Text Classification API in India

Text Classification at Scale

Classify reviews, support tickets, or user feedback into predefined buckets—sentiment, topic, intent, risk level. Run millions of classifications through unoblox, all billed in rupees.

Classification strategies

Zero-shot: Define labels in the prompt; no training needed. Few-shot: Provide 3–5 examples per label in the prompt. Fine-tuned: (Roadmap) Train a classifier on your data.

Start with zero-shot for speed; move to few-shot if accuracy drifts.

Budget-friendly models

Model	Input/Output (₹)	Use case
Qwen3 1.7B	₹0 / ₹0	Real-time, high-volume, simple labels
Qwen3.8-27B	₹16.32 / ₹48.96	Higher accuracy, still cheap
Claude Sonnet	₹201.6 / ₹1008	Complex, nuanced categories
GPT-4o	₹252 / ₹1008	Reasoning over ambiguous text

3-step classification pipeline

Step 1: Define categories and prompt

You classify customer feedback into: [Positive, Negative, Neutral, Feature Request, Bug Report]
Respond with ONLY the category name, no explanation.

Step 2: Batch-call the API

POST https://api.unoblox.ai/v1/chat/completions
Authorization: Bearer ub-gw-...

{
  "model": "qwen/qwen3-1.7b",
  "messages": [{"role": "user", "content": "[Your prompt + text to classify]"}],
  "max_tokens": 50
}

Step 3: Parse and store results — Extract the label from the response, batch-insert into your database, track confidence/errors.

Real costs: 100k classifications

• Qwen3 1.7B: ₹0 (freemium, throttled at ₹250/month if free tier).
• Qwen3.8-27B: ₹160–200 (depending on prompt+result length).
• No cold starts, no fine-tuning costs—just usage.

Accuracy tuning

Start with Qwen3 1.7B; if accuracy <85%, add one few-shot example per label. If still <90%, migrate to Qwen3.8-27B or Claude Sonnet.

Frequently asked questions

Q: How many labels can I classify at once? A: Practically 10–30 labels in the prompt before context fills up. For 100+ labels, build a two-tier taxonomy or use embeddings + nearest-neighbor instead.

Q: Can I fine-tune a model on unoblox? A: Not yet; on the roadmap. For now, use prompt engineering (few-shot examples, instruction refinement) to improve accuracy.

Q: How do I classify at high volume (1M+ items)? A: Batch requests concurrently (10–50 in flight is typical) and prefer Qwen3 1.7B or Qwen3.8-27B for throughput. Track failures and retry with backoff — at these price points, cost stays low even at millions of calls a month.

Q: How do I handle edge cases (ambiguous text)? A: Add a "Unclear" label, or set a confidence threshold—ask the model to explain reasoning in a follow-up prompt, then review by hand.

Q: Can I chain classifiers (e.g., intent → urgency → escalation)? A: Yes. Call the API multiple times in sequence, or build a prompt that does multi-stage reasoning in one call.

Q: How do I monitor accuracy over time? A: Sample 1% of results, have humans review, compare predicted vs. actual. Log misclassifications to a separate table; retrain prompts or switch models quarterly.

Get started in rupees → https://unoblox.ai/sign-in