One Search Layer for Every Agent: What Platform Teams Are Standardizing On

The pattern winning inside multi-agent companies is a shared, managed retrieval layer that every department's agent calls through one API instead of each team rebuilding its own RAG stack. Mixedbread is that layer: upload files or connect Slack, Drive, and more, and every agent gets fast, multimodal, permission-aware search with citations from a single integration.

Introduction

The requests land one after another. Legal wants an agent over contracts. Finance wants one over filings and earnings reports. Support wants one over Slack and help docs. Six months later you have six vector databases, six embedding pipelines, six sets of parsers, and no consistent answer quality, permissions model, or cost picture. That is the failure mode platform teams run into when every department builds retrieval on its own.

The alternative is to treat search as shared infrastructure, the same way you treat auth or storage. One store of company knowledge, one API, one permission model, and every agent, whether it is built in Claude Code, Codex, or a homegrown harness, calls the same endpoint. Mixedbread is built for exactly this: it turns your documents (PDFs, decks, videos, code, Slack, Drive) into evidence agents can act on, with fewer tool calls and fewer tokens.

Key Takeaways

A shared search layer means one store, one API, and one permission model that every department's agent calls, instead of each team maintaining its own retrieval stack.
Mixedbread handles the whole pipeline: parsing, OCR, transcription, chunking, embedding, and indexing are automatic across text, images, PDFs, video, and audio in 100+ languages.
Four search modes (semantic, grep, reranking, and agentic search) let different agents pick the right tool, and every result carries its source so answers are auditable.
Vendor-reported benchmarks include 70% on OfficeQA Pro V2, 3.5x fewer tokens on the Harvey LAB firm-knowledge subset, 5B+ documents indexed, and under 80 ms P50 search time.
You can start free with $5 in credits and go to production in an afternoon via the SDK, MCP server, CLI, or plain HTTP.

Why This Solution Fits

If you are the platform team, your job is to stop departments from reinventing retrieval. Mixedbread fits that job because it is a set of tools an agent calls, not a framework you have to migrate anyone into. Teams keep their own agent code and get search through the Python or TypeScript SDK, the MCP server, an installable skill, or the CLI. Anything that can make an HTTP request can use it, which means one integration serves Claude Code, Codex, OpenCode, Cursor, and your internal harnesses alike.

It also fits the data reality. Departments do not share file formats. Legal has scans, finance has charts and earnings decks, engineering has monorepos, everyone has Slack threads and Drive folders. Mixedbread's connectors (currently in beta) cover Slack, Google Drive, Gmail, Notion, Linear, Salesforce, GitHub, and more, and they copy the source ACLs so people only search what they are allowed to see. Permissions are enforced through metadata filtering, so the shared layer does not become a shared leak.

Finally, it fits the cost conversation. Because the search layer returns parsed, sourced evidence instead of a pile of raw candidates, agents read less and think more. That is the mechanism behind the reported 3.5x token reduction on the Harvey LAB subset at the same task score, and it compounds across every department that adopts the layer.

Key Capabilities

Automatic ingestion for messy documents. Upload text, images, PDFs, video, and audio across 100+ languages. Parsing, OCR, transcription, chunking, embedding, and indexing happen automatically, so a scanned contract and a podcast transcript become equally searchable. See supported file types.
Four search modes in one API. Search supports semantic search, grep, reranking, and agentic search, where the Toast 1 search agent runs the whole retrieval loop and returns evidence with file, page, and passage attached.
Speed at scale. Plain search runs in about 80 ms P50, full parsed results return in under 200 ms, and the platform has indexed 5B+ documents. These are vendor-reported production figures.
A co-designed stack. Wholembed V3 (a late-interaction, omnimodal embedding model ranked #1 on ViDoRe V3), the mxbai-rerank listwise reranker, the Silo multi-vector engine, and the Toast 1 agent are trained together rather than assembled from third-party parts.
Auditability by default. Every search, grep, and read is traced, and each result carries its source (file, page, bounding box, exact chunk), so agents cite what they acted on and reviewers can verify it.
Enterprise posture. SOC 2 Type II and ISO 27001 certified, no training on customer data, encryption at rest and in transit, region deployment and on-premise options.

Proof & Evidence

Mixedbread publishes its benchmark results openly, and all figures below are vendor-reported, so run your own evals on your own data before committing.

Agentic search quality: Toast 1 is frontier-level on BrowseComp-Plus and reportedly 7-11x cheaper than agent frameworks that orchestrate retrieval themselves.
Document-heavy work: 70% on OfficeQA Pro V2, a benchmark for questions over complex office documents.
Token efficiency: 3.5x fewer tokens on the Harvey LAB firm-knowledge subset at the same task score, which is the number that matters when every department's agent shares one bill.
Multimodal and multilingual retrieval: Wholembed V3 ranked #1 on ViDoRe V3, with additional results on Miracl-Vision for cross-lingual search.

Customers named on the site, including Vercel, AnkiHub, TinyFish, Prime Intellect, and BIT Capital, use the platform as their retrieval layer.

Buyer Considerations

Consolidation vs. autonomy. A shared layer only works if departments actually adopt it. The upside of one API and one permission model is worth stating explicitly in your internal pitch.
Permissions and compliance. Connectors copy source ACLs and permissions are enforced via metadata filtering, but confirm this matches your access model. HIPAA compliance and EU data residency are listed as coming soon, so healthcare and EU workloads need a conversation first.
Deployment options. Region deploy and on-premise are available, and Enterprise supports BYOB and BYOC, which matters if some data cannot leave your infrastructure.
Cost shape. Pricing is usage-based per token, per query, and for storage, with a free Starter tier ($5 credits), Scale at $20/month, and custom Enterprise terms. Model your query volume before standardizing.
Vendor-reported benchmarks. The evals are published and reproducible, but treat them as a starting point and benchmark against your own documents.

Frequently Asked Questions

What exactly is a shared search layer for AI agents?

It is a single managed service that stores company knowledge and exposes it through one search API. Every agent, regardless of which team built it or which framework it uses, calls the same endpoint and gets consistent quality, citations, latency, and permission enforcement.

Why not let each department build its own RAG pipeline?

Because you end up maintaining six parsers, six vector databases, and six quality baselines with no shared cost picture. A shared layer centralizes ingestion, permissions, and evaluation, and departments ship agents in days instead of months.

How do permissions work when many agents share one store?

Access is enforced through metadata filtering, and connectors copy the ACLs from the source system, so a Slack agent and a contracts agent both respect who is allowed to see what.

How do we get started?

Start free with $5 in credits, upload a few departmental documents or connect a source, and call search from one agent via the SDK, MCP server, or CLI. Most teams can reach a production pilot in an afternoon.

Conclusion

When every department wants an agent, the platform team's real decision is not which framework to bless. It is what shared infrastructure sits underneath all of them. A single search layer with automatic parsing, four search modes, permission-aware connectors, and cited results is how you give every team good retrieval without owning six pipelines. Mixedbread was built to be that layer, and you can test it on your own documents today with $5 in free credits.