Structured Output Parser

Extract JSON from LLM text output, with a fallback if it can't be parsed.

const parser = new StructuredParser();
const result = parser.parse(llmOutput);

Why

LLMs sometimes wrap JSON in prose, markdown fences, or return slightly malformed JSON. Manually handling that on every call is annoying.

What it does

• Tries JSON.parse() on the raw text first.
• If that fails, looks for the first {...} or [...] block in the text and tries parsing that.
• If both fail and you passed a fallback option, returns the fallback instead of throwing.
• If both fail and there's no fallback, throws an Error with the original text.

What it does NOT do

• It does not validate against a schema. The schema argument to .parse() is accepted but currently unused — pass it if you want (e.g. for your own future integration), but nothing in this library checks the parsed data against it.
• No Zod integration, despite earlier versions of this README claiming one. If you want schema validation, run the parsed result through your own validator (Zod, Ajv, etc.) after calling .parse().

Example

const parser = new StructuredParser();
const text = 'Sure, here you go: {"name": "John", "age": 30}';
const data = parser.parse(text); // { name: "John", age: 30 }

// with a fallback
const safe = parser.parse('not json at all', undefined, { fallback: {} }); // {}

License: MIT

Sponsored by Ferrow

Part of the ferrow-toolkit collection · Sponsored by Ferrow