Get a structured answer from the model

local_task answers in prose unless you give it a JSON Schema. With one, the answer is JSON the agent can use without parsing sentences.

Uses the llm helper from the Quickstart and meeting-notes.txt. The outputs were produced with qwen3.8:latest, so yours will read differently.

Pick the kind

kind chooses the instructions the server gives the model. instruction says precisely what to do, and the input is text for a short string, files for up to eight paths in the workspace, or both.

Kind	Answers with	A long input is
`summarize`	a headline and the key facts	split and read whole
`extract`	the requested fields, null if absent	split and read whole
`classify`	labels	refused (`input_too_large`)
`transform`	the input rewritten as instructed	refused
`free`	a direct answer to the instruction	refused

Constrain the answer with a schema

With jsonSchema, Ollama constrains the model's output to the schema and the server parses it. The answer arrives as result instead of answer:

llm local_task --tool-arg kind=classify \
  instruction="Is this message a complaint, a question or praise, and how urgent is it?" \
  text="Third time this month the delivery came a day after the promised date." \
  'jsonSchema={"type":"object","properties":{"type":{"enum":["complaint","question","praise"]},"urgency":{"enum":["low","medium","high"]}},"required":["type","urgency"]}' \
  | jq -c '{kind, result}'

Output

{"kind":"classify","result":{"type":"complaint","urgency":"high"}}

An enum keeps the answer to values the agent already handles.

Extract fields from a file

extract pulls named fields out of the input:

llm local_task --tool-arg kind=extract \
  instruction="The trial budget in euros, who approved it, and every action item with its owner and due date." \
  'files=["meeting-notes.txt"]' \
  'jsonSchema={"type":"object","properties":{"budgetEur":{"type":["number","null"]},"approvedBy":{"type":["string","null"]},"actions":{"type":"array","items":{"type":"object","properties":{"owner":{"type":"string"},"action":{"type":"string"},"due":{"type":["string","null"]}},"required":["owner","action","due"]}}},"required":["budgetEur","approvedBy","actions"]}' \
  | jq '.result'

Output

{
  "budgetEur": 18000,
  "approvedBy": "Ada",
  "actions": [
    {
      "owner": "Emre",
      "action": "talk to a second carrier and report back",
      "due": "30 January"
    },
    {
      "owner": "Lena",
      "action": "remove fax number from returns form in next release",
      "due": null
    }
  ]
}

Mark every field required, and allow null for one the input may not contain, so a missing value is null rather than a guess or an absent key.

When the answer is not valid JSON

A model can fail a schema it cannot hold. The call then fails with unparsable_output instead of returning a half-parsed object. Retry once with a flatter schema and fewer fields, or do the work without the local model. Some models cannot follow a schema at all: gpt-oss:20b was measured scoring 0 out of 100 on the numbered-row schema that local_map and structured local_task calls rely on.

Besides the answer, every call returns promptTokens and outputTokens, and chunks and reduceRounds, which exceed 1 and 0 only when a long input was split; see Summarise a document longer than the model's window.