Adaptive RAG for Banking Compliance

Inspiration

The project aims to improve banking compliance question answering by retrieving accurate evidence from regulatory documents and reducing irrelevant information.

What it does

Adaptive RAG retrieves relevant policy chunks dynamically and uses an LLM to generate answers grounded in the retrieved context with source references.

How we built it

We used Python, Sentence Transformers for embeddings, Qdrant for vector storage, and Groq for LLM-based answer generation. We evaluated Fixed Top-K, Similarity Threshold, and Adaptive Top-K retrieval strategies.

Challenges we ran into

Improving retrieval relevance, selecting the appropriate number of chunks, handling poorly matched queries, and ensuring answers remain grounded in source documents.

Accomplishments that we're proud of

We implemented an end-to-end RAG pipeline with adaptive retrieval, source metadata, grounded answer generation, and comparative retrieval experiments.

What we learned

We learned how embeddings, vector databases, adaptive retrieval, and LLMs work together, and why retrieval quality directly affects answer accuracy.

What's next for Adaptive RAG for Banking Compliance

Improve retrieval using reranking, hybrid search, and query reformulation, then evaluate answer faithfulness, retrieval relevance, latency, and overall performance against baseline methods.