Extraction endpoints

POST /v1/typed

Fill a schema from text (context), a conversation (messages) or an image (image, base64).

Request body

Parameter	Type	Default	Description
schema	object	required	Field name → field spec.
context	string	""	The document or text to read. Required unless messages or image is given.
messages	array	null	A conversation in OpenAI format: [{"role": "user", "content": "..."}], roles user, assistant, system. See Conversations. Not combinable with image.
image	string	null	Base64 image, with or without a data: prefix. See Images.
system	string	null	Instructions shared by all fields. Cannot be combined with a system message in messages.
max_new_tokens	integer	32	Answer budget per field, 1–512. Raise it for long text fields (150–300 for replies and summaries).
review_threshold	number	0.5	Fields with confidence below this are listed in needs_review (0–1).
secrets	object	null	Strings for {{secret.name}} placeholders in tool calls. Never logged or returned.
model	string	null	Model to use instead of the deployment's default. Loading another model takes time and memory; leave empty unless told otherwise.
thinking	boolean	false	Let the model reason freely before each answer. Only for Thinking models; others return HTTP 400.
thinking_budget	integer	128	Maximum reasoning tokens per field when thinking is on (1–4096).
dtype	string	"bfloat16"	Numeric precision for some model types; ignored otherwise.

Example

curl https://api.trusttyped.com/v1/typed \
  -H "Content-Type: application/json" \
  -d '{
    "system": "Copy values exactly. Never guess.",
    "context": "Invoice INV-0932. PT Sinar Jaya. Total due 3.885.000. Status: UNPAID.",
    "schema": {
      "invoice_no": {"type": "string", "extract": true},
      "vendor":     {"type": "string", "extract": true},
      "total":      {"type": "number", "integer": true, "max_digits": 10},
      "paid":       {"type": "bool"}
    },
    "review_threshold": 0.7
  }'

Response

A response object, with "mode": "text" or "mode": "image".

Streaming

POST /v1/typed/stream takes the same body and returns NDJSON (application/x-ndjson): progress events, then one final event with "done": true. See Streaming for the event format and client code.

Text-only endpoints

POST /v1/generate and POST /v1/generate/stream accept the same body without image, and their responses do not contain mode. Prefer /v1/typed for new integrations.