Advanced RAG Retrieval Benchmark — BM25, Dense & Hybrid Search

### Advanced RAG Retrieval Benchmark

Developed a Python-based Retrieval-Augmented Generation (RAG) retrieval benchmark to evaluate and compare different information retrieval approaches on a SaaS support knowledge base.

The project implements and compares three retrieval methods:

* **BM25:** Sparse keyword-based retrieval.

* **Dense Retrieval:** Semantic search using Sentence Transformers (`all-MiniLM-L6-v2`) and embedding similarity.

* **Hybrid Retrieval:** Combines dense and BM25 rankings using Reciprocal Rank Fusion (RRF).

The system was evaluated using a manually verified dataset containing both exact-term and paraphrased user queries.

Implemented evaluation metrics including:

* Recall@K

* Mean Reciprocal Rank (MRR)

* NDCG

* Hit Rate

The benchmark also includes graded relevance judgments to evaluate whether the retrieval system ranks highly relevant chunks above partially relevant information.

Finally, the three retrieval approaches were benchmarked on the same evaluation set, including an analysis of performance across exact-term and paraphrased queries.