How does the detector decide the format?

It runs a set of heuristics in order. It tries to parse the text as JSON, including after stripping a code fence; it looks for markdown signals like headings, bullets, and links; it detects code by language keywords and fenced blocks; and it checks for CSV-style delimited rows. The strongest match wins, and JSON that actually parses takes priority.

Why does it matter whether JSON parses?

A response can look like JSON but be invalid — a trailing comma, a missing quote, or surrounding prose. The detector tells you whether the JSON is genuinely parseable, which is the difference between safely calling JSON.parse and needing a repair or re-prompt step in your pipeline.

What does mixed-format mean?

Many responses combine prose with a fenced JSON or code block — for example an explanation followed by a code sample. The tool flags these so you know to extract the structured part rather than parsing the whole response, which would fail.

Yes. All detection happens in your browser with no API calls, so it is instant and your output never leaves the page. That makes it safe for confidential responses and free to use as often as you like.

LLM Output Structure Detector

Route LLM responses by what they actually are

When you process model output programmatically, the first question is always: what format is this? A response might be clean JSON, markdown, a code snippet, CSV rows, or plain prose — and treating one as another breaks your pipeline. This detector classifies the structural type, validates JSON for real, and flags responses that mix prose with a structured block so you can extract instead of parse.

How it works

The detector applies ordered heuristics. First it attempts to parse the text as JSON, both directly and after stripping a surrounding code fence, because valid JSON is the most actionable result. If that fails it scans for markdown signals — ATX headings, bullet and numbered lists, links, and emphasis. It detects source code by fenced language hints and common keyword density, and checks for consistent delimited rows that indicate CSV. The format with the strongest evidence is reported, along with a note on whether any embedded JSON or code block is present, which is the mixed-format case.

Tips and notes

Use the parse result, not just the type label: a response can look like JSON but fail to parse, and that distinction determines whether you can call JSON.parse directly or need a repair step. When a response is flagged as mixed-format, reach for an extraction step that pulls out the fenced block rather than parsing the whole thing. Detection is heuristic, so for short or ambiguous inputs it can misclassify — pair it with strict parsing in code. Everything runs locally and instantly, so drop it into any debugging workflow without cost or privacy concerns.