Local RAG/MCP Knowledge Base Assistant

📋 The Problem

• Growing Documentation: Knowledge scattered across files
• Information Retrieval: Hard to find answers without keywords
• Privacy Concerns: Cloud solutions may not comply with policies

Users → Search → Answer = 😫

✨ The Solution

A local, intelligent Q&A system using:

• RAG: Semantic search over documentation
• MCP: Dynamic document access
• Local LLM: Privacy-preserving answers (Ollama)

✨ Key Benefits

• ✅ Privacy-first (runs locally)
• ✅ No API costs
• ✅ Fast semantic search
• ✅ Intelligent document access
• ✅ Complete data control

🏗️ Architecture - Top Level

┌──────────────────────┐
│   User Interface     │ (CLI)
└──────────┬───────────┘
           │
     ┌─────┴─────┐
     ▼           ▼
  [RAG]       [MCP]
   Query      Tools
     │           │
     └─────┬─────┘
           ▼
    [Ollama LLM]

🏗️ Architecture - Storage

┌────────────────┐
│  FAISS Index   │ Vector Database
│  + MCP Tools   │
└────────┬───────┘
         │
    ┌────▼─────┐
    │   docs/  │
    │directory │
    └──────────┘

🔍 RAG Pipeline

1. Document Loading → Read .md, .txt, .pdf, .docx
2. Chunking → Split into 700-char chunks
3. Embedding → Use SentenceTransformers
4. Indexing → Build FAISS vector index
5. Query → Retrieve top 5 similar chunks
6. Prompt Building → Create context-aware prompt
7. LLM Generation → Get answer from model

🔍 Why FAISS?

• Fast vector similarity search
• Lightweight and memory-efficient
• No external dependencies
• Perfect for local deployments
• Millions of vectors supported

🔧 MCP - Model Context Protocol

MCP provides standardized interface for LLM tool access:

read_document(file_path)
list_documents()
search_documents(query)

🔧 MCP Benefits

• Tool Use by LLM
• Real-time document access
• Standardized interface
• Easy to extend
• Local tool execution

💻 Tech Stack

Language:      Python 3.10+
Vector DB:     FAISS
Embeddings:    SentenceTransformers
LLM:           Ollama (local)
MCP:           FastMCP

📁 Project Structure

src/
├── config.py           Configuration
├── main.py             CLI entry point
├── assistant.py        Main orchestrator
├── rag/
│   ├── ingest.py      Load documents
│   ├── chunk.py       Split text
│   ├── embed.py       Generate embeddings
│   ├── build_index.py Build FAISS index
│   └── query.py       Retrieve & generate
├── mcp/
│   ├── server.py      MCP tool definitions
│   └── client.py      MCP client wrapper
└── docs/              Documentation

🚀 Index Building (Setup)

$ python main.py build-index

1. Load documents
  ↓
2. Split into chunks
  ↓
3. Generate embeddings
  ↓
4. Build FAISS index
  ↓
5. Save files

🚀 Query Processing (Runtime)

User Question
  ↓
Embed question
  ↓
Search FAISS → Top 5 chunks
  ↓
LLM decides: Use MCP tools?
  ↓
Build prompt + context
  ↓
Call Ollama
  ↓
Return answer + sources

✨ Core Features

• Semantic Search: Find by meaning, not keywords
• Multi-format: .md, .txt, .pdf, .docx files
• Source Attribution: Shows document sources
• MCP Tools: LLM can read full documents
• No External APIs: Runs locally only
• Fast Retrieval: Sub-second search

⚙️ Configuration Options

CHUNK_SIZE = 700
CHUNK_OVERLAP = 100
EMBEDDING_MODEL = "all-MiniLM-L6-v2"
OLLAMA_MODEL = "qwen3:0.6b"
TOP_K = 5

🎬 Live Demo - Starting

$ python main.py

Output:

🤖 Company Knowledge Base
Ask questions about documentation
Type 'exit' to stop

🎬 Demo - Query 1

❓ What are company values?

🤖 Innovation, integrity, collaboration

📚 Sources:
  • Loan Rangers Team.md
  • Info Security.md

🎬 Demo - Query 2

❓ What documents do we have?

🤖 [Uses MCP list_documents]
  • Loan Rangers Team.md
  • Information Security.md
  • Services.md

🎬 Demo - Query 3

❓ Full security policy?

🤖 [Uses MCP read_document]
[Full document content...]

🔐 Security - Local vs Cloud

Cloud: Data → Internet → Server

• ⚠️ Network transmission
• ⚠️ External storage
• ⚠️ Subscription costs

Local: Data → Local System

• ✅ No transmission
• ✅ Local storage only
• ✅ No costs

🔐 Implementation Safeguards

• MCP Sandbox: Prevents path traversal
• Local Storage: Documents stay on device
• No Telemetry: No tracking
• Offline Ready: Works without internet

⚡ Performance Benchmarks

Index Building:   ~30s (one-time)
Query Embedding:  ~50ms
FAISS Search:     ~5ms
LLM Generation:   2-5s
Total Cycle:      2-6s

⚡ Tuning for Speed

# Faster (smaller model):
OLLAMA_MODEL = "qwen3:0.6b"

# Faster retrieval:
TOP_K = 3
CHUNK_SIZE = 500

🚢 Deployment - Single Machine

1. Install Ollama & Python deps
2. Copy docs/ to server
3. Build index
4. Run with nohup

$ nohup python main.py > log &

🚢 Scaling - Option 1: FastAPI

[HTTP Clients]			[HTTP Clients + Webllm]
       ↓        						 ↓
   [FastAPI]     				 [FastAPI]
       ↓         					 ↓
[Ollama + FAISS]      			  [FAISS]

🚢 Scaling - Option 2: Distributed

[Clients] → [Load Balancer]
             ↓
      [Multiple Retrievers]

🚢 Storage Scaling

Docs     Index      Build
10 MB    ~2 MB      ~5s
100 MB   ~20 MB     ~30s
1 GB     ~200 MB    ~5min

🔮 Phase 2: Enhanced Features

• ☐ Web UI (Streamlit)
• ☐ API endpoints
• ☐ Multi-language support
• ☐ Document versioning
• ☐ Fine-tuned embeddings

🔮 Phase 3: Advanced

• ☐ Conversation memory
• ☐ Multi-hop reasoning
• ☐ Metadata filtering
• ☐ Feedback loop
• ☐ Analytics dashboard

🔮 Phase 4: Enterprise

• ☐ User authentication
• ☐ Audit logging
• ☐ Role-based access
• ☐ LLM fine-tuning
• ☐ Cost analysis

📊 Why This Works

Aspect	Traditional	Our RAG
Understanding	Keywords	Semantic
Answers	Documents	Direct
Privacy	Cloud	Local
Cost	Subscription	One-time
Speed	Slow	Sub-second

✅ What You Have Now

• Local privacy-first knowledge base
• Fast semantic search (FAISS)
• Intelligent tool use (MCP)
• Maintainable Python code
• Foundation for enterprise features

🙋 Quick Reference

# Build index
python main.py build-index

# Run interactively
python main.py

# Check config
cat config.py

📚 Resources

• Code: MobilaName/local-rag-mcp
• FAISS: facebook/faiss
• Ollama: ollama.ai
• FastMCP: github.com/jlowin/fastmcp
• Transformers: huggingface.co

Thank You!