Irpapers – Visual embeddings vs. OCR trade-offs in scientific PDFs
Hey HN, we are releasing IRPAPERS to answer a highly pragmatic question: when building a RAG pipeline over PDFs, should you OCR the text or just embed the raw page images? Processing PDFs in production usually involves stringing together brittle OCR heuristics. While recent multimodal embeddings (like ColModernVBERT or ColPali) allow you to skip OCR entirely and retrieve directly from visual layouts, we wanted to measure if the computational overhead is actually worth the utility. The short answer: Transformer-based image pipelines won't be perfect for every use-case, but they fix exactly…
What it does
In the maker’s words, at launch
Hey HN, we are releasing IRPAPERS to answer a highly pragmatic question: when building a RAG pipeline over PDFs, should you OCR the text or just embed the raw page images? Processing PDFs in production usually involves stringing together brittle OCR heuristics. While recent multimodal embeddings (like ColModernVBERT or ColPali) allow you to skip OCR entirely and retrieve directly from visual layouts, we wanted to measure if the computational overhead is actually worth the utility. The short answer: Transformer-based image pipelines won't be perfect for every use-case, but they fix exactly what OCR breaks. Here is what we found benchmarking 3,230 pages of dense scientific literature: Complementary Bottlenecks: Text representations (BM25 + dense vectors) are highly efficient for exact lexical constraints (e.g., finding a specific acronym like "HyDE"). Conversely, image embeddings shine on spatial architecture diagrams and t-SNE plots where OCR serialization just turns into structural garbage. Multimodal Hybrid Search: Because these failure modes are almost perfectly orthogonal, fusing the two signals gives you the best performance out of the box. By combining them, we pushed top-1 recall to 49% (beating text alone at 46%). The Memory Constraint: Late-interaction image embeddings produce thousands of vectors per page, creating a massive storage bottleneck. To address this need, we evaluate MUVERA encoding. Under the hood, this compresses multi-vector representations into a single fixed-dimensional encoding via SimHash, allowing you to use standard HNSW indexing without the paralyzing memory overhead. In practice, if you are building a RAG workflow today, text-based context still provides higher downstream utility for the actual generation step (0.82 vs 0.71 alignment). Instead of picking one modality and dealing with its blind spots, start with hybrid text search as a sensible default, and inject multi-vector image embeddings to catch the visual edge-cases. We’ve open-sourced the benchmark and the evaluation recipes: Paper https://arxiv.org/abs/2602.17687 IRPAPERS dataset on HuggingFace at huggingface.co/weaviate/IRPAPERS and GitHub at github.com/weaviate/IRPAPERS Our experimental code is also available on GitHub at github.com/weaviate/query-agent-benchmarking Happy to answer any questions about the evaluation pipeline, the cold start problem of visual benchmarks, or the specific retrieval trade-offs we saw.
Does the same job
all alternatives →
- CACrop and OCR PDFs in the Browser2024 · pdf.darefail.com · ▲5
- TAtesserocr – A Python wrapper for the tesseract-ocr API2016 · github.com · ▲78
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 26d ago · cactuscompute.com
