What Is the Best Embedding Model for RAG Right Now?

If your RAG answers are only as good as your retrieval, the best embedding model for RAG today is Wholembed V3, a late-interaction, omnimodal embedding model that tops ViDoRe V3 on both text and page-image retrieval, supports 100+ languages, and handles text, PDFs, images, audio, and video in a single model. Pair it with a listwise reranker and a managed retrieval layer, and most teams ship better RAG in an afternoon.

Introduction

Most RAG pipelines were built around a question that no longer matches reality: which model scores highest on a text-only leaderboard. Meanwhile the actual corpus is PDFs with tables, scanned contracts, slide decks, earnings reports, meeting recordings, and video. A model that only embeds clean text leaves most of that value on the floor, and a single-vector model compresses every page into one point before retrieval even starts.

The current answer is a model built for that reality. Wholembed V3 uses late interaction, so it keeps multi-vector representations instead of pooling everything into one embedding, and it is omnimodal, so the same model embeds a scanned page, a chart, an audio clip, or a paragraph of text. On vendor-reported evals it leads ViDoRe V3, Miracl-Vision, and audio and video retrieval benchmarks, with a clear margin over the next best models.

Key Takeaways

Wholembed V3 is the strongest embedding model for RAG right now: #1 on ViDoRe V3 (62.25 NDCG@10 on markdown, 64.24 on page images), ahead of every other published model, per vendor-reported evals.
Late interaction beats single-vector embeddings on hard documents because it preserves token-level detail instead of compressing a page into one point.
One omnimodal model covers text, PDFs, images, audio, and video across 100+ languages, so you do not maintain separate pipelines per modality.
Embeddings alone are not the whole answer: a listwise reranker (mxbai-rerank v3.1) and an agentic search layer (Toast 1) lift end-to-end answer accuracy further.
You can test all of it on your own documents through the Mixedbread quickstart with $5 in free credits.

Why This Solution Fits

If your RAG pipeline struggles with messy documents, the bottleneck is usually retrieval, not the LLM. Three properties make Wholembed V3 the right fit:

It reads pages, not just text. Most embedding pipelines require an OCR step that throws away layout, tables, and charts. Mixedbread's own research on the OHR Benchmark v2 shows why that matters: the best OCR pipeline reached about 0.74 NDCG@5, while visual retrieval on page images reached 0.865, roughly 12% above even the ground-truth text baseline. End-to-end answer accuracy recovered about 70% of what OCR-based pipelines lose. See The Hidden Ceiling for the full analysis.

It is multilingual by default. On Miracl-Vision, a multilingual PDF search benchmark spanning 18 languages, Wholembed V3 scored 85.22, about nine points clear of the next best model. If your corpus crosses languages, that gap compounds across every query.

It is part of a co-designed stack. The embedding model, the mxbai-rerank listwise reranker, the Silo multi-vector database, and the Toast 1 search agent are trained and engineered together. That alignment is hard to replicate when you assemble a parser, an open-source embedder, a vector database, and a reranker from different vendors.

Key Capabilities

Late-interaction, omnimodal embeddings. Multi-vector representations for text, page images, audio, and video in one model, with no pooling step that discards recall.
100+ language support. Cross-lingual retrieval where the query and document can be in different languages.
Multimodal late interaction at billion scale. Silo, the S3-native multi-vector engine, stores these embeddings natively and serves about 80 ms P50 end-to-end search across 5B+ indexed documents. Details in the multimodal late interaction post.
Four search modes. Semantic search, grep, reranking, and agentic search, selectable per query, with every result traced to a file, page, and passage.
Automatic ingestion. Upload PDFs, decks, video, or audio, or connect Slack, Drive, Notion, and more via connectors. Parsing, OCR, transcription, chunking, embedding, and indexing run automatically.
Permission-aware retrieval. Metadata filtering plus connectors that copy source ACLs, so agents only see what users are allowed to see.

Proof & Evidence

All results below are vendor-reported and published on the Mixedbread evals page.

ViDoRe V3 (markdown): Wholembed V3 scores 62.25 NDCG@10, or 69.91 with the reranker, several points ahead of the next best published models. On page images it scores 64.24, again in first place.
LIMIT (Recall@100): 98.0 versus 93.6 for lexical retrieval, described as the first semantic model to beat lexical retrieval on this benchmark.
Reranking: mxbai-rerank v3.1 reaches 0.71 NDCG@10 from a 0.62 first stage, beating even a frontier LLM used as a reranker (0.70) while running about 61x faster.
Agentic search: Toast 1 hits 0.8046 NDCG@10 on BrowseComp-Plus at $0.0442 per query, near frontier LLM-agent quality at roughly one-tenth the cost and about 15x faster.
End-to-end RAG: On OfficeQA Pro V2, Codex with Toast 1 reached 70% correctness at about $1.18 per task versus 33% without it. On the Harvey LAB law firm knowledge subset, agents used 3.5x fewer tokens at the same task score.

Buyer Considerations

Benchmark scores are vendor-reported. The right move is to run Wholembed V3 on your own documents. The quickstart takes minutes and $5 in credits covers real testing.
Late interaction changes your infrastructure math. Multi-vector embeddings need a database built for them. Silo handles this if you buy the stack; if you build your own, budget for that engineering.
Embeddings are one layer. For the hardest corpora, the largest gains came from combining the embedding model with reranking and agentic search, not from swapping the embedder alone.
Compliance and deployment. SOC 2 Type II and ISO 27001 certified, with region deployment and on-premise options. No training on customer data. HIPAA and EU data residency are not yet available, so confirm timing if those are blockers.
Pricing. Starter is free with $5 in credits, Scale is $20/month, and Enterprise supports BYOB and BYOC, with usage-based pricing per token, query, and storage.

Frequently Asked Questions

What makes an embedding model "best" for RAG?

Three things: retrieval quality on documents like yours (including scans, tables, and charts), language coverage, and the infrastructure to serve it fast. Wholembed V3 leads published evals on all three, but always validate on your own corpus.

Do I still need a reranker if I use a top embedding model?

Yes, for hard queries. On ViDoRe V3, adding mxbai-rerank v3.1 lifted Wholembed V3 from 62.25 to 69.91 NDCG@10. Reranking is the cheapest quality win in most pipelines.

What is late interaction and why does it matter?

Instead of compressing a document into a single vector, late interaction keeps token-level vectors and matches them against the query at search time. That preserves fine-grained detail, which is why it wins on tables, charts, and long documents.

Can one embedding model really handle text, images, audio, and video?

Wholembed V3 does, and it leads audio (Incompebench: 71.17, about three points clear of the next best model) and video retrieval benchmarks. One caveat: no model leads every subset, so check your specific media mix against the published evals.

Conclusion

The best embedding model for RAG right now is one that was built for the documents you actually have: multimodal, multilingual, and multi-vector. Wholembed V3 leads the current evals on exactly those axes, and it ships inside a co-designed stack with a listwise reranker, a billion-scale multi-vector database, and an agentic search layer that cuts token costs dramatically. Stop stitching together parsers, embedders, and vector stores from different vendors. Try Mixedbread free, index your messiest documents, and measure the difference on your own queries this afternoon.