Deep Research

Free Perplexity/Gemini-level deep research for any AI coding agent. Combines NotebookLM's Gemini-powered research engine with iterative web search, domain-aware routing, and post-research artifact generation.

Activation

Triggered by /deep-research "topic" or intent detection ("research this", "deep dive into", "investigate").

Argument Parsing

• "topic" — required, the research query
• --domain maritime|legal|academic — optional domain boost
• --continue — resume previous research session
• --compare "X vs Y" — structured comparison mode

If no arguments provided, ask: "What topic should I research?"

Dependency Check

Before starting, determine available capabilities:

1. Try notebooklm status first. If command not found, try python3 -m notebooklm status or python -m notebooklm status. Use whichever works for all subsequent notebooklm commands.

   • If success: Full mode (NotebookLM + WebSearch)
   • If all fail: Fallback mode (WebSearch only)

2. If fallback mode, inform user:

   "NotebookLM not detected. Running in web-search-only mode. For full features (deep research, podcast, video, slides, quiz), install: pip install notebooklm-py && notebooklm login"

3. Store mode for pipeline decisions.

Pre-flight

Present this prompt and WAIT for user response before proceeding:

Research Topic: "{topic}"

Depth:
  1 = Quick (fast mode, 5-10 sources, ~1 min)
  2 = Moderate (fast + 1 WebSearch round, 10-15 sources)
  3 = Standard (deep mode, 20+ sources, ~3 min)  [default]
  4 = Deep (deep + 1 WebSearch round, 30+ sources)
  5 = Exhaustive (deep + 2 WebSearch rounds, 40+ sources)

Source focus:
  A = Everything (web + your notebooks)  [default]
  B = Academic only (papers, journals, research)
  C = News & current events (last 12 months)
  D = Technical docs & specs
  E = Custom (tell me what to focus on)

Enter choices (e.g. "3, A") or press Enter for defaults:

Parse response:

• Default if empty: depth=3, focus=A
• Extract depth (1-5) and focus (A-E)
• If focus=E, ask follow-up: "What should I focus on?"

Pipeline — Full Mode (NotebookLM available)

Step 1: Check existing notebooks

Run: notebooklm list --json

Search results for notebooks with titles matching the topic. If found, ask user: "Found existing notebook '{title}' with {n} sources. Use it? (y/n)"

• If yes: notebooklm use {id}, skip to Step 3
• If no: continue to Step 2

Step 2: Create research notebook

Run: notebooklm create "Research: {topic}" --json Parse notebook ID from output. Run: notebooklm use {id}

Step 3: Query existing sources (if focus=A and existing notebook)

If existing notebook was selected, run: notebooklm ask "What information do you have about {topic}?" --json Store response as baseline context.

Step 4: Launch deep research

Determine mode from depth:

• Depth 1-2: --mode fast
• Depth 3-5: --mode deep

Run: notebooklm source add-research "{topic_query}" --mode {fast|deep} --no-wait

Show progress: [2/5] Starting deep research... (mode: {mode})

Step 5: Wait for research completion

Run: notebooklm research wait --import-all --timeout 300

Show progress: [3/5] Waiting for sources... (processing...)

When complete, run: notebooklm source list --json Report: [3/5] {n} sources imported and ready.

Step 6: Extra WebSearch rounds (depth > 3)

If depth >= 4, run 1 WebSearch round:

• Generate 3-5 targeted search queries based on initial findings
• For each query: WebSearch, then read top 3 results with WebFetch
• Add discovered URLs as sources: notebooklm source add "{url}"

If depth = 5, run a second WebSearch round:

• Generate 3-5 NEW queries based on combined findings
• Same process: search, fetch, add as sources
• Wait for new sources: notebooklm source wait {id}

Show progress: [4/5] Running additional searches... ({n} extra sources)

Step 7: Domain boost (if --domain specified)

Run domain-specific searches from the Domain Boost section below. Add discovered URLs as sources to the notebook.

Step 8: Synthesize

Run structured synthesis query:

notebooklm ask "Based on ALL sources in this notebook, provide a comprehensive research report on '{topic}'.

Structure your response as:

1. EXECUTIVE SUMMARY (3-5 sentences, key takeaways)

2. MAIN FINDINGS (organized by theme, each finding citing specific sources)

3. KEY FINDINGS (bullet list of most important discoveries)

4. CONTRADICTIONS & OPEN QUESTIONS (where sources disagree or gaps exist)

5. For each claim, note which source(s) support it.

Be thorough, analytical, and cite sources precisely." --json

Parse response and extract references. Format into the Report Format below.

Pipeline — Fallback Mode (WebSearch only)

Used when NotebookLM is not available. Iterative WebSearch + WebFetch.

Round 1: Broad search

Generate 5 search queries from the topic (different angles). For each query:

1. WebSearch to collect top 5 URLs
2. WebFetch top 3 URLs to extract key findings
3. Store findings + source URLs

Show progress: [1/{total_rounds}] Broad search... ({n} sources found)

Round 2+ (if depth > 1): Deep dive

For each round (up to depth-1 additional rounds):

1. Analyze findings from previous round
2. Identify gaps and unanswered questions
3. Generate 3-5 new targeted queries
4. WebSearch + WebFetch for each
5. Merge new findings with existing

Show progress: [{round}/{total_rounds}] Deep dive round {n}... ({total} sources)

Domain boost in fallback

Same domain-specific queries as full mode, but results stay in-memory instead of being added to a NotebookLM notebook.

Synthesis in fallback

Compile all findings into the Report Format below. Synthesis is done by the agent directly from collected findings.

Limitations in fallback mode

Post-research actions NOT available without NotebookLM:

• Podcast generation
• Video generation
• Slide deck generation
• Quiz/flashcard generation
• Follow-up questions against persistent notebook

Available actions in fallback:

• Save as .md
• Go deeper (additional WebSearch rounds)
• Compare mode

Domain Boost

When --domain is specified, add these targeted searches to the pipeline.

Maritime (--domain maritime)

Additional search queries:

• "{topic}" site:imo.org
• "{topic}" SOLAS regulation
• "{topic}" STCW code
• "{topic}" ISM Code
• "{topic}" site:iacs.org.uk
• "{topic}" Maritime Safety Committee MSC circular
• "{topic}" class society guidelines

Source priority (highest to lowest):

1. IMO resolutions, circulars, guidelines
2. SOLAS/STCW/ISM Code references
3. IACS unified requirements
4. Class society publications (Lloyd's, DNV, BV, ABS)
5. Maritime industry papers (BIMCO, ICS, INTERCARGO)
6. General maritime publications

Citation format: include regulation references when available. Example: "SOLAS Ch. III, Reg. 19.3.3" or "STCW Code Table A-VI/1"

Legal (--domain legal)

Additional search queries:

• "{topic}" legislation regulation
• "{topic}" official journal law
• "{topic}" court decision ruling
• "{topic}" regulatory framework directive
• "{topic}" compliance requirements

Source priority:

1. Official government legislation/gazettes
2. Court decisions and rulings
3. Regulatory body publications
4. Legal databases and commentary
5. Law firm analysis and white papers

Academic (--domain academic)

Additional search queries:

• "{topic}" research paper study
• "{topic}" peer-reviewed journal
• "{topic}" systematic review meta-analysis
• "{topic}" site:arxiv.org
• "{topic}" site:researchgate.net
• "{topic}" IEEE conference proceedings

Source priority:

1. Peer-reviewed journal articles
2. Systematic reviews and meta-analyses
3. Conference proceedings
4. Preprints (arXiv, SSRN)
5. University research reports
6. Technical reports and white papers

Report Format

Format the final output as a Gemini Deep Research-style report:

# {Topic}