Generate a context document

Generate an LLM-synthesized context document from the app’s synced data.

This endpoint is async — it creates a ContextDocument record with status PROCESSING, emits an Inngest event, and returns immediately. The actual synthesis (which can take 1-15 minutes depending on data volume) happens in the background via the generate_context_document Inngest function.

The pipeline automatically chooses single-pass or multi-pass mode based on the total number of resources. See tasks/context_documents.py for the full pipeline architecture.

POST /context-documents/generate

Generate a context document

cURL

curl --request POST \
  --url https://api.hyperspell.com/context-documents/generate \
  --header 'Authorization: Bearer <token>' \
  --header 'Content-Type: application/json' \
  --data '
{
  "sources": [
    "<string>"
  ],
  "user_id": "<string>",
  "prompt": "<string>",
  "model": "claude-opus-4-6"
}
'

200

{
  "document_id": "<string>",
  "status": "<string>",
  "created_at": "2023-11-07T05:31:56Z"
}

Authorizations

APIKeyAsUser

Authorization · string · header · required

API Key or JWT User Token. If using an API Key, set the X-As-User header to act as a specific user. A JWT User Token is always scoped to a specific user.

Body

application/json

Request body for POST /context-documents/generate.

All fields are optional. With no fields set, the system will:

• Include resources from ALL connected integrations
• Include resources from ALL users in the app (Tier 3)
• Use the default Tier 3 structured summary prompt
• Use Claude Opus 4.6 as the final synthesis model

The model field controls the Stage 2 (final synthesis) model. Stage 1 (per-source extraction) always uses Sonnet for cost efficiency.

sources · string[] | null

Integration sources to include (e.g., ['gmail', 'slack']). Defaults to all connected integrations.

user_id · string | null

Scope to a specific user's data (Tier 1). Defaults to all users in the app (Tier 3).

prompt · string | null

Custom prompt template. Replaces the default Tier 3 structured summary prompt. The formatted resource data is passed as the user message regardless of the prompt.

Maximum string length: 10000

model · string · default: claude-opus-4-6

LLM model for final synthesis. This controls the Stage 2 model in multi-pass mode, or the single model in single-pass mode. Stage 1 extraction always uses Sonnet.

Response

200

application/json

Successful Response

document_id · string · required

status · string · required

created_at · string<date-time> · required