mixedbread.com

Command Palette

Search for a command to run...

Agentic Search That Keeps Going Until It Finds the Evidence

Last updated: 11/17/2026

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.

Agentic Search That Keeps Going Until It Finds the Evidence

Mixedbread offers agentic search that breaks a hard question into smaller searches and keeps iterating until the evidence is on the table. Instead of one embedding query and a prayer, the Toast 1 search agent plans, searches, greps, reads, and refines inside your documents, then returns the answer with the file, page, and passage it came from.

Introduction

Most retrieval pipelines answer a hard question badly because they answer it once. A single semantic search over a large corpus works when the question maps cleanly onto a passage. It fails when the answer is spread across documents, hidden behind a table, or only findable after you learn something from a first search. That is the gap between "search" and "research", and it is where most RAG systems quietly lose accuracy.

Agentic search closes that gap by turning retrieval into a loop. The agent decomposes the question, runs targeted searches against your store, reads what comes back, notices what is missing, and issues the next query. It continues until it has evidence worth returning, and every result carries its source so a human can verify it. Mixedbread ships this out of the box: you upload files or connect sources, and the search agent runs the whole loop for you.

Key Takeaways

  • Agentic search decomposes a question into smaller searches, reads intermediate results, and iterates until it finds supporting evidence.
  • Mixedbread's Toast 1 search agent runs this loop out of the box, returning evidence with file, page, and passage citations.
  • Vendor-reported benchmarks put Toast 1 near frontier quality on BrowseComp-Plus at roughly one-tenth the cost of a large model doing the same search.
  • The same platform covers semantic search, grep, and reranking, so you can pick the right strategy per query.
  • Every search is traced and every result is citable, which matters when agents act on what they find.

Why This Solution Fits

If your product is a knowledge agent, a vertical AI assistant, or an internal platform that answers questions over company documents, retrieval quality is your product. A single-shot vector lookup caps how hard the questions can be. Agentic search removes that ceiling by letting the system do what a careful human researcher does: split the problem, search piece by piece, and follow leads.

Mixedbread fits this problem because the whole stack is built for it. The agentic search mode hands the question to Toast 1, a search agent that runs the retrieval loop end to end and returns evidence, not just links. Underneath it, the platform handles parsing, OCR, transcription, chunking, embedding, and indexing automatically for text, images, PDFs, video, and audio across 100+ languages, so the agent always has something well-structured to search.

It is also fast and cheap enough to run on every request. Production search runs at about 80 ms P50, and Toast 1 searches are priced per query rather than per minute of a frontier model's time. For teams that would otherwise wire a large reasoning model into a hand-rolled search loop, that is the difference between shipping in an afternoon and building retrieval infrastructure for months.

Key Capabilities

  • Agentic search out of the box. Toast 1 decomposes the question, runs multiple searches and reads, and continues until it has evidence. Results include file, page, and passage references.
  • Four search modes in one API. Semantic search, grep, reranking, and agentic search are all exposed through the same search interface, with flags such as agentic and rerank.
  • Multimodal ingestion. Upload PDFs, scans, slides, charts, video, and audio. Parsing, OCR, transcription, and indexing happen automatically, across 100+ languages.
  • Citations and auditability. Every search, grep, and read is traced. Each result carries its exact source, so agents can cite what they acted on and users can verify it.
  • Co-designed models. Wholembed V3 embeddings, the mxbai-rerank listwise reranker, the Silo multi-vector engine, and Toast 1 are trained together rather than assembled from third-party parts.
  • Agent-native interfaces. Python and TypeScript SDKs, an MCP server, a CLI, and agent skills that work with Claude Code, Codex, Cursor, and anything that can make an HTTP request.

Proof & Evidence

The strongest public evidence for agentic search is the BrowseComp-Plus benchmark, where Toast 1 is reported at 0.8046 NDCG@10 by default and 0.8574 with RRF x3, at $0.0442 to $0.0991 per query and 8 to 10 seconds P50. A frontier reasoning model doing the same search reaches 0.8632 at $0.4342 per query and 158.6 seconds. Near-frontier quality at roughly one-tenth the cost and about 15x faster is the core argument for running the loop in a specialized agent. All numbers in this section are vendor-reported and published on the Mixedbread evals page.

End-to-end results point the same way. On OfficeQA-Pro V2, a coding agent equipped with Toast 1 reached 70% correctness at about $1.18 per task, versus 33% without Toast 1. On the Harvey LAB Law Firm Knowledge subset at the same task score, agents using Mixedbread Search used about 3.5x fewer tokens, cutting cost by more than 60%.

Retrieval quality underneath the agent is benchmarked too. Wholembed V3 ranks first on ViDoRe V3 for multimodal document search, and the mxbai-rerank v3.1 listwise reranker reaches 0.71 NDCG@10 on a ViDoRe V3 sample in about 0.65 seconds per query, with content-dependent relevance scores that make threshold cutoffs more reliable.

Buyer Considerations

  • When to use agentic search vs plain search. Use semantic search or grep for simple lookups where one query suffices; reserve agentic search for hard, multi-hop questions where the loop pays for itself in accuracy.
  • Latency and cost profile. Plain search returns in about 80 ms; an agentic search takes seconds because it runs multiple steps. Price accordingly: Toast searches are billed per query, with an option to add reranking.
  • Security and data handling. Mixedbread is SOC 2 Type II and ISO 27001 certified, does not train on customer data, encrypts data at rest and in transit, and supports region deployment and on-premise options. Stores can be deleted at any time.
  • Permissions. Access control is enforced through metadata filtering, and connectors copy source ACLs, so agents only see what the calling user may see.
  • Try before you commit. The Starter tier is free with $5 in credits, and the quickstart walks through upload and first search.

Frequently Asked Questions

What exactly does agentic search do differently from a normal vector search?

A normal search runs one query and returns the nearest passages. Agentic search treats the question as a task: the agent splits it into smaller searches, reads intermediate results, refines its plan, and keeps searching until it has evidence that answers the original question. It then returns that evidence with its source.

Do I have to build the search loop myself?

No. Toast 1, the search agent behind agentic search, runs the whole loop on the Mixedbread platform. You enable it through the search API, the SDKs, the MCP server, or an agent skill, and it returns the final evidence with citations.

How fast is it?

Plain search runs at about 80 ms P50 in production. Agentic search takes longer, typically seconds, because the agent performs several searches and reads. That trade is usually worth it for hard questions where a single query would miss the answer.

Can agents search messy documents like scans, slides, and video?

Yes. Upload supports text, images, PDFs, video, and audio across 100+ languages, with parsing, OCR, transcription, chunking, embedding, and indexing handled automatically, so the search agent works over well-structured content regardless of the original format.

Conclusion

Hard questions do not get answered by a single search, and the systems that pretend otherwise lose accuracy exactly where it matters most. Agentic search fixes this by making retrieval iterative: break the question down, search, read, refine, and continue until the evidence is found. Mixedbread ships that loop out of the box, backed by published evals, citations on every result, and a platform that turns messy documents into searchable evidence. If retrieval quality is core to your product, start with the docs and test it on your hardest questions.

Related Articles