fastbmRAG

FastbmRAG, an fast lightweight graph-based RAG optimized for biomedical literature.

Utilizing well organized structure of biomedical papers, fastbmRAG divides the construction of knowledge graph into two stages:

• Build a draft knowledge graph using only abstracts of scientific papers
• Refine the nodes and edges of knowledge graph using main texts

Our evaluations demonstrate that fastbmRAG is over 10x faster than existing graph-RAG tools and achieve superior coverage and accuracy to existing knowledge.

Overall, fastbmRAG provides a fast solution for quickly understanding, summarizing, and answering questions about biomedical literature on a large scale.

Install and build conda envirenment

fastbmRAG is implemented in Python 3.11 or higher.

Current version supports both ollama and vllm LLM backends.

For ollama backend: Please install ollama by visiting: https://ollama.com/download The ollama models should be installed before usage with a command like:

ollama run gemma3

For vllm backend: vllm will be automatically installed via pip. You can use any compatible model with vllm.

Backend selection: Use the --backend parameter to choose between 'ollama' and 'vllm':

• --backend ollama (default)
• --backend vllm (not support embedding now. It still needs ollama embedding model)

Build a conda environment for necessary dependencies:

git clone https://github.com/menggf/fastbmRAG.git
cd fastbmRAG
conda env create -f environment.yml
conda activate fastbmrag

Usage

There are two modes: 'update' and 'query', which are used to build local collection and query local collection, respectively.

Build local collection

The 'update' mode is used to create an new collection or add new documents to an existing collection of vector database. The documents should be a text file in CSV format.

It should have at least three columns: 'abstract', 'main_text' and 'paper_id'. If there are more columns, they are used for additional information. Each element of the 'main_text' column should be either a list of strings in the format '[str1, str2]' or a string separated by '\n'. 'paper_id' is a unique ID for each paper. If a paper ID exists in the collection, the corresponding paper will be ignored.

To update the collection, use the following command:

python main.py --job update --document examples/demo1.csv
			   --collection_name collection_name 
			   --working_dir directory_path
               --llm_update_model_name llm_model_name
               --embed_model_name embedding_model_name
               --backend ollama  # or --backend vllm

Here, ‘collection_name’ and ‘working_dir’ specify the collection name and directory to store collection. llm_model_name and embedding_model_name are the ollama model names.

Query the collection

It is used to query the collection with questions.

python main.py --job query --collection_name collection_name 
	           --working_dir directory_path 
	           --question 'your question'
               --llm_query_model_name llm_model_name
               --embed_model_name embedding_model_name
               --backend ollama  # or --backend vllm

Related publication

Guofeng Meng et.al., fastbmRAG: A Fast Graph-Based RAG Framework for Efficient Processing of Large-Scale Biomedical Literature. ArXiv, 2025, DOI:10.48550/arXiv.2511.10014