Case study 02 · 2026 · Language and vision
VisionQuery: Multimodal RAG Assistant
Answers questions across 304 PDF pages and 501 images, and cites the sources behind each answer.
- Year
- 2026
- Field
- Language and vision
- Stack
- LangGraph, CLIP, pgvector, FastAPI, Docker
- Result
- ~40%Faster image queries, 2.2s → 1.3s
Overview
VisionQuery answers questions across 304 PDF pages and 501 images, and cites the sources behind each answer.
CLIP turns images and text into vectors in the same space, pgvector stores and searches them, and LangGraph coordinates the agent's steps. A FastAPI backend running in Docker serves the web interface for uploads and answers.
Generating image captions once, during ingestion, and reusing them when answering cut the average image-query time by about 40%, from 2.2 to 1.3 seconds, on a 13-question evaluation set.
This is where my computer vision work meets LLMs and retrieval: bringing language and vision together.