Awesome Search KG

Mixedbread

2 min read

Mixedbread

Mixedbread is an AI infrastructure company whose platform ingests mixed document formats (PDFs, video, code, and connected sources such as Slack and Drive) and handles parsing, chunking, embedding, and indexing so that agents can retrieve context from them. It is best known in the retrieval community for its open-weight mxbai-* model line, whose reranker models are widely used as drop-in cross-encoders.

Website: https://mixedbread.ai

Products

From the company’s own site:

  • Wholembed V3 — its flagship embedding model, described as late-interaction, omnimodal, and covering 100+ languages (see Late Interaction, Multimodal Embeddings).
  • mxbai-rerank — the reranker line, in a listwise formulation (Listwise Relevance Evaluation).
  • Toast 1 — an agentic search agent that runs the retrieval loop itself (Agentic Search).
  • Silo — an S3-native multi-vector database engine.

Deployment options include regional and on-premise installations.

In This Vault

  • Hev meets Jev — mxbai-rerank-large-v2 (hosted) is one of the three purpose-built rerankers benchmarked, scoring 0.476 mean nDCG@10 across SciFact, NFCorpus, and FiQA at $3.50 per 1,000 queries — the most expensive and the lowest-scoring of the hosted rerankers in that particular run, though comfortably above the unreranked BM25 baseline of 0.404. Note that the run predates the current listwise mxbai-rerank generation.

Graph View

Backlinks