Deep Research

MANDATE: All research queries MUST load exa + valyu backends for comprehensive discovery.

Intelligent research agent with three modes: quick search, query expansion, and deep multi-query synthesis. Uses Tavily's AI-optimized search with smart stopping.

Required Backends

Before starting any research:

1. Load skill://exa-plus - Neural web + GitHub search
2. Load skill://valyu - Deep research + time travel analysis
3. Use lev find --scope=research for integrated search

Example:

# Load exa + valyu
lev find "authentication patterns 2026" --scope=research

# Augment with Tavily deep search
bun {baseDir}/scripts/research.mjs "authentication patterns 2026" --deep

Usage

# Quick mode - single query, fast results
bun {baseDir}/scripts/research.mjs "query" --quick

# Expand mode - iterative query refinement (default)
bun {baseDir}/scripts/research.mjs "query" --expand

# Deep mode - multi-query synthesis with comprehensive analysis
bun {baseDir}/scripts/research.mjs "query" --deep

Modes

Quick Mode (--quick)

Single Tavily search. Fast, returns top results with AI-generated answer.

• Best for: Simple factual questions, quick lookups
• Iterations: 1

Expand Mode (--expand) [default]

Iterative query refinement. Analyzes initial results and generates follow-up queries.

• Best for: Exploratory research, learning about a topic
• Iterations: 2-3 (stops when confident or no new info)

Deep Mode (--deep)

Multi-query parallel search with synthesis. Generates multiple angle queries, searches in parallel, and synthesizes findings.

• Best for: Comprehensive research, due diligence, complex topics
• Iterations: Up to 5 (configurable)

Options

• --quick: Quick single-query mode
• --expand: Iterative expansion mode (default)
• --deep: Deep multi-query synthesis mode
• --max-iter <n>: Maximum iterations (default: 5)
• --confidence <n>: Confidence threshold 0-100 (default: 85)
• --results <n>: Results per query (default: 5, max: 20)
• --topic <t>: Search topic - general (default) or news
• --json: Output raw JSON instead of markdown

Smart Stopping

The agent stops early when:

1. Confidence threshold reached - Sources consistently agree
2. No new information - Follow-up queries return redundant results
3. Max iterations hit - Safety limit reached

Output Format

# Research: [Query]

## Summary
[Synthesized findings with confidence score]

## Key Findings
- Finding 1
- Finding 2
...

## Sources
- [Title](url) - relevance%
...

## Research Trace
- Iteration 1: [query] → [n] results
- Iteration 2: [follow-up] → [n] new results
...

Configuration

API key is read from (in order):

1. TAVILY_API_KEY environment variable
2. ~/.clawdbot/clawdbot.json → skills.entries["tavily-search"].apiKey

Examples

# Quick fact check
bun {baseDir}/scripts/research.mjs "What is the current population of Tokyo?" --quick

# Explore a topic
bun {baseDir}/scripts/research.mjs "How does CRISPR gene editing work?"

# Deep research for decision making
bun {baseDir}/scripts/research.mjs "Best practices for Kubernetes autoscaling in production" --deep

# News research
bun {baseDir}/scripts/research.mjs "AI regulation updates 2024" --topic news --deep

Related Search Tools

Choose the right tool for your task:

Integration pattern:

# 1. Quick lookup
brave-search "keyword" --num 5

# 2. Deeper research
deep-research "keyword context" --deep

# 3. Recursive refinement
valyu research "keyword context" --turns 5 --threshold 0.85

# 4. Multi-perspective
lev-research "keyword" --template=technology_assessment

Notes

• Uses Tavily's advanced search depth for better results in expand/deep modes
• Deduplicates sources across iterations
• Tracks information density to detect diminishing returns
• Outputs structured markdown optimized for AI consumption