Prompt engineering, RAG, LangChain, etc

Based on the paper Principled Instructions Are All You Need for Questioning LLaMA-1_2, GPT-3.5_4

Model: Gemini-1.5-flash

Gemini API is utilized.

Finance Area

Data : BC Card AI Finance Data

Base Model: HuggingFace SmolLm2-1.7B-Instruct

Equipment: RTX 4090 24 GB

1. Ensemble Sparse Retirver and Dense Retriever

Dense Retriever (Vector Store): LangChain FAISS

Sparse Retriever (Traditional Search Method): BM25

rank_bm25 is a component of LangChain in sparse retriving process.

2. RAG condidering Department and Rank in a Corporation.

Use LangGraph and Decide to do RAG or Not.

Check groundedness score, context relevancy, response relevancy to measure the reponse and RAG performance.

Use both Gemini API and OpenAI API