README.md

metadata

title: Codey Bryant 3.0
emoji: 🤖
colorFrom: blue
colorTo: green
sdk: docker
app_file: app_hf.py
pinned: true
license: mit

🤖 Codey Bryant 3.0 - SOTA RAG Coding Assistant

Advanced RAG Architecture

Codey Bryant 3.0 implements state-of-the-art Retrieval-Augmented Generation (RAG) with four key innovations:

1. HyDE (Hypothetical Document Embeddings)

Generates hypothetical answers to improve retrieval relevance for vague queries.

2. Query Rewriting

Transforms casual/vague questions into specific, searchable queries.

3. Multi-Query Retrieval

Searches multiple query variations to increase recall.

4. Answer-Space Retrieval

Retrieves from both question AND answer embeddings for better context.

Technical Stack

• LLM: TinyLlama 1.1B (4-bit quantized)
• Embeddings: all-MiniLM-L6-v2
• Retrieval: FAISS + BM25 hybrid
• Datasets: OPC Educational + Evol-Instruct
• Framework: Gradio + Hugging Face Transformers

Performance Features

• Handles vague queries like "it's not working"
• Streaming responses
• Context-aware generation
• Hybrid dense-sparse retrieval
• Persistent artifact storage

Getting Started

1. Click "Initialize Assistant" (required once)
2. Ask Python coding questions
3. Get intelligent, context-aware responses

Example Queries

• "How to read a CSV file in Python?"
• "Why am I getting 'list index out of range'?"
• "Make this function faster..."
• "Help, my code isn't working!"
• "Best way to sort a dictionary by value?"

Why This Architecture?

1. HyDE: Addresses the "semantic gap" between queries and documents
2. Query Rewriting: Improves retrieval for conversational queries
3. Multi-Query: Increases recall for complex questions
4. Answer-Space: Provides better context for generation

📁 Repository Structure