Deep Research

PHASE 1: ITERATIVE MULTI-ENGINE SEARCH (Steps 1-10)

Search Strategy

Execute up to 10 searches across different engines. Rotate engines to avoid rate limits.

Engine Rotation Pattern:

Gap Analysis Between Searches

After each search round:

1. Analyze results - what aspects are covered?
2. Identify gaps - what's missing?
3. Refine query - add subtopics, different keywords
4. Choose next engine
5. Execute next search with modified query

Query Evolution Example:

Initial: "quantum computing applications"
After round 1: "quantum computing applications 2025" (add year)
After round 3: "quantum computing healthcare finance applications" (add domains)
After round 5: "quantum computing drug discovery market" (specific use cases)

Web-Search Operations Usage

Search Execution Rules

1. Never repeat same engine twice in a row (rate limit protection)
2. Use fetchGithubReadme when GitHub repos appear - high priority for code examples
3. Use fetchCsdnArticle/fetchJuejinArticle when Chinese resources found
4. Use fetchWeb for specific URLs that need detailed fetching
5. Track all sources - they may be used later
6. Stop early if saturation reached (no new relevant results)

PHASE 2: DOCUMENT PROCESSING

For each document type found during search:

Important: Process documents during search phase, not after. This gives better context for query refinement.

PHASE 3: SOURCE SELECTION

Deduplication

• Combine all sources from all 10 searches
• Remove exact duplicates
• Merge near-duplicates (same content, different URLs)

Scoring & Ranking

• Score each source by relevance to topic
• Consider: authority, recency, depth, uniqueness
• Rank all sources

Final Selection

• Select top 3-5 sources for deep analysis
• Prioritize sources with:
  • Comprehensive content
  • Authoritative authorship
  • Recent publication
  • Unique insights

PHASE 4: DEEP CONTENT FETCHING (Steps 11-15)

For EACH selected source:

1. Fetch full content using appropriate tool
2. Assign unique SOURCE_ID (e.g., OpenAI2024, Microsoft2025)
3. Extract key information
4. Create detailed note with:
   • Main points
   • Data/statistics (with SOURCE_ID)
   • Quotes (with SOURCE_ID and page number)
   • Unique insights
   • Analysis

PHASE 5: SYNTHESIS (Step 16)

• Compare findings across sources
• Identify patterns and consensus (with SOURCE_IDs)
• Note conflicts and gaps
• Use academic citation format

PHASE 6: FINAL REPORT (Step 17)

• Generate comprehensive report
• Include all findings with in-text citations
• Provide References section
• Academic writing style

Session Structure

Research is saved to: sessions/[research-topic]/

sessions/
└── [topic]/
    ├── 01-search-results.md      # All 10 searches with sources
    ├── 02-document-processing.md # Converted docs from markitdown
    ├── 03-source-selection.md   # Deduplicated & ranked sources
    ├── 04-source-1.md           # Deep analysis with SOURCE_ID
    ├── 05-source-2.md           # Deep analysis with SOURCE_ID
    ├── 06-source-3.md           # Deep analysis with SOURCE_ID
    ├── 07-source-4.md           # Deep analysis with SOURCE_ID
    ├── 08-source-5.md           # Deep analysis with SOURCE_ID
    ├── 09-synthesis.md          # Cross-reference with citations
    └── 10-final-report.md       # Final report with References

Source Note Format (Academic Style)

---