Launching today
multimodal-rag-showcase
Turns research PDFs, charts, and diagrams into RAG
2 followers
Turns research PDFs, charts, and diagrams into RAG
2 followers
Experimental multimodal RAG showcase with traceable answers, reviewed visual evidence, hybrid retrieval, and auditable evaluations.
Launch Team
Framer AI AgentsDesign and publish professional sites with AI
Promoted
Hey everyone!
I built this project because I wanted to explore a problem I kept seeing in RAG demos: PDFs are treated mostly as text, but a lot of valuable information lives inside charts, tables and diagrams.
So I started building a multimodal RAG system using Docling, local vision models, hybrid retrieval and source-aware citations. The challenging part wasn't retrieval itself, but making sure extracted evidence could be traced back to the original document and reviewed when uncertain. I also went a bit further with access controls, durable execution, observability and reproducible benchmarks. Mostly because I wanted to build and evaluate an actual engineering system, not another chatbot connected to PDFs lol.
This is an experimental engineering showcase, not a commercial product. The code and architecture are public on GitHub, and I'm considering making it open-source later. There are still plenty of limitations, particularly around visual extraction accuracy, answer quality and latency with self-hosted models.
I'm sharing it mainly to get technical feedback, exchange ideas and connect with others working on document intelligence and RAG. Would love to hear how others are approaching multimodal evidence extraction and evaluation!