Comment Triage for Creators (Local AI)

AI & LLM Tools

Privacy Note: Comments are classified 100% inside your browser (ONNX Runtime Web) by either a small MiniLM embedding model or the open-source Laya decision model. The first run downloads 22-422 MB depending on your choice, then it is cached. Nothing you paste is sent to a server.

Comments (one per line, up to 500)

0 comments

Model

Compares each comment with 30 built-in example comments. Fast, about 77% on our test set. Percentages are relative match scores and the toxic flag is a rough similarity check.

What is Comment Triage for Creators (Local AI)?

Comment Triage sorts a pasted batch of audience comments into question, praise, criticism, spam or other, and flags comments that look abusive. You choose between a small MiniLM embedding model with built-in example comments, and Laya, an open-source decision model that answers typed questions about a text in a single forward pass. Both run entirely in your browser.

How it works

The MiniLM option embeds each comment and picks the closest of five categories, using the centroid of 30 built-in example comments per set. The Laya options ask two typed questions per comment - one choice question and one yes/no question - and read probabilities from one ONNX forward pass. On our 60-comment test set the category accuracy was about 77% (MiniLM), 82% (Laya 278 MB) and 85% (Laya 422 MB). Labels are a first pass and not a verdict; review flagged items yourself.

Features & Benefits

Comments never leave your device - the model runs in a Web Worker via ONNX Runtime Web
Two typed questions per comment: its category and whether it insults or attacks someone
Three models to trade size for accuracy: 22 MB MiniLM, 278 MB Laya, or 422 MB Laya

Frequently Asked Questions

Is anything uploaded?

No. The model files download once from Hugging Face; your comments are processed locally and never sent anywhere.

How accurate is it?

Roughly 77-85% on a small test set of short comments, and none of the models is tuned on your audience. The Laya base model can be over-confident, and the MiniLM toxic flag is a rough similarity check. Use it to prioritise your reading, not to delete or ban automatically.

Which model should I pick?

Start with MiniLM (22 MB): it loads in seconds and runs in about a millisecond per comment. Choose a Laya build if you want the extra accuracy and can spare the download; both Laya builds are cached after the first load.

Related Tools

Local Vector Embeddings Tester (ONNX)

Compute cosine similarity locally using an ONNX embedding pipeline. Zero delays.

Audio Transcription

Transcribe audio to text entirely in your browser using Whisper. No upload, no server - your audio never leaves your device.

PII Redactor - Sensitive Data Masker

Redact emails, SSNs, credit cards, IP addresses, and phone numbers from text and logs. Add custom regex patterns. Runs entirely in your browser.

Popular Utilities

JSON Formatter & Validator

Format, validate, and minify JSON instantly in your browser. Your data never leaves your device.

JWT Decoder

Decode JWT tokens and inspect header and payload instantly in your browser. Your tokens never leave your device.

Word Counter

Count words, characters, sentences, and estimate reading time instantly in your browser. No sign-up required.

Learn More & Guides

May 12, 2026