Mixedbread Search
by Mixedbread (Berlin, founded 2023)
A search API built for the AI era: upload any file format to a “Store,” and Mixedbread automatically parses, understands, and indexes it — text, images, tables, and complex layouts across 100+ languages — so an LLM or agent can retrieve grounded context without a separate vector-DB pipeline. Public beta since 2025-10-01.
See https://www.mixedbread.com/
Stores — the core abstraction
A Store is a search index: create one, upload PDFs, images, Word/PowerPoint/Excel
files, code, or video, and Mixedbread’s infrastructure automatically understands the
content without manual preprocessing or document-parsing setup. Queries run as natural
language, not keyword matching, returning precisely ranked results optimized for
downstream AI consumption rather than for humans.
Agentic Search
Wraps the plain search endpoint in an LLM-driven loop: the agent runs parallel
sub-queries, analyzes metadata facets, inspects retrieved results, and decides whether to
search again before returning a final ranked chunk list — closer to a research assistant
than a single retrieval call.
Integrations
- API for direct programmatic queries
- MCP Server — lets Claude Desktop and other MCP-aware AI assistants search Stores
directly (see MCP Resources for the underlying protocol pattern) - CLI for bulk operations and CI/CD pipelines
- LangChain integration
- Agent Skills packages to reduce setup overhead
Performance claims
On BrowseComp-Plus benchmarks with Gemini 2.5 Flash: 38.19% answer accuracy — reported as
16% higher than competing systems — while using 8.26 LLM calls per question, 16% fewer
than the next-best alternative. On DeepResearch-style benchmarks, Mixedbread reports LLM
assistants reaching meaningfully better response accuracy versus existing search systems.
Embedding models
Also publishes open-source embedding/reranking models on Hugging Face, notably
mxbai-embed-large-v1 — the models underpinning Stores’ retrieval quality are part of
the same open research effort, not a fully closed black box.
Company
Founded 2023 in Berlin by Aamir Shakir and a co-founder, originating from building an
on-premise search engine for consultants and lawyers. Raised 900K, Jan 2024) and seed ($5.5M) rounds, most recently closing Feb 2025.
Related
- Gemini File Search — closest comparison: Google’s managed
multimodal RAG store inside the Gemini API; Mixedbread is provider-agnostic (MCP/API/
LangChain) rather than locked to one model family - MCP Resources — the protocol pattern Mixedbread’s MCP Server exposes
Stores through