0012026
Multimodal News RAG
Retrieval-augmented generation over 1,023 news article–image pairs, benchmarking text-only, caption-based, and actual-image retrieval.
- Built a news RAG pipeline combining text/image retrieval and vision-enabled generation.
- Created duplicate-aware data splits and benchmarked three retrieval approaches under fixed settings.
- Evaluated claim-level faithfulness with modality ablations, image-conflict tests, and bootstrap confidence intervals.
1,023 article–image pairs150 held-out articlesBootstrap confidence intervals