Why does step order matter?

Each step transforms the text and passes it to the next, so order changes the result. For example, stripping markdown before extracting JSON can break a fenced JSON block, while extracting JSON first then trimming whitespace gives clean parseable output. The builder runs steps top to bottom so you control the sequence.

What does the JSON extraction step do?

It pulls the first balanced JSON object or array out of the text — including from inside a fenced code block — and discards the surrounding prose. This is the common case where a model wraps a JSON response in explanation, and you want only the structured part.

How does PII anonymization work?

It replaces email addresses and phone-number-like sequences with placeholder tokens such as [EMAIL] and [PHONE] using pattern matching. It is a lightweight redaction aid for previews and logs, not a certified compliance tool, so review the result before relying on it for regulated data.

Is anything sent to a server?

No. Every transform runs locally in your browser, so your text and any sensitive content never leave the page. That makes it safe to experiment with redaction and cleaning on real output.

Output Post-Processing Pipeline Builder

Build a repeatable cleanup chain for LLM output

Raw model output rarely goes straight into your app. You usually want to strip boilerplate, pull out the JSON, normalize whitespace, or redact PII before display or storage. This builder lets you toggle those transforms and arrange them in order, then applies them in sequence so you can see exactly what your production post-processing chain would produce.

How it works

Each enabled step is a pure text transform applied in the order shown. Strip boilerplate removes common LLM lead-ins and sign-offs like “Sure, here is” and “Let me know if you need anything else”. Extract JSON scans for the first balanced object or array, including inside code fences, and keeps only that. Strip markdown removes headings, emphasis, links, and code markers. Trim whitespace collapses repeated spaces and blank lines. Anonymize PII swaps emails and phone-like sequences for placeholder tokens. The output of one step becomes the input to the next, so reordering changes the result and the preview updates live.

Tips and notes

Order is the most important control here. If you want clean JSON, put extract JSON near the top before any markdown stripping that could damage a fenced block. If you want readable prose, strip markdown and boilerplate first, then trim whitespace last. The PII step is a quick redaction aid using pattern matching — fine for previews and logs, but verify the result before treating it as compliance-grade. Everything runs locally, so you can safely test on real sensitive output. Once you settle on an order that works, replicate the same sequence of transforms in your own code.