Deep Research

Deep Research

Run an external research loop that turns a broad question into a cited dossier.

Loop: frame -> angle queries -> research threads -> gap queries -> rank -> synthesize.

Use this for research that needs source quality, disagreement, and follow-up rounds. Do not use it for one factual lookup.

0. Choose depth

Default to light unless the user asks otherwise.

If depth changes the cost or runtime materially, ask once:

Research at fast, light, or deep depth?

Completion: depth is chosen and the query count for each round is known.

1. Frame the request

Extract:

• Topic
• User's goal
• Constraints: time range, geography, industry, source types, exclusions
• Seed links or named sources
• 3–5 themes
• Named entities: people, companies, products, papers, projects, laws, competitors

Completion: you can say what decision, article, comparison, or understanding the research should support.

2. Create angle queries

Do not search the same phrase three ways. Create distinct angles.

Use these angle types:

• Direct: the obvious topic phrase
• Adjacent: parent categories, enabling technologies, related concepts
• Entity: named companies, authors, tools, papers, competitors
• Evidence: benchmarks, case studies, filings, datasets, pricing, adoption, incidents
• Counterpoint: criticism, failures, limitations, risks, skepticism
• Operator: implementation details, workflows, how people actually use it

For each query, write:

{"query":"...","angle":"direct|adjacent|entity|evidence|counterpoint|operator","why":"what this angle should uncover"}

Completion: round 1 has exactly the required number of non-overlapping queries.

3. Run discovery threads

For each query, run one research thread. Use parallel agents/tool calls if available; otherwise run them sequentially.

Each thread:

1. Searches the web for its query.
2. Opens the strongest sources, not just snippets.
3. Prefers primary sources: docs, papers, standards, filings, repos, changelogs, datasets, talks, interviews.
4. Uses credible secondary sources for context: journalism, analyst reports, explainers, technical blogs, user discussions.
5. Avoids SEO filler unless it points to a better source.
6. Returns at most 10 findings.

Finding shape:

{
  "title": "...",
  "url": "https://...",
  "source_type": "paper|docs|github|news|blog|report|video|forum|dataset|other",
  "evidence_quality": "primary|secondary|anecdotal|unknown",
  "relevance": "high|medium|low",
  "summary": "2-3 sentences on what this source contributes",
  "key_concepts": ["..."],
  "claim_to_verify": "optional"
}

Completion: every query produced findings, or you can state which query failed and why.

4. Gap-check between rounds

After each round except the last, deduplicate the findings so far and identify what is still missing.

Look for gaps:

• No primary source
• No counter-evidence
• Too much vendor language
• Missing implementation detail
• Missing adoption, market, pricing, or benchmark data
• Contradictory claims
• Outdated sources
• Newly discovered entity worth chasing

Generate exactly the next round's query count. Follow-up queries must target named gaps, not repeat round 1.

Completion: each follow-up query states the gap it is meant to close.

5. Rank sources

Deduplicate by canonical URL and obvious mirrors.

Score:

• 1.0 seed source from the user
• 0.9 high-relevance primary source
• 0.8 high-relevance secondary source
• 0.7 medium-relevance primary source
• 0.5 medium-relevance secondary source
• 0.2 low-relevance or anecdotal source

Keep the top 8–12 sources with detail. Demote the rest to one-line notes.

Completion: sources are ranked, duplicates are collapsed, and weak sources are not treated like strong evidence.

6. Write the dossier

Use this structure:

# <Topic> — Research Dossier

> One-line answer: what the research found and why it matters.

## Executive summary
The short synthesis. Include the main conclusion, strongest evidence, and biggest uncertainty.

## Research angles
- <query> — <angle> — <why it was included>

## Key findings
- <finding> — source: <title/link>

## Evidence quality
- **Well-supported:** <claims backed by primary or multiple credible sources>
- **Promising but unsettled:** <claims with partial evidence>
- **Weak or hype:** <claims that are vendor-led, anecdotal, or unsupported>

## Practical implications
What the user should do, compare, build, buy, avoid, monitor, or research next.

## Sources ranked
1. <Title> — <score> — <source_type> — <url> — why it matters

## Open questions
- <question> — what would answer it

Completion: every major claim links back to a source, uncertainty is explicit, and the output is synthesis rather than a link dump.

Failure modes

• One-pass summary: stop and run gap queries.
• Query clones: replace with distinct angle queries.
• No primary sources: add a docs/paper/filing/repo/dataset query.
• No skepticism: add a counterpoint query.
• Vendor fog: separate shipped facts from marketing claims.
• Link dump: rank and synthesize before answering.
• Fake confidence: label uncertainty and open questions.