Retrieval-Augmented Generation on ViDoRe Fin V3
58.4Accuracyvisual RAG
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| visual RAGRetrieval Basis=page images, Retrieval Model=ColEmbed-3B-v2, LLM=Gemini-3-Flash2026.05 | 58.4 | 41.1 | 273 | |
| MinerU-PopoRetrieval Basis=node title path and summary, LLM=Gemini-3-Flash2026.05 | 49.5 | 40.5 | 68 | |
| raw RAGRetrieval Basis=full text OCR, LLM=Gemini-3-Flash2026.05 | 48.7 | 32.4 | 225 |