RAG for Competitive Programming (rag_4_cpr)

RAG for Competitive Programming (rag_4_cpr) is a Retrieval-Augmented Generation (RAG) system designed to assist with solving competitive programming problems efficiently. The project aims to combine retrieval-based techniques with advanced LLM reasoning to deliver optimal solutions.

Dataset

The dataset is sourced from Project Euler (https://projecteuler.net/archives), specifically using the first 100 problems and solutions. Shout out to the Project Euler community for their exceptional resources.

Tools and Frameworks

• NVIDIA NIM API (https://build.nvidia.com/explore/discover): Used to integrate the meta/llama3-8b-instruct foundation model for generating responses.
• SQLite3: Database cpr.db is used to store problem-solution pairs, facilitating scalability through the pipeline in ./problems/main.py.
• Hugging Face: Embedding generation with the intfloat/e5-large-v2 model.
• Chroma Vector Store: Manages embeddings for retrieval tasks.
• LangChain: Facilitates prompt engineering, LLM integration, and retrieval workflows.

Processing Pipeline

1. Retrieve problem-solution pairs from the problem_solution table in cpr.db.
2. Format the data into a list of dictionaries, each containing an ID and combined problem-solution text.
3. Generate embeddings using the Hugging Face intfloat/e5-large-v2 model.
4. Initialize a Chroma vector store with the embeddings and associate metadata for problem retrieval.
5. Employ a custom prompt template to guide the meta/llama3-8b-instruct model to provide concise, performance-focused solutions.

Output Example

Below is an example output, showcasing the system's reasoning and selected solutions.