Generate editable sample JSON from a free-text request for schema authoring. Without attachments, use model knowledge and optional web search; with attachments, extract from the sources only (search does not relax that rule). Each sample is one instance in one language; set sample_count separately from request. Attachments force one sample. Requires editor; generation is billed. Returns a job_id and may already be paused or complete: relay pause questions through answer_job_question, otherwise poll get_job_status. Review returned samples and warnings before create_schema_from_sample; do not silently change facts or structure. For relationship modeling, multiple documents or hybrid extraction plus research, read enricher://docs/schema-from-sample and enricher://docs/documents.
View input schema
{
"type": "object",
"properties": {
"request": {
"default": "",
"description": "Entity type, desired fields, scope and size/depth budget. Required without attachments; optional source-mode instructions otherwise. Put the number of instances in sample_count, not in this text.",
"maxLength": 4000,
"title": "Request",
"type": "string"
},
"model": {
"default": "auto",
"description": "Model composite key; leave it on auto (the default). auto is YOUR model, or with use_platform_models=true the organization's best model for this task from its benchmark scores (quality/speed/cost weights), within the run's constraints (attachments, web search, thinking) — the result reports the pick as model_selection. Name a provider::model only when the user named one; that runs the job on the platform (billed). Keys: list_models. An explicit model bypasses the capability matching; provider combinations or quota may still fail.",
"title": "Model",
"type": "string"
},
"typical_objects": {
"anyOf": [
{
"items": {
"type": "string"
},
"maxItems": 20,
"type": "array"
},
{
"type": "null"
}
],
"default": null,
"description": "Up to sample_count concrete instances to anchor knowledge mode (e.g. ['Sanofi', 'Pfizer']), one per generated sample in order — slots beyond len(typical_objects) are named by the model's own instance roster. In source mode the attachment remains authoritative and this is ignored.",
"title": "Typical Objects"
},
"sample_count": {
"default": 1,
"description": "Number of same-type instances (1..20), default 1. Consider 3 varied instances when designing a schema. Attachments force 1. Inspect samples_note for under-delivery or the cap.",
"maximum": 20,
"minimum": 1,
"title": "Sample Count",
"type": "integer"
},
"naming_convention": {
"default": "auto",
"description": "auto | snake_case | camelCase",
"title": "Naming Convention",
"type": "string"
},
"attachment_ids": {
"anyOf": [
{
"items": {
"type": "string"
},
"type": "array"
},
{
"type": "null"
}
],
"default": null,
"description": "UUIDs from upload_attachment. Providing any attachment switches the call into source mode: transcribe the document or describe visible photo attributes only, with an interactive planner.",
"title": "Attachment Ids"
},
"auto_answer": {
"anyOf": [
{
"type": "boolean"
},
{
"type": "null"
}
],
"default": null,
"description": "Omit or false to pause for clarification; true authorizes standard interpretations and planner defaults without asking, in either mode.",
"title": "Auto Answer"
},
"enable_web_search": {
"anyOf": [
{
"type": "boolean"
},
{
"type": "null"
}
],
"default": null,
"description": "Knowledge mode only; source mode (attachments) remains source-only. On your model (the default), true asks you to ground the samples with your own web search tool and false not to; omitted returns web_search_undecided so you ask the user first. On platform models, true uses the provider's builtin search (omitted = false; ignored on an unsupported explicit model). Hybrid tasks: enricher://docs/documents.",
"title": "Enable Web Search"
},
"language": {
"anyOf": [
{
"type": "string"
},
{
"type": "null"
}
],
"default": null,
"description": "Output language code for the generated field names AND values (e.g. 'en', 'fr'); an explicit code applies even when attachments are in another language. Omitted (default) → the generator follows the language the request is written in (its text and any typical object), else the attachment's, else English.",
"title": "Language"
},
"existing_samples": {
"anyOf": [
{
"items": {
"additionalProperties": true,
"type": "object"
},
"maxItems": 19,
"minItems": 1,
"type": "array"
},
{
"type": "null"
}
],
"default": null,
"description": "Samples you already hold, to extend instead of starting over: the new samples reproduce the first one's exact field set and never repeat an instance already in the set; sample_count is then the number of NEW samples (set capped at 20 in total), request is optional context, and only the new samples are returned. Knowledge mode only (no attachments).",
"title": "Existing Samples"
},
"wait_seconds": {
"default": 120,
"description": "How long to wait for the first pause or completion before returning (0 = return the job_id immediately).",
"maximum": 600,
"minimum": 0,
"title": "Wait Seconds",
"type": "integer"
},
"reasoning_effort": {
"anyOf": [
{
"enum": [
"low",
"medium",
"high"
],
"type": "string"
},
{
"type": "null"
}
],
"default": null,
"description": "Thinking (reasoning) opt-in: low, medium or high; omitted = off. On the platform, auto then picks only a model whose reasoning the platform can pilot, and the effort is translated per provider. On YOUR model, apply the effort yourself when you run the requests.",
"title": "Reasoning Effort"
},
"use_platform_models": {
"default": false,
"description": "false (default): every LLM call of this job runs on YOUR model — the job parks each call as a pending request you answer with answer_llm_request; no credits are consumed. Attachments are no reason to leave this default: the requests name the files you uploaded and you read your own copies, whatever their format. true: the job runs on the platform (billed) and, with the model left on auto, on the organization's best model for the task — its highest blended benchmark score under the organization's quality/speed/cost weights, among the models that satisfy the run (attachments, web search, thinking). Naming a platform model also runs the job there, bypassing that pick: do it only when the user named one. Protocol: enricher://docs/llm-delegation.",
"title": "Use Platform Models",
"type": "boolean"
}
},
"title": "generate_sampleArguments"
}