> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ai-baseline.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Platform REST API preview

> Prelaunch organization-scoped collection queries with hosted identity or personal access tokens.

> [!WARNING]
>
> This is a prelaunch source contract for Platform API 0.3.0. It has local qualification,
> not a public deployment. There is no new public base URL yet. Your existing collection
> URLs and supported [REST v2 contract](https://docs.ai-baseline.com/guides/api-versioning.md) remain unchanged.

The new platform lets a customer discover their organization's collections and query
them through one API base URL. It returns evidence for your application to use,
not a generated answer. The authenticated collection UUID selects the collection;
you never provide a database, knowledge-domain deployment or backend host.

Use a hosted customer API access token or a personal access token (PAT) as exactly
one `Authorization: Bearer ...` header. Auth0 Management API credentials and identity
ID tokens are not customer API access tokens. PATs follow their chosen expiry and
current user access; removing membership, disabling access or revoking a PAT blocks
subsequent queries. A request already admitted may finish.

1. Discover your organizations with `GET /v1/organizations`.
2. List available collections with `GET /v1/organizations/{org}/collections`.
3. Send `POST /v1/organizations/{org}/collections/{collection}/query` with JSON:

```json theme={null}
{
  "query": "What evidence supports this claim?",
  "mode": "research",
  "effort": "medium",
  "retrieval": {"max_evidence": 20, "max_evidence_tokens": 16000}
}
```

All active organization members can query its ready collections. An inaccessible
organization or collection is not revealed by supplying its UUID. Frozen/archived
organizations cannot query content. Discovery is independent of any future personal
MCP tool selection.

The JSON result contains `request_id`, XML `evidence`, optional `evidence_instructions`,
`summary` and optional `retrieval_plan`. Effective metadata filters appear in the
summary and in individual retrieval-plan queries when provided by retrieval. A missing
filter field means unavailable; it does not establish that no filters were applied.
This slice does not return citation links, image delivery or an optional dashboard.

Query text is nonempty and at most 4,000 characters. Modes are `basic` (default),
`research` and `agentic_research`; effort is `low`, `medium` (default) or `high`.
Optional `metadata_filters` follow the existing retrieval filter syntax and require
research/agentic research. Requested evidence limits allow 1–100 entries and
1–64,000 tokens, defaulting to 20 and 16,000. Unknown controls and URL query parameters
are rejected; JSON requests have a 64 KiB limit.

The initial allowance is **10 admitted collection queries per fixed UTC minute per
organization**, shared across its users, collections and credentials. Adjacent-minute
bursts are possible. HTTP `429` includes `Retry-After`; discovery, management and upload
do not consume this allowance. Admitted attempts consume it even when retrieval fails.
This is a capacity limit, not a credit charge. No automatic retry is performed.
Malformed native query/filter input returns a bounded 400/413/422 error; unavailable
retrieval returns a sanitized 503 without private diagnostics. Responses are not cached.

[Preview OpenAPI](https://docs.ai-baseline.com/api-reference/platform-preview.openapi.json) describes this separate
prelaunch service. The current public `openapi.json` continues to describe the existing
supported API. Billing admission, hosted query rollout and shared MCP are not complete.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.