avoid repeating
• Set maxChars conservatively: 4000-8000 per fetch; increase only if needed
• Parallelize searches, serialize fetches (respect rate limits)

Advanced Patterns

For multi-phase research, see references/deep-dive.md.

For automated source credibility scoring, see scripts/credibility.py.