VisRAG

VisRAG 2.0: Evidence-Guided Multi-Image Reasoning in Visual Retrieval-Augmented Generation

📖 Introduction

EVisRAG (VisRAG 2.0) is an evidence-guided Vision Retrieval-augmented Generation framework that equips VLMs for multi-image questions by first linguistically observing retrieved images to collect per-image evidence, then reasoning over those cues to answer. EVisRAG trains with Reward-Scoped GRPO, applying fine-grained token-level rewards to jointly optimize visual perception and reasoning.

VisRAG is a novel vision-language model (VLM)-based RAG pipeline. In this pipeline, instead of first parsing the document to obtain text, the document is directly embedded using a VLM as an image and then retrieved to enhance the generation of a VLM. Compared to traditional text-based RAG, VisRAG maximizes the retention and utilization of the data information in the original documents, eliminating the information loss introduced during the parsing process.

🎉 News

• 20260709: Our work EVisRAG is accepted by ACM MM 2026 🎉.
• 20251207: Released all benchmarks on HuggingFace.
• 20251118: Both EVisRAG and VisRAG can be easily reproduced within UltraRAG v2.
• 20251022: We upload all evaluation benchmarks in VisRAG Collections
• 20251014: Released EVisRAG-3B on HuggingFace.
• 20251014: Released EVisRAG (VisRAG 2.0), an end-to-end Vision-Language Model. Released our Paper on arXiv. Released our Model on HuggingFace. Released our Code on GitHub
• 20241111: Released our VisRAG Pipeline on GitHub, now supporting visual understanding across multiple PDF documents.
• 20241104: Released our VisRAG Pipeline on Hugging Face Space.
• 20241031: Released our VisRAG Pipeline on Colab. Released codes for converting files to images, which could be found at visrag_scripts/file2img.
• 20241015: Released our train data and test data on Hugging Face which can be found in the VisRAG Collection on Hugging Face. It is referenced at the beginning of this page.
• 20241014: Released our Paper on arXiv. Released our Model on Hugging Face. Released our Code on GitHub.

✨ VisRAG Pipeline

EVisRAG

EVisRAG is an end-to-end framework that equips VLMs with precise visual perception during reasoning in multi-image scenarios. We trained and released VLRMs with EVisRAG built on Qwen2.5-VL-7B-Instruct, and Qwen2.5-VL-3B-Instruct.

VisRAG-Ret

VisRAG-Ret is a document embedding model built on MiniCPM-V 2.0, a vision-language model that integrates SigLIP as the vision encoder and MiniCPM-2B as the language model.

VisRAG-Gen

In the paper, we use MiniCPM-V 2.0, MiniCPM-V 2.6, and GPT-4o as the generators. Actually, you can use any VLMs you like!

⚙️ Setup

EVisRAG

git clone https://github.com/OpenBMB/VisRAG.git
conda create --name EVisRAG python==3.10
conda activate EVisRAG
cd VisRAG
pip install -r EVisRAG_requirements.txt

VisRAG

git clone https://github.com/OpenBMB/VisRAG.git
conda create --name VisRAG python==3.10.8
conda activate VisRAG
conda install nvidia/label/cuda-11.8.0::cuda-toolkit
cd VisRAG
pip install -r requirements.txt
pip install -e .
cd timm_modified
pip install -e .
cd ..

Note:

1. timm_modified is an enhanced version of the timm library that supports gradient checkpointing, which we use in our training process to reduce memory usage.

⚡️ Training

EVisRAG

To train EVisRAG effectively, we introduce Reward-Scoped Group Relative Policy Optimization (RS-GRPO), which binds fine-grained rewards to scope-specific tokens to jointly optimize visual perception and reasoning abilities of VLMs.

Stage1: SFT (based on LLaMA-Factory)

git clone https://github.com/hiyouga/LLaMA-Factory.git 
bash evisrag_scripts/full_sft.sh

Stage2: RS-GRPO (based on Easy-R1)

bash evisrag_scripts/run_rsgrpo.sh

Notes:

1. The training data is available on Hugging Face under EVisRAG-Train, which is referenced at the beginning of this page.
2. We adopt a two-stage training strategy. In the first stage, please clone LLaMA-Factory and update the model path in the full_sft.sh script. In the second stage, we built our customized algorithm RS-GRPO based on Easy-R1, specifically designed for EVisRAG, whose implementation can be found in src/RS-GRPO.

VisRAG-Ret

Our training dataset of 362,110 Query-Document (Q-D) Pairs for VisRAG-Ret is comprised of train sets of openly available academic datasets (34%) and a synthetic dataset made up of pages from web-crawled PDF documents and augmented with VLM-generated (GPT-4o) pseudo-queries (66%).

bash visrag_scripts/train_retriever/train.sh 2048 16 8 0.02 1 true false config/deepspeed.json 1e-5 false wmean causal 1 true 2 false <model_dir> <repo_name_or_path>

Note:

1. Our training data can be found in the VisRAG collection on Hugging Face, referenced at the beginning of this page. Please note that we have separated the In-domain-data and Synthetic-data due to their distinct differences. If you wish to train with the complete dataset, you’ll need to merge and shuffle them manually.
2. The parameters listed above are those used in our paper and can be used to reproduce the results.
3. <repo_name_or_path> can be any of the following: openbmb/VisRAG-Ret-Train-In-domain-data, openbmb/VisRAG-Ret-Train-Synthetic-data, the directory path of a repository downloaded from Hugging Face, or the directory containing your own training data.
4. If you wish to train using your own datasets, remove the --from_hf_repo line from the train.sh script. Additionally, ensure that your dataset directory contains a metadata.json file, which must include a length field specifying the total number of samples in the dataset.
5. Our training framework is modified based on OpenMatch.

VisRAG-Gen

The generation part does not use any fine-tuning, we directly use off-the-shelf LLMs/VLMs for generation.

📃 Evaluation

EVisRAG

bash evisrag_scripts/predict.sh
bash evisrag_scripts/eval.sh 

Notes:

1. The test data is available on Hugging Face under EVisRAG-Test-xxx, as referenced at the beginning of this page.
2. To run evaluation, first execute the predict.sh script. The model outputs will be saved in the preds directory. Then, use the eval.sh script to evaluate the predictions. The metrics EM, Accuracy, and F1 will be reported directly.

VisRAG-Ret

bash visrag_scripts/eval_retriever/eval.sh 512 2048 16 8 wmean causal ArxivQA,ChartQA,MP-DocVQA,InfoVQA,PlotQA,SlideVQA <ckpt_path>

Note:

1. Our test data can be found in the VisRAG Collection on Hugging Face, which is referenced at the beginning of this page.
2. The parameters listed above are those used in our paper and can be used to reproduce the results.
3. The evaluation script is configured to use datasets from Hugging Face by default. If you prefer to evaluate using locally downloaded dataset repositories, you can modify the CORPUS_PATH, QUERY_PATH, QRELS_PATH variables in the evaluation script to point to the local repository directory.

VisRAG-Gen

There are three settings in our generation: text-based generation, single-image-VLM-based generation and multi-image-VLM-based generation. Under single-image-VLM-based generation, there are two additional settings: page concatenation and weighted selection. For detailed information about these settings, please refer to our paper.

python visrag_scripts/generate/generate.py \
--model_name <model_name> \
--model_name_or_path <model_path> \
--dataset_name <dataset_name> \
--dataset_name_or_path <dataset_path> \
--rank <process_rank> \
--world_size <world_size> \
--topk <number of docs retrieved for generation> \
--results_root_dir <retrieval_results_dir> \
--task_type <task_type> \
--concatenate_type <image_concatenate_type> \
--output_dir <output_dir>

Note:

1. use_positive_sample determines whether to use only the positive document for the query. Enable this to exclude retrieved documents and omit topk and results_root_dir. If disabled, you must specify topk (number of retrieved documents) and organize results_root_dir as results_root_dir/dataset_name/*.trec.
2. concatenate_type is only needed when task_type is set to page_concatenation. Omit this if not required.
3. Always specify model_name_or_path, dataset_name_or_path, and output_dir.
4. Use --openai_api_key only if GPT-based evaluation is needed.

🔧 Usage

EVisRAG

Model on Hugging Face: https://huggingface.co/openbmb/EVisRAG-7B

from transformers import AutoProcessor
from vllm import LLM, SamplingParams
from qwen_vl_utils import process_vision_info

def evidence_promot_grpo(query):
    return f"""You are an AI Visual QA assistant. I will provide you with a question and several images. Please follow the four steps below:

Step 1: Observe the Images
First, analyze the question and consider what types of images may contain relevant information. Then, examine each image one by one, paying special attention to aspects related to the question. Identify whether each image contains any potentially relevant information.
Wrap your observations within <observe></observe> tags.

Step 2: Record Evidences from Images
After reviewing all images, record the evidence you find for each image within <evidence></evidence> tags.
If you are certain that an image contains no relevant information, record it as: [i]: no relevant information(where i denotes the index of the image).
If an image contains relevant evidence, record it as: [j]: [the evidence you find for the question](where j is the index of the image).

Step 3: Reason Based on the Question and Evidences
Based on the recorded evidences, reason about the answer to the question.
Include your step-by-step reasoning within <think></think> tags.

Step 4: Answer the Question
Provide your final answer based only on the evidences you found in the images.
Wrap your answer within <answer></answer> tags.
Avoid adding unnecessary contents in your final answer, like if the question is a yes/no question, simply answer "yes" or "no".
If none of the images contain sufficient information to answer the question, respond with <answer>insufficient to answer</answer>.

Formatting Requirements:
Use the exact tags <observe>, <evidence>, <think>, and <answer> for structured output.
It is possible that none, one, or several images contain relevant evidence.
If you find no evidence or few evidences, and insufficient to help you answer the question, follow the instruction above for insufficient information.

Question and images are provided below. Please follow the steps as instructed.
Question: {query}
"""

model_path = "xxx"
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True, padding_side='left')

imgs, query = ["imgpath1", "imgpath2", ..., "imgpathX"], "What xxx?"
input_prompt = evidence_promot_grpo(query)

content = [{"type": "text", "text": input_prompt}]
for imgP in imgs:
    content.append({
        "type": "image",
        "image": imgP
    })
msg = [{
          "role": "user",
          "content": content,
      }]

llm = LLM(
    model=model_path,
    tensor_parallel_size=1,
    dtype="bfloat16",
    limit_mm_per_prompt={"image":5, "video":0},
)

sampling_params = SamplingParams(
    temperature=0.1,
    repetition_penalty=1.05,
    max_tokens=2048,
)

prompt = processor.apply_chat_template(
    msg,
    tokenize=False,
    add_generation_prompt=True,
)

image_inputs, _ = process_vision_info(msg)

msg_input = [{
    "prompt": prompt,
    "multi_modal_data": {"image": image_inputs},
}]

output_texts = llm.generate(msg_input,
    sampling_params=sampling_params,
)

print(output_texts[0].outputs[0].text)

VisRAG-Ret

Model on Hugging Face: https://huggingface.co/openbmb/VisRAG-Ret

from transformers import AutoModel, AutoTokenizer
import torch
import torch.nn.functional as F
from PIL import Image
import os

def weighted_mean_pooling(hidden, attention_mask):
    attention_mask_ = attention_mask * attention_mask.cumsum(dim=1)
    s = torch.sum(hidden * attention_mask_.unsqueeze(-1).float(), dim=1)
    d = attention_mask_.sum(dim=1, keepdim=True).float()
    reps = s / d
    return reps

@torch.no_grad()
def encode(text_or_image_list):
    
    if (isinstance(text_or_image_list[0], str)):
        inputs = {
            "text": text_or_image_list,
            'image': [None] * len(text_or_image_list),
            'tokenizer': tokenizer
        }
    else:
        inputs = {
            "text": [''] * len(text_or_image_list),
            'image': text_or_image_list,
            'tokenizer': tokenizer
        }
    outputs = model(**inputs)
    attention_mask = outputs.attention_mask
    hidden = outputs.last_hidden_state

    reps = weighted_mean_pooling(hidden, attention_mask)   
    embeddings = F.normalize(reps, p=2, dim=1).detach().cpu().numpy()
    return embeddings

model_name_or_path = "openbmb/VisRAG-Ret"
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True)
model = AutoModel.from_pretrained(model_name_or_path, torch_dtype=torch.bfloat16, trust_remote_code=True)
model.eval()

script_dir = os.path.dirname(os.path.realpath(__file__))
queries = ["What does a dog look like?"]
passages = [
    Image.open(os.path.join(script_dir, 'test_image/cat.jpeg')).convert('RGB'),
    Image.open(os.path.join(script_dir, 'test_image/dog.jpg')).convert('RGB'),
]

INSTRUCTION = "Represent this query for retrieving relevant documents: "
queries = [INSTRUCTION + query for query in queries]

embeddings_query = encode(queries)
embeddings_doc = encode(passages)

scores = (embeddings_query @ embeddings_doc.T)
print(scores.tolist())

VisRAG-Gen

For VisRAG-Gen, you can explore the VisRAG Pipeline on Google Colab which includes both VisRAG-Ret and VisRAG-Gen to try out this simple demonstration.

📄 License

• The code in this repo is released under the Apache-2.0 License.
• The usage of VisRAG-Ret model weights must strictly follow MiniCPM Model License.md.
• The models and weights of VisRAG-Ret are completely free for academic research. After filling out a "questionnaire" for registration, VisRAG-Ret weights are also available for free commercial use.

📧 Contact

EVisRAG

• Yubo Sun: [email protected]
• Chunyi Peng: [email protected]

VisRAG

• Shi Yu: [email protected]
• Chaoyue Tang: [email protected]

📈 Star History