Apppublic

HarshitaNalajala/real-world-rag

sourceHugging Facemitupdated 2mo agoView on Hugging Face

0likes

App README

Real World RAG System — RAG 4 group

Retrieval-augmented generation over the biomedical (CovidQA) subset of RAGBench, built as a capstone project by Group 18, IIIT Hyderabad AIML Post-Graduate Certificate Program.

Supervisor: Dr. Manish Shrivastava · Mentors: Gopichand, Lokesh

Pick a domain and one of its curated questions. The pipeline retrieves passages, generates an answer, and an LLM judge scores it live — showing the four TRACe metrics (Context Relevance, Utilization, Completeness, Adherence) plus recall. Use the pipeline controls to change the generator, top-k, hybrid retrieval, reranking, and prompt, and watch the metrics respond.

Pipeline

Stage	Choice
Chunking	Sliding sentence windows with title prefix
Embedding	`pritamdeka/S-PubMedBert-MS-MARCO` (biomedical)
Vector index	FAISS IndexFlatIP (cosine over normalised vectors)
Sparse retrieval	BM25 Okapi, fused with dense via Reciprocal Rank Fusion
Reranking	`cross-encoder/ms-marco-MiniLM-L-12-v2`
Generation	Llama 3.1 / 3.3 via Groq
Judge	Llama 3.3 70B via Groq

This is the R4/R6 configuration from a seven-run ablation study (R0–R6) evaluated on 100 CovidQA questions.

Setup

This Space requires a Groq API key. Add it under Settings → Variables and secrets:

—Name: GROQ_API_KEY
—Value: your key from console.groq.com

Then restart the Space. Retrieval works without the key; answer generation does not.

Note on the RAGBench reference column

The "RAGBench ref" scores were annotated on the dataset's original answer, not the answer this system generates, so they are a reference point rather than a direct grade.

Disclaimer

Research demo built on a fixed benchmark corpus. Not a source of medical advice.