NotebookLM MCP Structured

# Usage Guide - Structured Prompts & Source Fidelity > **NotebookLM MCP Structured v1.0.0** - Client-side prompt structuring for professional document analysis. This guide covers usage patterns optimized for **source fidelity** - ensuring responses come ONLY from your uploaded documents with proper citations. > For installation and quick start, see the main [README](../README.md). --- ## Core Concept: Structured Prompts Unlike the original MCP server, this fork **automatically transforms** your questions into structured prompts that enforce: 1. **Source constraints**: NotebookLM uses ONLY your uploaded documents 2. **Citations**: Every claim includes document source attribution 3. **Explicit gaps**: Missing information declared as `[NOT FOUND IN DOCUMENTS]` 4. **Consistent format**: Output structured by question type ### How It Works ``` Your question: "What are the main findings?" Claude structures it as: ┌─────────────────────────────────────────────┐ │ RESPONSE INSTRUCTIONS │ │ │ │ TASK: What are the main findings? │ │ │ │ OPERATIONAL CONSTRAINTS │ │ - Use ONLY information from documents │ │ - DO NOT add external knowledge │ │ - If not found: "[NOT FOUND IN DOCUMENTS]" │ │ │ │ REQUIRED OUTPUT FORMAT │ │ For each finding: │ │ - FINDING: [description] │ │ - SOURCE: [document name/section] │ │ - QUOTE: "direct quote" │ │ │ │ BEGIN STRUCTURED RESPONSE │ └─────────────────────────────────────────────┘ NotebookLM responds with source-faithful answer ``` **You just ask naturally** - Claude handles the structuring automatically. --- ## Use Cases ### Legal Document Analysis Perfect for ensuring responses come only from case documents: ``` "Compare the rulings in Case A vs Case B" → Structured as comparison with source citations "Extract all clauses related to liability" → Structured as extraction with document references "What does the contract say about termination?" → Source-faithful answer or [NOT FOUND IN DOCUMENTS] ``` ### Research & Literature Review ``` "What methodology did the authors use?" → Direct quotes from methodology section "Summarize the key findings across all papers" → Findings attributed to specific documents "What are the limitations mentioned?" → Explicit list with source citations ``` ### Professional Fact-Checking ``` "Does the document support the claim that X?" → YES with quote, NO, or [NOT FOUND IN DOCUMENTS] "What evidence is provided for Y?" → Listed evidence with source attribution "Are there any contradictions between documents?" → Comparison analysis with specific citations ``` --- ## Question Types Claude automatically detects your question type and applies the appropriate structure: ### Comparison Questions **Triggers**: "compare", "vs", "versus", "difference", "similarities" ``` "Compare approach A vs approach B in the documents" Output structure: - Elements being compared - Similarities (with sources) - Differences (with sources) - Synthesis ``` ### List Questions **Triggers**: "list", "identify", "which", "what are the" ``` "List all the requirements mentioned" Output structure: 1. Item - SOURCE: [document] - "quote" 2. Item - SOURCE: [document] - "quote" ... ``` ### Analysis Questions **Triggers**: "analyze", "examine", "evaluate", "assess" ``` "Analyze the risk factors discussed" Output structure: - Subject of analysis - Observations (with evidence) - Evidence (direct quotes) - Conclusions (sourced) ``` ### Explanation Questions **Triggers**: "explain", "why", "how does" ``` "Explain how the process works according to the manual" Output structure: - Concept identified - Explanation (sourced) - Examples from documents - Related information ``` ### Extraction Questions **Default for all other questions** ``` "What is the deadline mentioned?" Output structure: - Data point - SOURCE: [document/section] - QUOTE: "exact text" ``` --- ## Language Support The structured fork works with **any language Claude supports**. **Tested**: Italian - works perfectly with Italian users and documents. **How it works**: 1. Claude detects your language from context 2. Structuring instructions adapt to your language 3. NotebookLM responds in the same language 4. No configuration needed **Other languages**: Should work automatically. If you use a language other than Italian, please share feedback on GitHub! --- ## Best Practices ### 1. Ask Specific Questions More specific questions → better structured responses ``` ❌ "Tell me about the project" ✅ "What are the project milestones and their deadlines?" ``` ### 2. Trust the `[NOT FOUND]` Response When you see `[NOT FOUND IN DOCUMENTS]`, it means: - The information genuinely isn't in your documents - Source fidelity is working correctly - NotebookLM isn't making things up ### 3. Use Multiple Questions for Complex Topics ``` 1. "What does document A say about X?" 2. "What does document B say about X?" 3. "Are there contradictions between the two?" ``` ### 4. Session Management - Same `session_id` maintains context across questions - Sessions auto-cleanup after 15 minutes - Use `reset_session` to start fresh --- ## Authentication ### Automatic Connection Verification This fork automatically checks authentication before operations: 1. You ask a question 2. Server verifies NotebookLM connection 3. If expired/missing: - Chrome running → Message to close Chrome - Chrome closed → Auto-opens login browser 4. After login, original request proceeds **No manual intervention needed** in most cases. ### Manual Authentication ``` "Check NotebookLM health" → get_health "Setup NotebookLM auth" → setup_auth (opens browser) "Switch NotebookLM account" → re_auth (clears and re-authenticates) ``` ### Rate Limits - Free accounts: 50 queries/day - Google AI Pro/Ultra: 5x higher limits - Use `re_auth` to switch accounts when limit reached --- ## Notebook Management ### Add Notebooks ``` "Add this NotebookLM notebook: [URL]" ``` ### Switch Notebooks ``` "Use the Legal Docs notebook" "Switch to Research Papers notebook" ``` ### Organize with Metadata ``` "Add notebook with description 'Contract analysis Q4 2024' and tags: legal, contracts, Q4" ``` --- ## Example Workflow ### Legal Research Session ``` 1. "Add the Case Law notebook: [URL]" 2. "What does Smith v. Jones say about negligence?" → Structured response with citations 3. "Compare that with the ruling in Brown v. Davis" → Comparison with sources from both cases 4. "Are there any contradictions?" → Analysis with specific page references 5. "Extract all mentions of 'duty of care'" → List with direct quotes and sources ``` ### Research Literature Review ``` 1. "Select the Literature Review notebook" 2. "What methodology did each study use?" → Structured list per study 3. "What are the main findings across all papers?" → Findings attributed to specific papers 4. "What limitations are mentioned?" → Source-faithful list 5. "Are there gaps in the research?" → Analysis based only on what papers state ``` --- ## Troubleshooting ### NotebookLM Timeout - Never use decorative lines (`===` or `---`) in custom prompts - This fork handles this automatically ### Responses Include External Knowledge - Verify you're using `ask_question` tool - Check that structuring hasn't been accidentally disabled ### Missing Citations - Ask more specific questions - General questions may produce less structured responses See [Troubleshooting Guide](troubleshooting.md) for more. --- ## Credits - Original `notebooklm-mcp`: [Gérôme Dexheimer](https://github.com/PleasePrompto/notebooklm-mcp) - Client-side structuring approach: Paolo Dalprato