Endpoint reference

POST/ingest

Push one unit of content. Runs the full pipeline and queues per tenant, so content applies in the order it is sent.

{
  "teamId": "wZ4mKq",
  "ingestionId": "b41c9e2a7f6d4c118a3e",
  "sourceId": "call-8821",
  "status": "queued"
}

GET/ingest/{ingestionId}

Poll until the queue has applied the content and finished its deferred maintenance. Pass teamId as a query parameter if the token reaches more than one project.

{
  "teamId": "wZ4mKq",
  "ingestionId": "b41c9e2a7f6d4c118a3e",
  "status": "completed",
  "settled": true,
  "blocked": false,
  "readiness": {
    "raw": 1.0,
    "comprehension": 1.0,
    "consolidation": 1.0,
    "parked": 0
  }
}

• Gate on settled, not on status. settled is true only when every stage has drained with nothing parked. status can read "completed" while distillation is still outstanding, and reading then returns a half-written graph.
• blocked: true means every remaining row parked. Stop polling and inspect the feed rather than waiting out the timeout.
• readiness is a live progress vector, enough to show a real progress bar instead of a spinner.
• The queue is FIFO, so the last ingestion id of a feed covers the whole feed. There is no need to poll every id.

PUT/ingest/group

Replace the entire contents of a group in one call. Every point currently in the group is superseded by the items you send, so this is the call for a source that is re-synced rather than appended to: a document that changed, a thread re-fetched in full, a nightly export of a record.

An item takes the same fields as a single /ingest, minus group (the route already names it) and minus append (a replace is the whole content by definition): content (required), label, id, timestamp and author.

{
  "group": "acme-pilot",
  "items": [
    { "content": "Nadia confirmed the pilot ships March 14. Budget is 32k.",
      "label": "Account review call",
      "timestamp": "2026-02-03T09:00:00Z",
      "author": "Nadia Haddad" },
    { "content": "Budget for the Acme pilot moved to 40k.",
      "label": "Email from Nadia",
      "timestamp": "2026-07-28T16:00:00Z",
      "author": "Nadia Haddad" }
  ]
}

Accepted and queued exactly like POST /ingest, and polled the same way through GET /ingest/{ingestionId}.

POST/recall

Returns ranked evidence with its sources and no generated text, for running a separate model over the context.

{
  "teamId": "wZ4mKq",
  "results": [ /* typed matches, each with its source and score */ ],
  "evidence": [
    "Nadia confirmed the pilot ships March 14 (Account review call, 2026-02-03)",
    "Budget for the Acme pilot moved to 40k (Email from Nadia, 2026-07-28)"
  ],
  "truncated": false,
  "timelineIsVerbatim": false,
  "intent": "Lookup",
  "degradedArms": [],
  "usage": {
    "recallTokensIn": 812,
    "recallTokensOut": 96,
    "recallCostUsd": 0.000214
  }
}

• degradedArms names any retrieval arm that stayed degraded after its retry, which tells a partial search apart from a complete one that found little. An empty array means every arm ran.
• truncated means the token budget cut the result set. Both results and evidence are cut identically.

POST/answer

Recall plus a grounded answer. It shares most of /recall's fields but NOT limit or maxTokens: the answerer sizes its own context, which is what budget names. Sending either is ignored rather than honoured.

{
  "teamId": "wZ4mKq",
  "answer": "March 14, 2026. Nadia confirmed it on the February 3 account review call.",
  "evidence": [
    "Nadia confirmed the pilot ships March 14 (Account review call, 2026-02-03)"
  ],
  "abstained": false,
  "answerer": "gemini-3.5-flash",
  "intent": "Lookup",
  "recordKept": 12,
  "recordTotal": 12,
  "degradedArms": [],
  "usage": {
    "answerTokensIn": 2140,
    "answerTokensOut": 38,
    "answerCachedTokensIn": 1792,
    "answerCostUsd": 0.000431,
    "recallTokensIn": 812,
    "recallTokensOut": 96,
    "recallCostUsd": 0.000214
  }
}

• abstained: true means the memory did not hold the answer, so the model declined to generate one. Handle it as a normal outcome.
• recordKept and recordTotal separate "the record never loaded" from "the record loaded and the answer still missed": a retrieval problem versus a reasoning one.
• In agent mode the response carries toolCalls instead of evidence, listing the retrieval steps the loop took.

Entities, relationships and hops

POST/configurations