
Agents only retrieve what their owner may read
Argent runs OpenFGA to decide, document by document, what a retrieval may hand an AI agent. The agent inherits its human owner's access, nothing more, and anything the engine cannot answer comes back empty.
What Argent does with OpenFGA
Vector search finds what is relevant. OpenFGA decides what this person is allowed to see. Only that reaches the model.
OWASP's Top 10 for LLM applications names the risk directly: "Inadequate or misaligned access controls can lead to unauthorized access to embeddings containing sensitive information." Its first mitigation: "Implement fine-grained access controls and permission-aware vector and embedding stores" (OWASP LLM08:2025).
OpenFGA is a CNCF project that moved to the Incubating maturity level on October 28, 2025 (CNCF).
- The owner's access, never more. An agent's relationships are those of its human owner. An agent with no owner, or one Argent cannot map to exactly one person, gets nothing.
- Inheritance that matches your folders. Owners, viewers, groups inside groups, folders inside folders and collections: share a folder with a group and every member can retrieve what is under it, with no per-document grant.
- Fail closed. An engine error returns an empty set with
fail_closed: true, never the unfiltered candidates. Errors are never cached as an allow. - Gate at the gateway too. When the tenant's RAG authorization module is on, the gateway also asks OpenFGA whether the agent's owner may
invokethe upstream endpoint the agent is calling, and denies on any doubt.
How it connects
OpenFGA ships with Argent and runs next to it. Your pipeline asks Argent, and Argent asks OpenFGA.
| Engine | Self-hosted OpenFGA with its own Postgres database, in the Compose overlay and the Kubernetes chart; its playground is off, for air-gapped sites |
| Model | Provisioned by Argent at start-up: user, group, folder, collection, document |
| Filter | POST /v1/rag-authz/filter-retrieval: candidate ids in, the authorized subset out |
| Point check | POST /v1/rag-authz/check: one allow or deny |
| Grants | POST and DELETE /v1/rag-authz/tuples, written by your ingestion pipeline |
| SDKs | Go FilterRetrieval, Node filterRetrieval, Python filter_retrieval |
| Tenant | Taken from the credential only, never from the request body |
What you configure
Three changes to a RAG pipeline you already run, plus one option.
- Switch it on. Turn on RAG authorization in the install and set the SDK credential your pipeline will present.
- Mirror your sharing. When your pipeline indexes a document, it writes who owns or can view it, and which folder or collection it sits in. When it removes a document, it deletes those grants. Writes fail loudly, so a missed grant is seen.
- Filter before the prompt. Between vector search and building the prompt, call
filter_retrievalwith the candidate ids and send onlyallowedto the model. If the answer is empty, the context is empty. - Optional: gate the endpoint. Turn on the tenant's RAG authorization module and grant each person the upstream endpoints they may call.
A lower admin tier cannot overwrite or remove a grant a higher tier wrote, and each refusal is signed.
Proof of what was withheld
The leak that did not happen is evidence too.
- Retrieval log. For each retrieval: who asked, which documents reached the model, which were withheld, and any fail-closed event, with names, not ids.
- Console. The RAG Authorization page under Guardrails lists retrievals and shows each one's withheld documents and fail-closed warnings.
- Who can read it. SuperAdmin, TenantAdmin and Auditor. Grants are written by administrators and the pipeline only.
- Metrics. Decision and fail-closed counts are exported as metrics, so your monitoring can alert on a rise in fail-closed answers.
Where this fits
Pages that go deeper.
- Data loss prevention: retrieval filtering is one layer of data loss prevention; Sensitive Data Redaction covers what is typed into prompts, and rolling out copilots and chat safely puts both to work.
- Policy for requests: everything else an agent sends is decided by the built-in OPA engine or by your own.
- Capabilities: Model Control, Agent Permissions, Watch Before You Enforce and all capabilities.
- Goals: model control and routing, agent identity and ownership, audit and accountability and all goals.
- Use cases: controlling agent tool access, securing AI coding agents and AI in regulated and air-gapped environments.
- Where it runs: air-gapped on-premises, private cloud or your public cloud VPC. Compare them on deployment options.
- Industries: Financial Services, Insurance, Healthcare, Public Sector and all industries.
- More integrations: every system Argent connects to, grouped by type, on the integrations page.
Want this running on your own use cases?
All downloads