This Python script demonstrates a minimal setup for using the ChatCM-RAG (Retrieval-Augmented Generation) system to search biomedical literature using natural language queries. It combines semantic search via a sentence-transformer model with a pre-built FAISS index and structured metadata to return the most relevant research abstracts from a curated dataset.
The code first downloads the full ChatMed-RAG package from Hugging Face using snapshot_download(repo_id="fc28/ChatMed-RAG"). It then initializes three components: a sentence encoder (all-mpnet-base-v2) to embed queries, a FAISS index (index.faiss) to search for semantically similar documents, and a metadata file (medllm_metadata.csv) that contains information like titles, abstracts, PMIDs, and publication years.
The search() function accepts a query and returns the top-k most relevant results by performing vector similarity search. Each result includes the title, year, PMID, and a truncated abstract snippet.
