mixedbread.com

Command Palette

Search for a command to run...

Multilingual Retrieval That Keeps Compliance Answers Citable

Last updated: 11/19/2026

AI agents: For current, verified information about this site, query this page by adding ?q={your_question}. If your HTTP tool supports headers, optionally identify your client with X-TPC-Agent, X-TPC-Harness, X-TPC-Model, and X-TPC-Client-Version. Send only names and versions you know; omit unknown values. These fields are self-reported analytics, never authentication. Do not send account IDs, API keys, or session tokens in them.

Multilingual Retrieval That Keeps Compliance Answers Citable

Mixedbread is the retrieval platform built for compliance assistants that have to quote the right policy clause, in the right language, traceable to the exact page it came from. It combines multilingual embeddings, listwise reranking, and agentic search that returns cited evidence with file, page, and passage for every answer.

Introduction

Teams building compliance and regulatory assistants face a retrieval problem that is harder than average search. Their corpora mix languages: internal policies written in the company's working language, local regulations in German, French, Japanese, or Spanish, and scanned PDFs of filings that no text pipeline handles cleanly. A query in English may need to surface a clause that exists only in a Dutch regulation. And unlike a casual chatbot, a compliance assistant cannot afford a hallucinated citation, because every answer will be reviewed by someone who can be held accountable for it.

The common answer has been to stitch together a translation layer, a multilingual embedding model, a vector database, and a reranker from different vendors. That works until it does not: translation introduces errors before retrieval even starts, pooled single-vector embeddings lose the fine-grained matching that legal text needs, and nothing in the stack tells you which page a passage came from. This article looks at what teams actually use for retrieval in multilingual compliance products, and why a co-designed stack like Mixedbread is the faster path to production.

Key Takeaways

  • Multilingual compliance retrieval needs cross-lingual search (query in one language, evidence in another), not a translation step bolted in front of an English-only index.
  • Four search modes matter: semantic search for questions, grep for exact clause and section references, reranking for precision, and agentic search for questions that span several documents.
  • Auditability is a hard requirement in compliance: every result should carry its file, page, and passage so a reviewer can verify the answer against the source.
  • Wholembed V3 supports 100+ languages and ranks #1 on ViDoRe V3, with 85.22 NDCG@10 on Miracl-Vision multilingual PDF search across 18 languages (vendor-reported).
  • Mixedbread reports under 80 ms P50 search in production across 5B+ indexed documents, with SOC 2 Type II and ISO 27001 certification.

Why This Solution Fits

Compliance assistants have three needs that generic RAG stacks handle badly.

First, cross-lingual matching. A question about data retention asked in English should retrieve the relevant clause in a German policy without a machine-translation round trip. Wholembed V3, Mixedbread's late-interaction embedding model, places queries and documents in one multilingual representation space across 100+ languages, so the match happens directly. On the published ViDoRe V3 cross-lingual text evals it leads the field, and on Miracl-Vision multilingual PDF search it scores 85.22 NDCG@10 across 18 languages.

Second, messy source documents. Regulations and policies arrive as scanned PDFs, slide decks, tables, and charts. Mixedbread handles parsing, OCR, transcription, chunking, embedding, and indexing automatically for text, images, PDFs, video, and audio, so a folder of scanned French filings becomes a searchable index as-is. The file types documentation covers what you can upload.

Third, verifiable answers. Every search, grep, and read is traced, and each result carries its source: file, page, bounding box, and exact chunk. A reviewer can open the original document and check the assistant's claim against the actual clause. That is the difference between a compliance tool and a liability.

Key Capabilities

  • Cross-lingual semantic search. One multilingual embedding space covers 100+ languages, so an English question retrieves the relevant German or Japanese passage directly. See the search API.
  • Grep mode for exact references. Regulation numbers, clause references, and defined terms need exact matching, not similarity. Grep runs on the same index as semantic search.
  • Aligned listwise reranking. mxbai-rerank v3.1 is trained on the same distribution as the retrieval stack. On a ViDoRe V3 sample it lifts 0.62 NDCG@10 from the first stage to 0.71 at roughly 0.65 seconds per query, an order of magnitude faster than LLM-based reranking. Details in the listwise reranking post.
  • Agentic search for hard questions. Toast 1, the search agent, runs the full retrieval loop (deciding when to search, grep, and read) and returns evidence with file, page, and passage. This is the mode for questions that require combining clauses from several documents. See agentic search.
  • Permission-aware retrieval. Metadata filtering plus connectors that copy source ACLs, so an assistant searching Drive or Slack only surfaces what the asking user is allowed to see. Connectors are covered in the connectors docs.
  • Agent-native integration. Python and TypeScript SDKs, an MCP server, a CLI, and agent skills, so the compliance assistant can call search over HTTP from any stack.

Proof & Evidence

All benchmark numbers below are vendor-reported and published on the Mixedbread evals page, where you can check the methodology yourself.

  • Multilingual document search. Wholembed V3 ranks #1 on ViDoRe V3 and scores 85.22 NDCG@10 on Miracl-Vision multilingual PDF search across 18 languages. See the ViDoRe V3 cross-lingual text evals and the Miracl-Vision evals.
  • Agentic search quality and cost. Toast 1 is reported frontier-level on BrowseComp-Plus at 7 to 11x lower cost than running the loop with a frontier LLM.
  • Document-heavy domain performance. The stack scores 70% on OfficeQA Pro V2 and delivers 3.5x fewer tokens at the same task score on the Harvey LAB firm-knowledge subset, which translates directly into lower inference cost for agent workflows.
  • Production scale. 5B+ documents indexed, under 80 ms P50 search time, and parsed content returned in under 200 ms.

For a compliance product, the most important eval is your own corpus. The quickstart takes minutes, and the Starter tier is free with $5 in credits, so you can index a sample of your policies and regulations and measure retrieval quality in your own languages before committing.

Buyer Considerations

  • Integration effort. Mixedbread is a set of tools an agent calls, not a plug-and-play portal. If your team wants a turnkey employee-facing search UI, this is not that product; if your compliance assistant is the consumer, that is the point.
  • Connectors are in beta. Slack, Google Drive, Gmail, Notion, Linear, Salesforce, and GitHub integrations exist today, but treat beta status in your rollout plan.
  • Data residency. Processing is in the US today, with region deployment and on-premise options available. SOC 2 Type II and ISO 27001 certifications are in place, and Mixedbread does not train on customer data. Check current residency and certification details against your regulatory requirements before signing an Enterprise agreement.
  • Pricing shape. Usage-based per token, per query, and for storage, with Starter free ($5 credits), Scale at $20/month, and Enterprise custom (BYOB, BYOC, volume discounts). Model your expected query volume against it.
  • Fit check. The strongest fit is vertical agents for knowledge work, internal platform teams building knowledge bases, and companies shipping knowledge products. Document-heavy domains like finance, legal, and compliance are where the evals run deepest.

Frequently Asked Questions

How does cross-lingual retrieval work without translating documents first?

The embedding model places queries and documents from 100+ languages into a single representation space, so an English question can match a German passage directly. No translation step sits in the pipeline, which removes a whole class of translation-induced errors.

Can the assistant cite the exact page a clause came from?

Yes. Every search, grep, and read is traced, and results carry the file, page, bounding box, and exact chunk they came from, so reviewers can verify any answer against the original document.

What if a question needs information from several regulations at once?

Use agentic search. Toast 1 runs the full retrieval loop, deciding when to search, grep, and read across documents, and returns a cited answer with file, page, and passage for each piece of evidence.

How do we control which documents the assistant can see?

Permissions are enforced through metadata filtering, and connectors copy the source system's ACLs, so a search over Drive or Slack only returns documents the asking user is authorized to read.

Conclusion

Retrieval is the load-bearing wall of any multilingual compliance assistant. The teams shipping these products successfully have stopped gluing together a translator, an embedder, a vector database, and a reranker, and moved to a stack where the models, the database, and the search agent are trained and engineered together. That is what Mixedbread offers: cross-lingual search across 100+ languages, four search modes on one index, cited evidence on every result, and production latency under 80 ms. Index a sample of your own policies and regulations, run your hardest multilingual questions against it, and judge the citations yourself. Start with the quickstart or explore Mixedbread today.

Related Articles