Personal Knowledge Platform

A personal research and knowledge management system powered by Claude.ai + MCP + LlamaIndex.

Features

• Perplexity-style search - Search your docs AND the web with citations
• NotebookLM-style summaries - Multi-doc synthesis and podcast scripts
• SEO/AEO content generation - Optimized outlines and snippets
• $0/month LLM costs - Uses your existing Claude.ai subscription
• 100% local & private - All data stays on your Mac

Architecture

Claude.ai (You) ← MCP Protocol → Local Backend (LlamaIndex + FastAPI)
                                        ↓
                                  PostgreSQL + pgvector

Key Insight: Local backend is "dumb" (search/index only). Claude.ai is the intelligent layer.

Quick Start

1. Start PostgreSQL

docker-compose up -d

2. Install Dependencies

cd backend
pip install -r requirements.txt
python -m spacy download en_core_web_sm

3. Upload Documents

# Copy your PDFs, notes, etc. to:
cp ~/Documents/*.pdf data/documents/

4. Configure Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "personal-knowledge": {
      "command": "python",
      "args": [
        "/Users/YOUR_USERNAME/personal-knowledge-platform/backend/mcp_server.py"
      ],
      "env": {
        "SEARXNG_URL": "https://searx.be"
      }
    }
  }
}

Replace YOUR_USERNAME with your actual username!

5. Restart Claude Desktop

Quit and restart Claude Desktop app.

6. Test

Ask me in Claude.ai:

• "What tools do you have?" (should list 9 tools)
• "Upload this document: /path/to/file.pdf"
• "Search my knowledge base for X"
• "Search the web for Y and cite sources"

Available Tools (9)

Tool	Purpose
search_knowledge_base	RAG search your documents
web_search	Search web (SearxNG/Tavily)
upload_document	Add new documents
generate_seo_outline	SEO content structure
generate_aeo_snippets	Answer Engine Optimization
extract_keywords	SEO keyword extraction
analyze_content_seo	Content quality analysis
reindex_documents	Rebuild search index
list_indexed_documents	Show indexed files

Usage Examples

Example 1: Research with Citations

You: "What are the benefits of intermittent fasting? Search my docs and web."

I do automatically:

1. Call search_knowledge_base(query="intermittent fasting")
2. Call web_search(query="intermittent fasting 2026")
3. Synthesize with citations

I respond: Cited answer with sources from your docs + web.

Example 2: SEO Article

You: "Write SEO article about 'Personal Knowledge Management' using my notes."

I do automatically:

1. generate_seo_outline(topic="Personal Knowledge Management")
2. search_knowledge_base(query="knowledge management")
3. web_search(query="PKM best practices 2026")
4. Write 2000-word optimized article

Example 3: Multi-Doc Summary

You: "Summarize my 5 research papers as a podcast script."

I do automatically:

1. search_knowledge_base() for each paper
2. Synthesize into conversational dialogue

Configuration

Environment Variables

Edit .env file:

# Database
DATABASE_URL=postgresql://kb_user:kb_pass@localhost:5432/knowledge_base

# Web Search (choose one)
SEARXNG_URL=https://searx.be
# TAVILY_API_KEY=your_key_here  # Optional premium search ($29/mo)

# Storage
DOCUMENTS_DIR=../data/documents
STORAGE_DIR=../data/storage

# Embedding Model (local, no API costs)
EMBEDDING_MODEL=BAAI/bge-small-en-v1.5
EMBEDDING_DEVICE=cpu

Supported Document Types

• PDF (.pdf)
• Word (.docx)
• Text (.txt)
• Markdown (.md)

Cost Analysis

Component	Monthly Cost
LLM (Claude.ai)	$0 (existing subscription)
Embeddings (local)	$0
Web Search (SearxNG)	$0
Web Search (Tavily - optional)	$29
Infrastructure (Mac)	$0
Total	$0-29/mo

vs. Commercial Alternatives:

• Perplexity Pro: $20/mo
• NotebookLM Plus: $20/mo (future)
• API costs: $50-200/mo
• This solution: $0/mo

Troubleshooting

"No documents indexed yet"

# Manually reindex
cd backend
python -c "
from mcp_server import initialize_components, indexer
initialize_components()
result = indexer.index_documents('../data/documents')
print(result)
"

Or ask me in Claude.ai: "Reindex all documents"

"Module not found"

cd backend
pip install -r requirements.txt
python -m spacy download en_core_web_sm

PostgreSQL not running

docker-compose up -d
docker ps  # Verify container is running

MCP server not visible in Claude Desktop

1. Check config file path: ~/Library/Application Support/Claude/claude_desktop_config.json
2. Verify Python path in config is correct
3. Restart Claude Desktop (fully quit and reopen)
4. Check MCP server logs in Claude Desktop developer tools

Development

Run MCP Server Manually (for debugging)

cd backend
python mcp_server.py

Test Components Directly

# Test indexer
from rag.indexer import DocumentIndexer
indexer = DocumentIndexer(
    "postgresql://kb_user:kb_pass@localhost:5432/knowledge_base",
    "../data/storage"
)
result = indexer.index_documents("../data/documents")
print(result)

# Test retriever
from rag.retriever import KnowledgeRetriever
retriever = KnowledgeRetriever(indexer.get_index())
results = retriever.search("your query")
print(results)

# Test web search
import asyncio
from web.search import WebSearcher
searcher = WebSearcher()
results = asyncio.run(searcher.search("test query"))
print(results)

Roadmap

• Add support for more document types (EPUB, HTML)
• Implement conversation history tracking
• Add automated web scraping
• Build citation graph visualization
• Add notebook export (Obsidian, Notion)
• Multi-language support
• Audio/video transcription integration

Architecture Details

Why No LLM in Local Backend?

Critical Design Decision: The local backend uses LlamaIndex WITHOUT any LLM configured.

from llama_index.core import Settings
Settings.embed_model = HuggingFaceEmbedding(...)
# Settings.llm is NEVER set - Claude.ai is the LLM

Why:

• Avoids API costs (embeddings are local)
• Claude.ai (via MCP) does all reasoning/generation
• Backend only does retrieval/search/indexing
• Keeps architecture simple and cost-effective

Data Flow

1. User asks question in Claude.ai
2. Claude.ai calls MCP tool (e.g., search_knowledge_base())
3. MCP Server receives request via stdio
4. Backend executes search (vector similarity, no LLM)
5. Results returned to Claude.ai as JSON
6. Claude.ai synthesizes answer with citations
7. User gets intelligent response

License

MIT

Support

For issues or questions, open a GitHub issue or check the documentation.