gbalabanli/web-deepsearch

Web Deep Research

Scope

• Research external knowledge from web sources.
• For project-internal codebase questions, use local file analysis workflows instead.
• Prefer primary sources first (official docs, standards, papers, vendor docs).
• Use community and blog sources as supporting evidence, not anchors.

Phase 1: Frame the Problem

1. Rewrite the user question into one decision statement.
2. Define the output type: compare, choose, explain, plan, or risk review.
3. Decompose into 3-5 sub-questions that collectively answer the full request.
4. Note explicit constraints: budget, timeline, scale, compliance, and environment.

Phase 2: Generate the Search Plan

For each sub-question, create 3-5 query variants:

• Baseline query
• Official-doc query (site:docs... or vendor domain)
• Comparison query (X vs Y)
• Risk or limitation query
• Recent-update query (after:YYYY)

Rules:

• Build 12-20 total queries.
• Run in parallel batches of 4-8 queries.
• If results are weak, re-query only missing gaps with targeted follow-ups.

Phase 3: Collect and Rank Evidence

1. Merge all results.
2. Deduplicate by URL, domain, and content overlap.
3. Rank each source on:
   • Credibility (0-10): official docs, standards, papers > vendor blogs > forums.
   • Freshness (0-10): newer sources score higher when topic changes quickly.
   • Relevance (0-10): direct answers with implementation detail score highest.
4. Compute weighted quality score: score = credibility*0.5 + freshness*0.2 + relevance*0.3
5. Keep highest-quality sources and drop low-value duplicates.

Target evidence quality:

• At least 8 solid sources for complex requests.
• At least 50% authoritative sources.
• At least 30% recent sources for fast-moving topics.

Phase 4: Synthesize Findings

For each sub-question:

• Summarize what sources agree on.
• Highlight disagreements and explain why they differ (context, date, assumptions).
• Extract practical implications (cost, complexity, performance, risk).

Cross-question synthesis:

• Connect findings across sub-questions.
• Identify decision drivers and blocking unknowns.
• Separate "known with confidence" from "needs validation".

Phase 5: Produce Decision-Ready Output

Use this structure:

# Deep Research: <Question>

## Executive Summary
2-3 short paragraphs with the main answer and rationale.

## Research Overview
- Sub-questions: <n>
- Queries executed: <n>
- Sources used: <n> (<authoritative %>, <recent %>)
- Iterations: <n>

## Findings
### 1) <Sub-question>
Narrative synthesis with citations [1][2].
- Key insight
- Key insight
- Key insight

### 2) <Sub-question>
...

## Synthesis
- Consensus points
- Contradictions and context
- Remaining uncertainty

## Recommendations
### Critical
1. <Action + why + evidence>
2. <Action + why + evidence>

### Important
1. <Action + why + evidence>
2. <Action + why + evidence>

### Optional
1. <Action + why + evidence>

## Sources
[1] Title - Organization, Date. URL
[2] Title - Organization, Date. URL

Output rules:

• Every factual claim must map to citations.
• Recommendations must be actionable and prioritized.
• Explicitly list assumptions when evidence is incomplete.

Phase 6: Refine

Run at least one gap check before finalizing:

• Any sub-question with fewer than 3 meaningful citations?
• Any major contradiction not explained?
• Is guidance still ambiguous for next-step execution?
• Are recommendations missing risk or rollback notes?

If gaps remain:

1. Run targeted re-queries.
2. Update findings and recommendations.
3. Publish with a clear "Research Gaps" section if uncertainty remains.

Citation Rules

• Use numbered citations in-line: [1], [2], [3].
• Keep a source list at the end with full URL and date.
• Prioritize primary sources for key claims.
• If citing secondary sources, cross-check with at least one independent source.

Quality Bar

Deliver only when all are true:

• Sub-questions are fully covered.
• Evidence quality thresholds are met, or exceptions are explained.
• Output provides a clear decision path, not only a summary.
• Tradeoffs and risks are explicit.
• Sources are transparent and verifiable.