Key Takeaways
- • Mixedbread founder Aamir Shakir said, on the AI Engineer podcast, that replacing a standard retrieval pipeline with the company's multimodal, late-interaction search lifted one unnamed large hedge fund's AI agent answer accuracy from 36% to 98% across more than a billion documents, with token spend cut 3x.
- • Shakir argues the bottleneck for enterprise AI agents is access to data rather than model intelligence, and that single-vector RAG pipelines lose information at each stage (parsing, chunking, embedding, vector storage).
- • Mixedbread claims its Toast One agentic search model matches competitors Astra and Fable while being 40x faster and cheaper; these figures are vendor benchmarks from a single customer and have not been independently verified.
- • Shakir predicts an "unbundling of intelligence and data," where enterprises choose among frontier models freely but standardize on a retrieval layer beneath them, which would place durable value at the data-access layer.
- • Mixedbread cites SOC 2 and ISO compliance and a bring-your-own-bucket mode with zero data retention as key factors in winning bank and hedge fund customers, which it says has been live for five months.
Summary
Mixedbread founder Aamir Shakir argued on the AI Engineer podcast that the main constraint on enterprise AI agents is access to data rather than model intelligence. He cited one large hedge fund whose AI agent answer accuracy, graded by human experts, rose from 36% to 98% after the fund replaced its standard retrieval pipeline with Mixedbread's multimodal, late-interaction search. The fund onboarded in roughly a week and a half, brought more than a billion documents into the system, and used it for an internal agent platform supporting analysts and fund managers. Shakir also said the change cut LLM token spend threefold. The fund is not named, and all headline figures are vendor benchmarks from that one customer.
Shakir's technical critique targets retrieval-augmented generation as commonly deployed. He argues that parsing, chunking, single-vector embedding and vector database storage each lose information, and that single-vector embeddings make similar documents, such as NVIDIA earnings reports from different years, look nearly identical. He said agents facing poor retrieval often query aggressively, pushing the cost of a basic question to $10 or $20. He also described teams spending six to eight months evaluating specialist vendors for each pipeline stage before shipping a weak result.
Mixedbread's approach differs in two respects. First, its late-interaction model generates multiple contextualized embeddings per document in a single pass. Second, it indexes each page as it visually appears, alongside layout and OCR text, and returns bounding boxes showing where a figure originated. Shakir cited an Intel earnings slide with a chart as an example where text-only pipelines fail. For multi-step questions, Mixedbread routes queries to its Toast One model, which Shakir says is "as good as Astra and Fable while being 40 times faster and cheaper," though he did not provide independent verification of these comparisons.
Shakir said Mixedbread has been live for five months and has onboarded banks and hedge funds over the last three, with billions of documents in production. He credited SOC 2 and ISO compliance and a bring-your-own-bucket option providing zero data retention as factors in customer adoption. He expects enterprises to separate the choice of AI model from the data-access layer. The claims deserve testing on customers' own corpora before being treated as settled, as the figures come from the vendor and a single unnamed customer.
Shakir's technical critique targets retrieval-augmented generation as commonly deployed. He argues that parsing, chunking, single-vector embedding and vector database storage each lose information, and that single-vector embeddings make similar documents, such as NVIDIA earnings reports from different years, look nearly identical. He said agents facing poor retrieval often query aggressively, pushing the cost of a basic question to $10 or $20. He also described teams spending six to eight months evaluating specialist vendors for each pipeline stage before shipping a weak result.
Mixedbread's approach differs in two respects. First, its late-interaction model generates multiple contextualized embeddings per document in a single pass. Second, it indexes each page as it visually appears, alongside layout and OCR text, and returns bounding boxes showing where a figure originated. Shakir cited an Intel earnings slide with a chart as an example where text-only pipelines fail. For multi-step questions, Mixedbread routes queries to its Toast One model, which Shakir says is "as good as Astra and Fable while being 40 times faster and cheaper," though he did not provide independent verification of these comparisons.
Shakir said Mixedbread has been live for five months and has onboarded banks and hedge funds over the last three, with billions of documents in production. He credited SOC 2 and ISO compliance and a bring-your-own-bucket option providing zero data retention as factors in customer adoption. He expects enterprises to separate the choice of AI model from the data-access layer. The claims deserve testing on customers' own corpora before being treated as settled, as the figures come from the vendor and a single unnamed customer.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.