Godspeed RAG Chatbot

A Retrieval-Augmented Generation (RAG) chatbot built using the Godspeed Framework, designed to intelligently respond to queries using the latest Godspeed Systems documentation. It leverages embeddings and a local FAISS vector database to deliver fast, relevant, and source-traceable answers.

---

Key Features

1. Query Handling via mcp.handle-query Tool

• Purpose: Core component responsible for receiving user queries and returning relevant context and sources.
• How it Works:
  • Hosted on stdio and automatically connected to the Saarthi assistant.
  • When the "Godspeed Mode" is activated in Saarthi, user queries are forwarded to this tool.
  • The tool searches the FAISS vector store for relevant content and returns:
    • Extracted context
    • Source document paths

2. GitHub Documentation Ingestion

• Source: Godspeed Documentation GitHub Repo
• How it Works:
  • Repo URL and branch are configured.
  • The ingestion script checks for diffs via GitHub API and SHA hash comparison to determine modified files.
  • New or modified files are chunked and embedded using a Google Gemini embedding model, then upserted into FAISS.

3. FAISS Vector Store (Local)

• Why FAISS:
  • FAISS is fast and efficient, particularly with local storage.
  • Minimizes latency compared to remote vector databases.
• Operations:
  • Upsert, search, delete vectors
  • Stores document path as doc_id for efficient tracking of file changes

4. User Query HTTP Endpoint

• Accepts user queries via HTTP and returns context and results in real-time.
• Useful for testing or integrating the chatbot into frontend applications.

5. Document Upload HTTP Endpoint

• Accepts raw documentation files for ingestion.
• Uses SHA comparison to avoid redundant vector operations.

6. LLM Configuration HTTP Endpoint

• Allows dynamic configuration of embedding/LLM keys for better modularity.
• Currently supports Google Gemini Embeddings.

7. Cron Job for Repo Sync

• Frequency: Checks every minute.
• Sync Trigger: Detects if 24 hours have passed since the last update.
• Action: If a day has passed, it re-fetches GitHub data and re-ingests changed files.

8. API Key Management via mcp.handle-api-key

• Integration: API key input can be done via Saarthi through the MCP layer.
• Ensures secure and flexible key injection for embedding models.

---

🧱 Architecture

                          ┌────────────┐
                          │   Saarthi  │
                          └─────┬──────┘
                                │
                                ▼
                      ┌────────────────────┐
                      │ mcp.handle-query   │◄────┐
                      └────────┬───────────┘     │
                               │                 │
          ┌────────────────────▼────────────┐    │
          │  FAISS Vector Store (Local)     │    │
          └─────────────────────────────────┘    │
                               ▲                 │
        ┌──────────────────────┴────────────┐    │
        │ ingest_and_update_docs.ts         │◄───┘
        └────────────────┬──────────────────┘
                         │
       ┌─────────────────▼───────────────────────┐
       │ GitHub Repo: gs-documentation           │
       └─────────────────────────────────────────┘

---

🚀 Running the Project

Clean Start (Windows)

start_clean.bat

Clean Start (Linux/macOS)

./start_clean.sh

Both scripts:

• Kill any process on port 10947
• Run godspeed serve to start the backend

---

Limitations

Embedding Model Limitation

• Currently supports only Google Gemini Embeddings
• Other models like OpenAI, Cohere, etc., are not yet integrated

Local FAISS Index

• Pros: Fast and efficient
• Cons: Not scalable for distributed or cloud-native architectures

Single Documentation Source

• Currently tied to a specific GitHub repo (gs-documentation)
• No built-in support for multiple sources or domains

No GUI/Web Interface

• Operates via CLI, MCP, or HTTP endpoints
• No default front-end for interacting with the bot

---

🧪 Test Scenarios

• Query handling through Saarthi and mcp.handle-query
• Documentation updates via GitHub sync (cron job)
• File-level change detection using SHA and GitHub diffs
• Document ingestion through HTTP endpoint

---

🧩 Future Improvements

• 🧠 Embedding model selection and modular configuration
• 📊 Improve logging and usage analytics