knowledge-mcp

A Model Context Protocol (MCP) server that provides scoped RAG workspaces ("Notebooks") backed by Postgres + pgvector and TEI.

Overview

This server enables an agent to:

1. Create named "Notebooks" (Postgres-backed collections).
2. Ingest documents (PDF, Markdown, Text) into specific notebooks.
3. Query specific notebooks using vector search (RAG).
4. Synthesize findings across a notebook.

Designed to replicate the NotebookLM experience: clean, focused, bounded context.

Stack

• Language: Python 3.11+
• Framework: mcp SDK
• Vector DB: Postgres + pgvector
• Embeddings: Text Embeddings Inference (TEI) - BAAI/bge-base-en-v1.5

Tools

create_notebook

Creates a new isolated workspace (Postgres-backed notebook).

• name: string (e.g., "project-alpha")

add_source

Ingests a document into the notebook.

• notebook: string
• content: string (raw text or local path)
• source_name: string
• format: text or pdf_path

query_notebook

Performs a semantic search/RAG generation against the notebook.

• notebook: string
• query: string

Configuration

Env vars:

• DATABASE_URL: Postgres connection string (e.g., postgresql://postgres:password@postgres.knowledge-mcp.svc:5432/knowledge)
• TEI_URL: URL to TEI (e.g., http://text-embeddings.tei.svc.cluster.local:8080)

TODO

• Add PDF → Markdown/text conversion step to improve extraction quality.
• Add OCR pipeline for scanned PDFs.