Deep Research

Deep Research

Autonomous iterative research agent. Recursively explores a topic tree: generates search queries → retrieves sources → extracts learnings → generates follow-up questions → goes deeper. Each depth level halves the breadth (exponential focus). All learnings accumulate into a final cited report.

Algorithm: Based on dzhng/deep-research (18K⭐, by David Zhang/Aomni), fused with Exa deep search, Serper SERP, subagent parallelism, and scientific rigor from K-Dense-AI.

When to Use

• User says "deep research", "investigate", "deep dive", "thorough analysis"
• Any question requiring 15+ sources and multi-angle synthesis
• Technology evaluations, competitive analysis, market sizing
• "What's the current state of X?" with real depth
• NOT for quick lookups (use research or exa-search directly)

Parameters

Higher breadth = wider coverage. Higher depth = more follow-up drilling. Total searches ≈ breadth × (1 + breadth/2 + breadth/4 + ...).

Core Algorithm

deepResearch(query, breadth=4, depth=3, learnings=[]):
  1. Generate `breadth` SERP queries from query + accumulated learnings
  2. For each query (parallel, concurrency=3):
     a. Search via Exa deep + Serper (dual-source)
     b. Extract learnings + follow-up questions from results
     c. If depth > 1:
        nextQuery = research_goal + follow_up_questions
        deepResearch(nextQuery, breadth/2, depth-1, allLearnings)
     d. Else: return learnings + URLs
  3. Deduplicate learnings, return all

Execution

Step 0: Clarify (optional)

If the topic is vague, ask 1-2 sharpening questions:

• "What's your goal — learning, deciding, or writing?"
• "Any specific angle or time period?"

If user says "just do it" — proceed with defaults.

Step 1: Plan Research Tree

Break the topic into research sub-questions (these become the initial breadth queries):

Topic: "State of nuclear fusion energy in 2026"
→ Sub-questions:
  1. What fusion approaches are closest to net energy gain?
  2. Which companies have the most funding and progress?
  3. What are the engineering bottlenecks?
  4. What's the regulatory landscape?

Step 2: Execute the Recursive Search

For EACH sub-question at each depth level, run dual-source search:

Exa Deep Search (semantic, high-quality)

EXA_API_KEY=$(pass api/exa 2>/dev/null)
curl -s -X POST "https://api.exa.ai/search" \
  -H "x-api-key: $EXA_API_KEY" \
  -H "content-type: application/json" \
  -d '{
    "query": "<sub-question>",
    "type": "deep",
    "numResults": 5,
    "effort": "base",
    "contents": {
      "text": { "maxCharacters": 5000 },
      "highlights": true,
      "summary": true
    },
    "excludeDomains": ["wikipedia.org"],
    "startPublishedDate": "2025-01-01"
  }' | jq '{results: [.results[] | {title, url, summary, highlights, text}]}'

Serper Google SERP (factual, current)

SERPER_API_KEY="${SERPER_API_KEY:-$(pass api/serper2 2>/dev/null || pass api/serper 2>/dev/null)}"
curl -s -X POST "https://google.serper.dev/search" \
  -H "X-API-KEY: $SERPER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"q": "<sub-question>", "num": 8}' | \
  jq -r '.organic[] | {title, link, snippet}'

For academic topics, add Scholar

curl -s -X POST "https://google.serper.dev/scholar" \
  -H "X-API-KEY: $SERPER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"q": "<sub-question>", "num": 5}' | jq '.organic[]'

Step 3: Extract Learnings

After each search batch, extract structured learnings. For each result set:

What to extract:

• Concrete facts with numbers, dates, names (not vague summaries)
• Entities: people, companies, products, institutions
• Exact metrics, statistics, dollar amounts
• Contrarian or surprising findings
• Contradictions between sources

What to generate:

• 3-5 learnings per search (dense, specific, entity-rich)
• 2-3 follow-up questions that go deeper based on what was found

Step 4: Recurse Deeper

For each follow-up direction, construct the next query:

Next query = "Previous goal: <research_goal>
Follow-up directions:
- <follow_up_question_1>
- <follow_up_question_2>"

Recurse with breadth = ceil(breadth/2) and depth = depth - 1.

This naturally focuses: level 1 is broad (4 queries), level 2 is narrower (2 queries per branch), level 3 is laser-focused (1 query per branch).

Step 5: Deep-Read Key Sources

After all recursion completes, pick the 3-5 most important URLs and extract full content:

# Exa content extraction
curl -s -X POST "https://api.exa.ai/contents" \
  -H "x-api-key: $EXA_API_KEY" \
  -H "content-type: application/json" \
  -d '{
    "urls": ["<url1>", "<url2>", "<url3>"],
    "text": { "maxCharacters": 10000 },
    "summary": true,
    "highlights": { "query": "<original_topic>" }
  }'

Step 6: Synthesize Final Report

Combine ALL accumulated learnings into a comprehensive report:

# <Topic>: Deep Research Report
*Generated: <date> · Sources: <N> · Depth: <D> levels · Confidence: High/Medium/Low*

## Executive Summary
3-5 sentences. The answer, not the methodology.

## 1. <Major Theme>
Findings with inline citations: "Company X raised $2B ([Source](url))"
- Key data points with attribution
- Cross-referenced claims

## 2. <Major Theme>
...

## 3. <Major Theme>
...

## 4. Contrarian Views & Uncertainties
What one source says that others don't. Flag speculation.

## Key Takeaways
- Actionable insight 1
- Actionable insight 2
- Actionable insight 3

## Research Gaps
What couldn't be found. What needs primary research.

## Sources
1. [Title](url) — one-line summary
2. ...

## Methodology
- Queries generated: N
- Sources analyzed: M
- Depth levels: D
- Search engines: Exa (semantic) + Serper (Google SERP)

Parallelism with Subagents

For breadth ≥ 4, use subagent parallelism to research sub-questions concurrently. Each subagent handles one branch of the research tree:

Launch 3-4 subagents in parallel:
  Agent 1: Research sub-questions 1-2 (depth=2)
  Agent 2: Research sub-questions 3-4 (depth=2)
  Agent 3: Cross-cutting themes + contrarian views
Main session: Synthesize all findings into final report

Each subagent searches, reads, extracts learnings, and returns structured findings. The main session deduplicates and synthesizes.

For Maximum Depth: Exa Research API

For very complex topics where you want Exa's own agent to do deep research:

# Fire async research task
curl -s -X POST "https://api.exa.ai/research" \
  -H "x-api-key: $EXA_API_KEY" \
  -H "content-type: application/json" \
  -d '{
    "instructions": "<detailed research question>",
    "model": "exa-research-pro",
    "outputSchema": {
      "type": "object",
      "properties": {
        "findings": { "type": "array", "items": { "type": "object", "properties": {
          "claim": { "type": "string" },
          "source": { "type": "string" },
          "confidence": { "type": "string" }
        }}},
        "summary": { "type": "string" },
        "gaps": { "type": "array", "items": { "type": "string" } }
      }
    }
  }'

# Poll for results
RESEARCH_ID="<from above>"
curl -s "https://api.exa.ai/research/$RESEARCH_ID" \
  -H "x-api-key: $EXA_API_KEY" | jq '{status, output}'

Use this as ONE input alongside your own iterative search — don't delegate everything to it.

Quality Rules

1. Every claim needs a source. Unsourced → flagged or dropped.
2. Cross-reference. Single-source claims get [unverified] tag.
3. Recency bias. Prefer sources < 12 months old. Flag stale data.
4. Entities over vibes. "Revenue grew 40% to $2.1B" > "Revenue grew significantly."
5. Contradictions are gold. When sources disagree, present both sides.
6. Separate fact from inference. Label estimates, projections, opinions.
7. No hallucination. If data wasn't found, say "insufficient data" — never fabricate.
8. Contrarian check. Actively search for opposing views on key claims.

Output

• Short topics (< 3 pages): post in chat
• Long reports: save to .git/reports/<topic-slug>-YYYYMMDD.md
• Always print file:///absolute/path for generated files

Quick Reference: Effort Levels

Examples

"Deep research: state of nuclear fusion energy"
"Investigate Rust vs Go for high-frequency trading systems"
"Deep dive: best strategies for bootstrapping B2B SaaS in 2026"
"Thorough analysis of CRISPR gene therapy competitive landscape"
"Research the current state of AI code editors — who's winning and why"