perfecXion/intentguard-legal

IntentGuard — Legal & Compliance

    

Production-ready vertical intent classifier for LLM chatbot guardrails. Classifies user messages as `allow`, `deny`, or `abstain` to keep legal chatbots on-topic and within professional boundaries.

Research Article | perfecXion.ai | Finance Model | Healthcare Model | Legal Model

IntentGuard Model Family

Overview

The Problem

Legal chatbots carry significant liability risk — unauthorized practice of law, confidentiality breaches, and jurisdictional issues demand strict topic boundaries. An LLM that happily answers questions about celebrity gossip or sports while branded as a legal assistant undermines professional credibility and trust.

The Solution

IntentGuard uses a tiny DeBERTa-v3-xsmall model (22M parameters, 2.5MB quantized) to classify user intent in <30ms on CPU with three-way classification:

• —Allow — On-topic legal query, pass to the LLM
• —Deny — Off-topic, block with a polite redirect
• —Abstain — Ambiguous, escalate to secondary classifier or human review

Performance

Model Details

Core Topics (Allow)

Contracts, litigation, employment law, intellectual property, criminal law, family law, real estate law, immigration, corporate law, compliance, privacy law, civil rights, estate planning, bankruptcy

Hard Exclusions (Deny)

Sports, entertainment, cooking, gaming, celebrity gossip, fashion, travel/leisure, fiction writing, relationship advice

Usage

Python (ONNX Runtime)

import onnxruntime as ort
from transformers import AutoTokenizer
import numpy as np

tokenizer = AutoTokenizer.from_pretrained("perfecXion/intentguard-legal")
session = ort.InferenceSession("model.onnx")

text = "What are the requirements for filing a patent application?"
inputs = tokenizer(text, return_tensors="np", max_length=128, truncation=True, padding="max_length")

logits = session.run(None, {
    "input_ids": inputs["input_ids"],
    "attention_mask": inputs["attention_mask"]
})[0]

labels = ["allow", "deny", "abstain"]
prediction = labels[np.argmax(logits)]
confidence = float(np.max(np.exp(logits) / np.sum(np.exp(logits))))

print(f"Intent: {prediction} (confidence: {confidence:.3f})")
# Output: Intent: allow (confidence: 0.996)

Docker

docker pull ghcr.io/perfecxion/intentguard:legal-1.0
docker run -p 8080:8080 ghcr.io/perfecxion/intentguard:legal-1.0

curl -X POST http://localhost:8080/v1/classify \
  -H "Content-Type: application/json" \
  -d '{"messages": [{"role": "user", "content": "What are my rights as a tenant?"}]}'

pip

pip install intentguard

from intentguard import IntentGuard
guard = IntentGuard.load("legal")
result = guard.classify("What are the requirements for filing a patent?")
print(result)  # Intent(label='allow', confidence=0.996)

Example Classifications

Citation

@misc{thornton2025intentguard,
  title={IntentGuard: A Production-Grade Vertical Intent Classifier for LLM Guardrails},
  author={Thornton, Scott},
  year={2025},
  publisher={perfecXion.ai},
  url={https://perfecxion.ai/articles/intentguard-vertical-intent-classifier-llm-guardrails.html},
  note={Model: https://huggingface.co/perfecXion/intentguard-legal}
}

Quality Metrics

License

Apache 2.0

Links

• —Research Article: IntentGuard: A Production-Grade Vertical Intent Classifier
• —Publisher: perfecXion.ai
• —Finance Model: perfecXion/intentguard-finance
• —Healthcare Model: perfecXion/intentguard-healthcare
• —Docker Image: ghcr.io/perfecxion/intentguard:legal-1.0