| name | deep-research |
| description | IterDRAG iterative retrieval-augmented generation pattern — search, summarize, reflect loop for deep research tasks |
| tags | ["research","search","rag","deep-research","analysis"] |
Deep Research — IterDRAG Pattern
When assigned a research task, follow this iterative loop until the answer is comprehensive and well-sourced.
The Loop
QUERY → DECOMPOSE → SEARCH → SUMMARIZE → REFLECT → (gaps?) → SEARCH again → REPORT
Stop when: all sub-questions have authoritative answers, OR 3 iterations complete, OR new searches return no new information.
Phase 1: Decompose the Query
Before searching, break the research question into 3–5 independently searchable sub-questions that together cover the full topic.
Example — "What are best practices for Connect-RPC in Go?":
- What is Connect-RPC and how does it differ from standard gRPC?
- What are the recommended project structure conventions?
- How should error handling work in Connect-RPC handlers?
- What are the performance best practices for streaming RPCs?
- How does Connect-RPC handle backward compatibility?
Phase 2: Search (per sub-question)
For each sub-question:
- Formulate a precise search query (use technical terms, avoid ambiguity).
- Execute search via available tool (web search, document retrieval, code search).
- Collect 3–5 relevant sources. Prefer: official docs > source code > RFC/specs > engineering blogs > forum posts.
- Note the URL/reference, title, and key claims from each source.
Phase 3: Summarize
After searching all sub-questions:
- Write a structured summary per sub-question.
- Cite sources inline for every claim.
- Flag contradictions between sources explicitly.
- Mark what is authoritative (spec, official docs) vs. community opinion.
Phase 4: Reflect on Gaps
Ask:
- What questions remain unanswered?
- What claims lack good sourcing?
- Are there edge cases or failure modes not covered?
- Is the information current? (check publication dates)
If gaps found → formulate new targeted queries and return to Phase 2.
If confident → proceed to Phase 5.
Phase 5: Produce the Report
# [Topic] — Research Report
## Summary
[2–3 sentences: key findings]
## Findings
### [Sub-question 1]
[Answer with inline citations]
### [Sub-question 2]
[Answer with inline citations]
## Key Recommendations
- [Actionable recommendation 1]
- [Actionable recommendation 2]
## Sources
- [URL or doc name] — [why relevant]
## Gaps and Caveats
- [What remains uncertain or needs verification]
Configuration
| Mode | When to use |
|---|
| Local (Ollama) | Private data, offline, cost-sensitive. Models: llama3, mistral, deepseek-r1 |
| Cloud | High accuracy on complex topics. Models: claude-sonnet-4-6, gpt-4o |
Search providers (no key required): DuckDuckGo instant answers.
Search providers (API key): Brave Search API, Tavily API, SerpAPI.
Quality Signals
Good output:
- Every claim has a source
- Contradictions are surfaced, not hidden
- Distinguishes spec/official vs. community practice
- Gaps section is honest
Poor output:
- Confident claims without sources
- No alternatives or tradeoffs discussed
- Single-source conclusions on contested topics
- Missing a gaps/caveats section
