Personal Knowledge-Base MCP Server

A Personal Knowledge-Base MCP Server that provides semantic search over a student-owned document collection using the Model Context Protocol (MCP), Gemini embeddings, and Qdrant.

Project Overview

This project exposes a personal knowledge base as callable MCP tools.

Instead of relying on keyword matching, the system converts user queries into vector embeddings and retrieves semantically relevant document chunks from Qdrant.

Architecture

User / MCP Client
       |
       v
MCP Server (FastMCP)
       |
       +----------------------+
       |                      |
       v                      v
 search_notes()        get_document()
       |
       v
Gemini Embedding API
       |
       v
Qdrant Vector Database
       |
       v
Ranked Chunks
       |
       v
Source + Page + Score + Text
## Features

* PDF document ingestion
* Page-by-page text extraction
* Recursive text chunking
* Gemini `gemini-embedding-001` embeddings
* Qdrant vector storage
* Semantic similarity search
* Source and page citations
* Confidence threshold for low-relevance queries
* Full-document retrieval
* Indexed-source listing
* MCP Inspector support

## MCP Tools

### `search_notes`

Searches the knowledge base using semantic similarity.

Arguments:

* `query`: search question or topic
* `top_k`: maximum number of results

Returns:

* similarity score
* source filename
* page number
* relevant text chunk

### `get_document`

Returns the complete text of an indexed PDF document.

Argument:

* `doc_id`: document filename

Example:

```text
Complex_Variables_Project_Report.pdf

list_source_documents

Lists all indexed source documents.

Example output:

1. Complex_Variables_Project_Report.pdf

Project Structure

Personal-Knowledge-MCP/
├── documents/
│   └── Complex_Variables_Project_Report.pdf
├── services/
│   ├── chunking.py
│   ├── embedding.py
│   ├── pdf_reader.py
│   └── qdrant_service.py
├── .env
├── .gitignore
├── evaluation.py
├── ingest.py
├── requirements.txt
└── server.py

Setup

1. Create and activate virtual environment

python -m venv .venv
.venv\Scripts\Activate.ps1

2. Install dependencies

pip install -r requirements.txt

3. Configure Gemini API key

Create a .env file in the project root:

GEMINI_API_KEY=your_api_key_here

Never commit .env to Git.

4. Start Qdrant

The project uses local Qdrant at:

http://localhost:6333

Example Docker command:

docker run -d --name qdrant -p 6333:6333 -p 6334:6334 qdrant/qdrant

Document Ingestion

Place the PDF inside:

documents/

Run:

python ingest.py

The ingestion pipeline performs:

PDF
 ↓
Page extraction
 ↓
Chunking
 ↓
Gemini embeddings
 ↓
Qdrant storage

Each stored chunk contains:

text
page
source

Running the MCP Server

Start the MCP Inspector:

mcp dev server.py

The MCP server uses STDIO transport.

Available tools:

search_notes
get_document
list_source_documents

Retrieval Evaluation

A small evaluation set of five queries was used to check whether at least one expected relevant page appeared within the top three retrieved results.

Evaluation result:

Tests: 5
Successful hits: 5
Hit@3: 100%

Example evaluation queries included:

• What is a complex variable?
• What are the Cauchy-Riemann equations?
• How does the Laplace transform help engineering systems?
• What is the difference between Laplace and Fourier transforms?
• How is FFT used for audio noise reduction?

Confidence Filtering

The search tool uses an initial similarity threshold of:

0.60

For example, relevant queries produced scores around:

0.79
0.76
0.75

while an unrelated query produced scores around:

0.52

Therefore low-scoring results are filtered and the tool returns:

No confident match found.

Technologies

• Python
• FastMCP
• Model Context Protocol (MCP)
• Google Gemini Embeddings
• Qdrant
• PyMuPDF
• LangChain Text Splitters
• Docker
• MCP Inspector

Current Knowledge Source

The current demonstration corpus is:

Complex_Variables_Project_Report.pdf

The document contains 7 pages and was split into 30 chunks for the indexed personal_knowledge collection.

Security

• API keys are stored in .env
• .env is excluded through .gitignore
• Secrets should never be committed to source control

Future Improvements

• Support Markdown and TXT documents
• Add document-level persistent IDs
• Improve duplicate-chunk handling
• Expand the evaluation dataset
• Add more retrieval metrics
• Support multiple document collections
• Add optional Qdrant Cloud deployment