Is OCR quality the silent killer of your RAG performance? 🧵👇. Our deep dive into 8500+ enterprise documents reveals how OCR errors create a "hidden ceiling" for RAG, and how multimodal approaches offer a path forward.

The takeaway: OCR caps your RAG. Multimodal retrieval offers a path to higher accuracy by overcoming text extraction challenges. Combining multimodal retrieval with quality OCR text is the best way forward for now.

What it means for real word setting? Find out more here:

4/🧵 Finding 4: Vision-only generation is promising but not ready for multi-doc RAG. Feeding page images directly to the LLM for generation yielded the lowest accuracy (worse than high-quality OCR text). Multimodal excels at finding docs visually today; generating from multiple

4/🧵 Finding 3: Better retrieval directly means better answers. Using standard OCR text for generation resulted in ~26% fewer correct answers than with perfect text. But using multimodal retrieval (while still feeding OCR text to the LLM) recovered 70% of that lost accuracy.

3/🧵 Finding 2: Multimodal retrieval breaks the OCR ceiling. By "seeing" page images, multimodal retrieval outperformed all text methods, including perfect ground-truth text. It achieved an average NDCG@5 nearly 12% higher than perfect OCR.

2/🧵 Finding 1: OCR creates a real performance ceiling for text-based RAG retrieval. Even the best OCR solutions fall ~4.5% short of ground-truth text (NDCG@5) on complex enterprise documents. It's a bottleneck you can't ignore. (Interesting side note: BM25 outperformed

Full blog post:

RT @juliuslipp: new bottles out now!!! come get yours today!.

RT @juliuslipp: together with @aaxsh18, @xmlee97 and @ rui. the only thing missing is you 🫰.

Read more in our blog post: And check models on @huggingface:

Great Performance vs. Latency Trade-Off

Multistep Training for Better Understanding

ToolRet Benchmark. mxbai-rerank-v2 also achieves SOTA on the large-scale ToolRet test, showing advanced code & tool usage understanding. Ideal for MCP or any scenario that needs the best-fitting tools.

SOTA on BEIR Benchmark. Despite being multilingual, mxbai-rerank-v2 still outperforms both open & closed-source English models—showcasing strong language understanding and domain generalisability.

What’s New?. • 100+ languages, extended context (8–32K).• GRPO: RL for better understanding.• Up to 8× faster.• Use cases: MCP, Code, SQL, JSON, Documents. • Flexible & Open: Self-host or use our API.

Baked-in Brilliance: Reranking Meets RL 🍞. Meet mxbai-rerank-v2, our second-gen rerankers built on Qwen2.5 (thanks, @Alibaba_Qwen) & refined with GRPO from @deepseek_ai. They outperform open & closed-source models while staying fully open. 👇

RT @juliuslipp: 5% of all HF downloads go to mxbai-embed :o

RT @juliuslipp: just revamped the hiring page of Mixedbread a bit. we're looking for amazing people to join us doing:.- full stack (next.j….

RT @notrab: I posted a new tutorial on how to process embeddings with @mixedbreadai using @redpandadata and @tursodatabase ✨. Link in reply….