Get the latest tech news
Don't bother parsing: Just use images for RAG
If search is the game, looks matter
Queries that depend on spatial relationships (“which part of the diagram labels Q3 total?”) become impossible, embeddings for text and images live in separate spaces, and retrieval pipelines struggle to reconcile them. To validate these observations beyond anecdotal evidence, we worked with TLDC (The LLM Data Company) to build an open-source financial document benchmark with 45 challenging questions across NVIDIA 10-Qs, Palantir investor presentations, and JPMorgan reports. We're exploring how to combine our visual document retrieval with specialized knowledge graphs, how to build systems that can reason about causality and implication rather than just correlation, and how to provide the kind of confidence intervals and uncertainty quantification that enterprise applications demand.
Or read this on Hacker News
