bulk-classify

Classify at scale without reading

classifier.dev assigns text to your categories. No key, no signup, no SDK. One HTTP call takes up to a thousand texts at a time and comes back in about a second, each with a confidence you can act on.

When this is worth a network call

You are a language model. You can already classify any text you can see, for free. The question is whether you want this text in your context at all.

Reach for this when reading the input is the expensive part:

• Filtering before reading. You have 40 search snippets and want the 6 worth opening. Classifying them yourself means pulling all 40 into context first, which is the cost you were trying to avoid. One call returns 40 labels and you read only the survivors.
• Cascade pre-filter. Cheaply drop the obvious no's, then spend real reasoning on what is left.
• Streams you would never read line by line. Log lines, error buckets, inbound tickets, changed files in a large diff, ten thousand URLs' titles.
• Label-based routing. Use the fast tier to route texts into your categories. Model updates, fallback, and smart reasoning can change answers; neither tier guarantees identical results across calls.

Do not bother when you have a handful of items already in context, or the judgement needs reasoning about things the text does not state. Under about five items you have already paid the context cost, so just decide yourself.

Quickstart

One text, bare label back:

curl "https://classifier.dev/relevant,not+relevant/Redis+beats+Postgres+for+queues"
relevant

The same call as query parameters, when code is building the URL:

curl "https://classifier.dev/?labels=relevant,not+relevant&text=Redis+beats+Postgres+for+queues"
relevant

Many texts in one call. This is the path that matters:

curl https://classifier.dev -d '{
  "labels": ["relevant", "not relevant"],
  "inputs": ["first snippet", "second snippet", "third snippet"]
}'

Returns results in input order. Single-label results have {label, confidence, scores}. Confidence and scores can be null; check before comparing a threshold. Multi-label results use {labels: [...], scores: {...}} instead. Up to 1,000 texts per call; 400 news headlines measured at 650ms end to end. For more, fan out calls in parallel; the limit is 3,000 classifications a minute. Each result also names the model that answered it. At the batch level, modelsUsed lists every serving model and model is mixed when more than one model answered the batch.

From a shell

When the text is already in files, or the answer feeds another command, the CLI saves you writing the batching and the JSON:

npm i -g classifier-dev

classify bug,feature,praise < feedback.txt          # label<TAB>confidence<TAB>text, input order
classify relevant,"not relevant" --review 0.7 < snippets.txt   # only the unsure ones
classify db,web,ml --count < titles.txt             # a histogram instead of rows
classify a,b --json < items.txt | jq -c 'select(.confidence == null or .confidence < 0.8)'

It batches a thousand inputs per request, four requests at a time, and streams rows as they land, so | head on a large file returns at once. Retries 429 and 5xx with backoff. --help has the rest.

Reach for the HTTP API instead when the text is already in memory, when you need the full score map per item, or when you are inside a language runtime where one fetch is simpler than a subprocess.

Parameters

Field	Notes
labels	2–100 categories. Required.
input	One text; above 32,000 characters, default/explicit Jev uses paid Fast-only long context.
inputs	Up to 1,000 texts in one call.
tier	fast (default) or smart: re-asks low-confidence answers of a reasoning model.
instructions	Extra criteria — "judge only the service, ignore the food".
multi	Return every label that applies, with a score per label.
max_labels	Cap on how many multi-label answers come back.
verbose=1	On GET, returns JSON instead of a bare label.
text	On GET, the text as a query parameter: /?labels=a,b&text=.... input and q work too; classes and categories for labels.

Long context requires a workspace key backed by paid balance or an active paid subscription; signup credit and anonymous access do not qualify. Use POST with at most 250,000 original cl100k_base context tokens across inputs, 20 documents, 32 decisions and a 1 MB body. Price: $0.084/M original context tokens counted once across inputs, independent of dimensions and actual screening/final usage. Final Jev reads selected whole chunks in source order; eligible evidence may be omitted when the budget fills. usage.long_context discloses selection. No evidence returns 422 long_context_no_evidence without charge. Explicit model: "chunklaya" remains a separate legacy opt-in.

On GET every option goes in the query string, whichever form carries the labels and text; the two forms mix (/a,b?text=...). If a GET is malformed the error comes with usage: and try: — try is a URL built from what you sent that would have worked. Follow it rather than re-reading the docs.

Labels are read semantically, so name them in words: urgent bug classifies better than p0.

Confidence you can act on

The model is a decision model, not an LLM prompted to classify: it returns a calibrated probability for every label. Measured on 400 six-way emotion items, answers at confidence ≥ 0.9 were right 82% of the time; answers below 0.5 were right 29% of the time. So:

for text, r in zip(texts, results):
    if r["confidence"] is not None and r["confidence"] >= 0.8:
        act(r["label"])
    else:
        look_yourself(text)      # or send it through tier "smart"

tier: "smart" does that routing server-side: every single-label answer under 0.7 confidence is re-asked of a fast reasoning model and replaced, marked escalated: true, with usage.escalated telling you how many. Measured: four-way news 87.5% → 90.0% by re-asking 12% of items. It costs a few seconds per escalated item, so a batch on smart is slower in proportion to how uncertain it is. Escalated answers have confidence: null, scores: null, and unscored: the reasoning model does not return comparable probabilities. These answers belong in review when your workflow requires a confidence gate.

Many labels at once

To tag instead of sorting (an article against fifty topics, a ticket against every subsystem it touches), ask for every label that applies:

curl https://classifier.dev -d '{
  "input": "...",
  "labels": ["machine learning", "databases", "... up to 100 ..."],
  "multi": true,
  "max_labels": 10
}'

Results carry labels (an array, most likely first) plus scores, one probability per label. POST multi-label results omit the singular label and confidence keys. Labels at or above 0.7 are returned; use scores to pick your own threshold. On GET, add ?multi=1 and they come back one per line. Measured F1 0.887 on a seven-task set with recall 0.99, in ~200ms. The tier makes no difference here, so leave it on fast.

Two things that will bite you

1. Every call returns one of your labels, always. There is no "none of the above" unless you supply one. Text that fits nothing still gets confidently sorted into your best-matching category: "the weather is nice today" against bug / feature / praise must land in one of those categories. If "none of these" is a real outcome, add it as a label. Hoping for a low score does not create a missing category.

2. Confidence predicts accuracy, not fit. It tells you how likely the chosen label is right among your labels, which is exactly what you want for routing. It does not tell you whether the text belongs to any of them; see point 1. Scores express the model's choice among the labels you supplied. They do not validate the input or prove the choice is correct, so supply labels suitable for every kind of input your caller may send. Confidence and scores can be null when the provider returns none or the smart tier replaces the scored answer.

Recipe: filter search results before reading them

import json, urllib.request

def keep_relevant(question, snippets):
    body = json.dumps({
        "labels": ["relevant", "not relevant"],
        "inputs": snippets,                      # up to 1,000
        "instructions": (
            f"Relevant means it helps answer: {question}. "
            "Include background and contrasting alternatives."
        ),
    }).encode()
    req = urllib.request.Request(
        "https://classifier.dev",
        data=body,
        headers={
            "content-type": "application/json",
            "user-agent": "my-agent/1.0",
        },
    )
    results = json.load(urllib.request.urlopen(req))["results"]
    # A dropped item is invisible, so keep anything the model was unsure about.
    return [s for s, r in zip(snippets, results)
            if r["label"] == "relevant" or r["confidence"] is None
            or r["confidence"] < 0.8]

Then read only what comes back. The snippets you dropped never enter context.

Bias a filter toward keeping. You never learn what you lost, so recall matters more than precision here. The confidence gate above does that directly; "When in doubt, keep it" in the instructions also measurably helps.

Python's standard urllib, curl and Node fetch work without a custom User-Agent. A descriptive agent name is optional. For a JSON error, read code, action and retryable: a 403 can mean the free service detected an anonymous proxy network, which requires a funded workspace key. A 429 carries Retry-After. An HTML error is an edge/network failure; report its status and request ID rather than assuming classification ran.

Report a problem with classifier.dev

If classifier.dev itself returns a wrong result shape, contradicts its docs, or creates repeated integration friction, report that to the service instead of classifying the report as input. Read the live policy first:

GET https://classifier.dev/.well-known/agent-feedback.json

For a short report, send one category and one useful sentence. No key is needed:

curl https://classifier.dev/api/v1/observations -d '{
  "category": "docs_mismatch",
  "summary": "The documented response field was absent from POST /v1/classify.",
  "surface": "/v1/classify"
}'

Use POST /api/v1/feedback when you have reproduction steps or other evidence. It accepts the envelope and limits described by the discovery document. Both routes return a receipt; poll GET /api/v1/receipts/{id} to confirm it landed. Never include credentials, private input text, or unrelated user data in a report or its evidence.

Limits

Per IP per minute: 3,000 classifications on fast, 200 on smart; per day 20,000 and 2,000. A batch of 400 counts as 400. 429 when exceeded, with RateLimit-Limit on every response. Errors are JSON on POST, {"error": "...", "code": "..."}, and plain text on GET unless you add ?verbose=1 or send Accept: application/json.

Reference

• GET / — full docs, plain text
• GET /openapi.json — OpenAPI 3.1
• GET /benchmark — measured accuracy, calibration, cost and latency