Hugging Face Ecosystem RAG Assistant

An advanced Retrieval-Augmented Generation (RAG) assistant specialized in providing expert-level technical support for the Hugging Face ecosystem: transformers, peft, accelerate, trl, and datasets.

This project addresses the "hallucination" problem in LLMs regarding fast-moving library APIs by grounding responses in official documentation retrieved directly from GitHub.

🔗 Live Demo

Try it now on Hugging Face Spaces: HF Ecosystem RAG Assistant

(No local setup required)

---

🚀 Key Features

• Semantic Library Routing: Automatically classifies queries to specific libraries to narrow the search space and prevent cross-library noise.
• Multi-Query Fusion: Generates 3+ search variations for every user prompt to ensure high-recall retrieval even with vague technical terminology.
• Transparent Reasoning: Features a streaming interface that reveals the agent's internal "thought process" (Routing selection -> Query expansion -> Context extraction).
• Automated Ingestion Pipeline: Built-in ETL that clones repositories, parses Markdown headers, and builds/updates the ChromaDB vector index automatically.

---

🛠 Tech Stack

• Orchestration: LangChain (LCEL)
• Vector Store: ChromaDB
• Embeddings: sentence-transformers/all-MiniLM-L6-v2 (Local/CPU-optimized)
• LLM: Step-3.5-Flash (via OpenRouter)
• UI Framework: Gradio 5.0+ (Streaming support)
• Data Handling: GitPython (ETL), Git LFS (Database storage)

---

📁 Project Structure

huggingface-ecosystem-RAG/
├── data/                 
│   ├── hf_vector_db/     # Persistent ChromaDB index (Stored via Git LFS)
│   └── temp_docs/        # Temporary storage for cloned HF repos
├── src/
│   ├── app.py            # Gradio Web UI & Streaming entry point
│   ├── rag_core.py       # Core RAG logic: Routing, Fusion, and Chains
│   ├── ingest.py         # ETL pipeline for documentation indexing
│   └── config.py         # Global configuration & Path management
├── .env                  # API Keys (OpenRouter, LangSmith)
└── requirements.txt      # Project dependencies

---

⚙️ Setup & Installation

1. Prerequisites

Ensure you have Git LFS installed to pull the pre-built vector database.

git clone https://github.com/your-username/huggingface-ecosystem-RAG.git
cd huggingface-ecosystem-RAG
git lfs pull
pip install -r requirements.txt

2. Configuration

Create a .env file in the root directory:

OPENROUTER_API_KEY=your_openrouter_key
# Optional: Enable LangSmith for tracing
LANGCHAIN_TRACING_V2=true
LANGCHAIN_API_KEY=your_ls_key

3. Execution

python src/app.py

Note: If the data/ folder is empty, the system will automatically trigger the ingestion pipeline to build the index from scratch.

---

🧠 Implementation Details

Semantic Routing & Fallback

The agent employs a custom router to identify the target library meta-data. If the retrieval returns zero results within a specific library, a Global Fallback mechanism triggers a wider search across the entire ecosystem to find relevant context.