KnowledgeManager MCP

What is this project?

KnowledgeManager MCP is a highly efficient local server that implements the Model Context Protocol (MCP). It acts as a bridge, providing Large Language Models (LLMs) with direct, tool-based access to a curated knowledge base built from local PDF documents.

What does it do?

It automatically ingests PDF files from a specified data directory, processes them into semantic text chunks, and creates a fast, in-memory vector database. It then exposes two primary tools to any MCP-compatible LLM client (such as Claude Desktop):

1. Query Knowledge Base: Allows the LLM to ask general questions and retrieve the most relevant information across the entire document corpus.
2. Search Specific Documents: Allows the LLM to target its questions to specific PDF files by name, narrowing down the search scope for higher accuracy.

Additionally, it silently maintains a local SQLite database (logs.db) to audit and track every tool execution, query, and output performed by the AI.

How it works (Technical Level)

When the server starts src.py:

1. Database Initialization: It creates or connects to a SQLite database (logs.db) for secure usage logging.
2. Document Loading: It asynchronously scans the ../data directory for .pdf files and loads them in parallel using LangChain's PyPDFLoader.
3. Chunking: The extracted text is smartly chunked using RecursiveCharacterTextSplitter (1000 character chunks with 200 character overlaps) to preserve context continuity.
4. Remote Embedding Generation: It uses HuggingFaceInferenceAPIEmbeddings to generate rich text embeddings via the BAAI/bge-large-en-v1.5 model on the HuggingFace API.
5. Vector Indexing: It indexes these embeddings into an in-memory FAISS (Facebook AI Similarity Search) database.
6. Tool Exposure: Finally, the standard MCP server boots up via FastMCP over stdio transport. When the LLM requests a search, the server performs asynchronous similarity searches against the FAISS index, intelligently filters results based on query length and confidence scores, and constructs an answer for the LLM.

Technology Used & Benefits

• Model Context Protocol (FastMCP): Provides a standardized, zero-configuration way to expose Python code as LLM tools.
• LangChain: Simplifies the pipeline for loading, splitting, and managing complex text documents.
• FAISS: Provides blazingly fast in-memory similarity search without the overhead of spinning up a dedicated vector database service like Pinecone or Weaviate.
• HuggingFace Inference API: By offloading the heavy embedding generation to a cloud API, the host machine uses virtually zero CPU/GPU resources for inference, allowing it to run smoothly in the background on any hardware.
• SQLite: A self-contained, serverless database that provides out-of-the-box audit logs without complex setup requirements.
• Python Asyncio: Used extensively throughout document loading and vector querying to ensure the server remains highly responsive and non-blocking.

Why is it useful to have/use?

• Zero-Hallucination Answers: LLMs lack access to your private files natively. This system directly connects them to your PDFs, forcing the AI to base its answers on your actual documents.
• Granular Auditing: In enterprise or professional environments, knowing exactly what the AI was searching for and what context it received is critical. The built-in logging provides complete transparency.
• Resource Efficiency: By avoiding heavy local embedding models and avoiding permanent, disk-hugging vector databases, this project is lightweight enough to be run 24/7 on a laptop while still providing enterprise-grade search speed.
• Intelligent Filtering: The dynamic filtering logic actively adjusts confidence thresholds based on whether the AI is asking a simple or complex question, stripping out irrelevant noise and improving AI performance.