ColPali Retrieval Optimization Demo

This repository demonstrates how to optimize ColPali-based PDF retrieval using Qdrant, with a focus on improving retrieval speed while maintaining quality. The project includes two Jupyter notebooks with experiments and results.

Repository Structure

• ColPali as a reranker I.ipynb Covers the initial setup and data preparation:

  • Combining ViDoRe, UFO, and DocVQA datasets into a single collection.
  • Uploading vectors to a Qdrant collection.

• ColPali as a reranker II.ipynb Focuses on retrieval experiments:

  • Using pooled vectors for first-stage retrieval and ColPali for reranking.
  • Comparing retrieval quality and speed between pooled and original vectors.
  • Exploring optimizations such as cutting tokens and binary quantization.