MatrixArk Contact

Context API reference

One boundary for LLM context.

The MatrixArk context API is a single request-time surface. Send raw events and lightweight hints; MatrixArk extracts, compiles, and stores context, then returns a token-budgeted ContextPack you can drop straight into a prompt — and replay later by id. You never model hash keys, indexes, or storage routing yourself.

Talk to MatrixArk How TemporalStore serves it

The API has three verbs that mirror the context lifecycle: ingest an event, build a context pack for a prompt, and replay a pack by id. Extraction, freshness, permission checks, and storage routing happen behind the boundary.

1. Ingest an event 
POST /v1/context/events

{
  "tenant": "company_a",
  "source": "cursor",
  "scope_hint": { "team": "platform", "project": "project_1" },
  "kind": "tool_result",
  "content": "Alice approved another GPU batch up to $80k until June 30.",
  "source_ref": "s3://company-a/tools/purchase-approval-913.json",
  "observed_at": "2026-06-14T16:20:00Z"
}

# 202 Accepted
{ "event_id": "evt_5f31a9", "status": "extracting" }
2. Build a token-budgeted ContextPack 
POST /v1/context/packs

{
  "tenant": "company_a",
  "raw_query": "Can we buy another GPU batch this week?",
  "hints": { "team": "platform", "project": "project_1", "max_prompt_tokens": 1200 },
  "as_of": "2026-06-14T18:00:00Z"
}

# 200 OK -> a replayable manifest, not a raw blob
{
  "context_pack_id": "pack_7731c0",
  "as_of": "2026-06-14T18:00:00Z",
  "sections": [
    { "id": "sec_latest_approval", "text": "GPU batch approved to $80k, valid until 2026-06-30.",
      "source_ref": "s3://company-a/tools/purchase-approval-913.json", "valid": true },
    { "id": "sec_open_spend", "text": "Committed spend this quarter: $61k of $80k." }
  ],
  "blocked": [
    { "id": "evt_2201", "reason": "stale: superseded by approval-913" }
  ],
  "token_estimate": 940,
  "replay_id": "pack_7731c0"
}
3. Replay a pack by id 
GET /v1/context/packs/pack_7731c0

# 200 OK -> the exact context the model saw, plus why each item was chosen
{
  "context_pack_id": "pack_7731c0",
  "as_of": "2026-06-14T18:00:00Z",
  "included": ["sec_latest_approval", "sec_open_spend"],
  "blocked": [{ "id": "evt_2201", "reason": "stale: superseded by approval-913" }],
  "query": "Can we buy another GPU batch this week?",
  "token_estimate": 940
}

Request and response shape

Simple inputs at the edge; bounded, auditable context on the way out.

Requests stay forgiving

Every call carries a tenant and either an event to ingest or a raw query to answer. Scope hints (team, project, matter, ticket) are optional; MatrixArk extracts scope, entities, time, validity, and permissions and compiles them into TemporalStore records. You do not send hash keys, index names, or timestamp layouts.

Responses stay bounded

A ContextPack is a compact manifest: included section ids and text, blocked records with reasons, source refs, a token estimate against your max_prompt_tokens budget, and a replay_id. Nothing is an unbounded scan — time windows are validated and result counts are capped so serving stays low-latency.

Related

Serving core TemporalStore The time-aware engine that stores and serves what this API compiles. Enterprise MatrixKV Transactional metadata — permissions and approvals the pack references. Enterprise MatrixDB Hot context state and cache metadata behind the serving path.