Practical overview
When it fits
Use VARAG when building or comparing RAG retrieval for scanned, multimodal or layout-rich documents, and select the retrieval method according to whether extracted text, cross-modal image matching or whole-page visual context matters.
Who it is for
- Developers prototyping retrieval over documents that mix text with images, diagrams, tables or complex layouts.
- Teams comparing OCR-based, cross-modal and page-image retrieval approaches.
Documented uses
- Extract and index text from scanned PDFs or images with Docling OCR.
- Search text and images in a shared embedding space with JinaCLIP-based Vision RAG.
- Retrieve visually rich document pages using ColPali page embeddings and late interaction.
- Combine image-embedding candidate retrieval with ColPali reranking in a hybrid approach.
Requirements & limits
- Simple RAG with OCR is described for text-heavy scanned material.
- Vision RAG uses JinaCLIP to search across text and images in a shared vector space.
- ColPali retrieves whole pages as images and uses token-to-patch late interaction.
- Hybrid ColPali first retrieves candidate pages with image embeddings, then reranks them with ColPali.
- The documented setup uses Python 3.10 and offers pip or Poetry installation; OCR dependencies are installed separately.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up VARAG and its OCR dependencies
Prepare the Python environment and install VARAG with OCR support for scanned PDFs or images.
Before you start
- Conda
- Python 3.10
- The VARAG repository
Steps
Clone the repository and change into its directory.
Create and activate a Conda environment named varag-venv with Python 3.10.
Install the project with pip install -e . or poetry install.
For OCR dependencies, install with pip install -e .["ocr"].
Expected outcome
A configured project environment with the optional OCR dependencies installed.
Guide limitations
- These setup instructions do not by themselves describe running a retrieval workflow.
Run the documented ColPali comparison demo on Modal
Launch VARAG's ColPali model comparison demo with heatmaps on Modal.
Before you start
- The VARAG repository and its installed dependencies
- Modal installed with pip install modal
- A Modal account configured with modal setup
- An hf-secret configured in the Modal dashboard with a Hugging Face token
Steps
Install Modal with pip install modal.
Configure the Modal account with modal setup.
Set up the hf-secret in the Modal dashboard with a Hugging Face token.
Run python -m modal run examples/inference_colpali/modal_demo_heatmaps_comparing_colpali_models.py::comparision_demo.
Expected outcome
The documented ColPali comparison demo is launched using Modal GPU resources.
Guide limitations
- The repository describes this as a ColPali comparison demo, not a general deployment procedure.
Compare options for Retrieval-augmented applications
These projects document a common task and share a project role. This is a source-based comparison, not an execution benchmark or a guarantee of interchangeability.
| Project | Documented purpose | Platform | Access and requirements |
|---|---|---|---|
| Varag | Extract and index text from scanned PDFs or images with Docling OCR. | Cloud | |
| Langsearch | Crawl website or local data for a knowledge base. | CLI, Self-hosted, Server | |
| Hybrid Search Rag | Combine document retrieval and language-model generation in a configurable workflow. |
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 16 October 2026
- Original website title
- adithya-s-k/VARAG · Official homepage title · retained in its source language