MCP-Based Enterprise Knowledge Search Assistant

This repository contains a proof-of-concept implementation of an enterprise knowledge assistant that compares a traditional RAG-style experience with an MCP-driven retrieval flow.

Architecture

• Frontend: Claude Desktop style chat experience

• Backend: FastAPI + Python 3.12

• Retrieval: Elasticsearch-style hybrid search abstraction with BM25 + dense vector similarity

• LLM: Ollama with gemma3:latest

• MCP: Custom Python MCP-style tool server exposing domain-specific retrieval tools

Key features

• Question routing to the correct MCP tool

• Hybrid retrieval with ranked chunks and source citations

• Ollama integration for answer generation

• Evaluation metrics logging to JSON

• Docker Compose support for local development

Running locally

1. Install dependencies:

   • pip install -e .

2. Start Ollama and ensure gemma3:latest is available.

3. Start the API:

   • uvicorn app.core.app:app --reload

4. Call the chat endpoint:

   • POST /chat with { "question": "What is the maternity leave policy?" }

Sample questions

• What is the maternity leave policy?

• How do I submit a travel reimbursement request?

• What are the annual appraisal guidelines?

• What should I do for a security incident?

Evaluation

Run the evaluation helper to write a metrics record:

from app.services.evaluation import EvaluationService

service = EvaluationService()
print(service.compare("What is the maternity leave policy?"))