Site AI Assistant Integration

A modern, production-ready AI assistant platform for integrating advanced language models with your website's knowledge base.
Easily build, manage, and query a multilingual knowledge base with robust evaluation and a sleek, responsive frontend.

---

Features

• Knowledge Base Initialization

  • Extracts and chunks content from your website using a base URL or sitemap file upload.
  • Semantic chunking and FAISS vector search for efficient, accurate retrieval.

• Multilingual Language Detection

  • Detects user query language (Arabic, English, French, etc.) using FastText.
  • Automatically adapts prompts and responses for Arabic/Persian or English.

• LLM Integration

  • Uses Groq's Llama-3.3-70b-versatile model for chat responses.
  • Context-aware answers based on retrieved knowledge base content.

• Validation & Evaluation Metrics

  • Computes retrieval metrics (Precision, Recall, F1), LLM metrics (ROUGE, BERTScore), and cross-encoder relevance.
  • Metrics are displayed in the frontend and logged to Comet ML for experiment tracking.

• Modern Frontend

  • Responsive Bootstrap UI with sidebar for knowledge base setup, chat, and collapsible metrics section.
  • Real-time chat and feedback.

---

Project Structure

• backend/: FastAPI app, API routes, vector store, evaluation, and utilities.
• models/: Pre-trained FastText language detection model.
• static/: CSS, JS, and images for the frontend.
• templates/: Jinja2 HTML templates.

---

Quickstart

1. Install uv:

   pip install uv

2. Create a Virtual Environment:

   uv venv

3. Sync Dependencies:

   uv sync --all-extras

4. Configure Environment Variables:
   Copy .env.sample to .env and fill in your API keys and base URL.

5. Run the Application:

   uvicorn backend.main:app --reload

6. Open in Browser:
   Visit http://127.0.0.1:8000

---

Usage

Knowledge Base Initialization

• Use the sidebar form to enter a base URL or upload a sitemap file.
• The backend extracts, chunks, and indexes your website content for retrieval.

Chat & Language Detection

• Enter your question in any supported language.
• The system detects the language and adapts the prompt for the LLM.
• Arabic and Persian queries get native-language prompts; others default to English.

Validation & Experiment Tracking

• Metrics: Precision, Recall, F1, ROUGE, BERTScore, Cross-Encoder Relevance, Latency.
• Comet ML Integration:
  • Logs all evaluation metrics and parameters for experiment tracking.
  • Configure your Comet ML credentials in .env.

---

Environment Variables

• GROQ_API_KEY: Groq LLM API key.
• BASE_URL: Default website for sitemap extraction.
• COMET_ML_API_KEY, COMET_ML_PROJECT_NAME, COMET_ML_WORKSPACE: Comet ML tracking.
• (Optional) HUGGINGFACE_API_KEY, TAVILY_API_KEY, FIRECRAWL_API_KEY for extra integrations.

---

License

MIT License.
See LICENSE for details.

---

Contact

For questions, suggestions, or contributions, please open an issue or pull request.

---