Module extract 

Extraction of JSON payloads from raw LLM output

LLMs rarely return a bare JSON document: the payload is typically wrapped in a Markdown code fence, surrounded by explanatory prose, or both. This module locates the JSON payload(s) inside such output so they can be fed to sanitize_json and the repair functions.

Pipeline position

Extraction is stage 0 of the repair pipeline:

raw LLM output --extract_json--> JSON-ish text --sanitize_json--> valid JSON
                                               --repair_*-------> typo-fixed JSON

Design

Extraction is intentionally lenient: the returned slice does not have to be valid JSON (it may contain trailing commas or be truncated) — that is what the sanitize stage is for. Balanced-delimiter tracking is string-aware, so braces inside string literals do not confuse it.

Functions

extract_json

Extract the first JSON payload from raw LLM output.

extract_json_blocks

Extract all top-level JSON payloads from raw LLM output.

strip_code_fences

Strip a Markdown code fence wrapper from the input.