Utilities for extracting structured data from LLM responses.

Provides lightweight JSON-mode extraction and schema validation so consuming applications can request structured output without pulling in heavy dependencies like Instructor.

Supported Formats

• {:json_schema, schema} — Decode JSON from the response, validate against a schema map
• {:json_schema, schema, opts} — Same, with custom validator and options

• {:custom, fun} — Apply a custom extraction function (Response.t() -> {:ok, term()} | {:error, term()})

Usage

Structured output is typically requested via response_format in the dispatch opts:

{:ok, response} = LlmCore.send(prompt, :extraction,
  response_format: {:json_schema, %{type: "object", properties: %{name: %{type: "string"}}}}
)

response.structured
#=> %{"name" => "value"}

Or process a response directly:

{:ok, response} = Structured.process({:ok, raw_response}, {:json_schema, schema})

Custom Validators

{:json_schema, schema, validator: fn value ->
  if value["score"] > 0.5, do: {:ok, value}, else: {:error, "score too low"}
end}