---
language:
  - en
  - de
base_model: sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
pipeline_tag: text-classification
tags:
  - onnx
  - intent-classification
  - education
  - setfit
  - quantized
library_name: onnxruntime
license: mit
---

# Search Intent Classifier

Binary text classifier that decides whether a student's search query needs an AI-generated answer (`trigger_ai`) or can be satisfied with search results alone (`skip_ai`).

## Model Details

| Property | Value |
|---|---|
| Base model | [sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2](https://huggingface.co/sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2) |
| Architecture | BERT (12 layers, 384 hidden size) |
| Task | Binary text classification |
| Quantization | Dynamic INT8 (weights QInt8, activations QUInt8) |
| Format | ONNX Runtime |
| Languages | English, German (multilingual) |

## Labels

| Label | ID | Meaning |
|---|---|---|
| `skip_ai` | 0 | Query can be answered with search results (e.g. "Lecture 4", "Exercise Sheet 2") |
| `trigger_ai` | 1 | Query needs an LLM-generated explanation (e.g. "explain recursion", "wie funktioniert hashing") |

## Usage

```python
import numpy as np
import onnxruntime as ort
import joblib
from transformers import AutoTokenizer

# Load model components
tokenizer = AutoTokenizer.from_pretrained("sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2")
session = ort.InferenceSession("model_quantized.onnx")
head = joblib.load("model_head.joblib")

def predict(query: str) -> str:
    inputs = tokenizer(query, return_tensors="np", truncation=True, max_length=512, padding=True)
    outputs = session.run(None, {
        "input_ids": inputs["input_ids"],
        "attention_mask": inputs["attention_mask"],
        "token_type_ids": inputs["token_type_ids"],
    })
    embedding = outputs[0].mean(axis=1)
    label_id = head.predict(embedding)[0]
    return "trigger_ai" if label_id == 1 else "skip_ai"

print(predict("explain recursion"))                          # trigger_ai
print(predict("how does quicksort work step by step"))       # trigger_ai
print(predict("was ist ein Deadlock"))                       # trigger_ai
print(predict("Lecture 4"))                                  # skip_ai
print(predict("Exercise Sheet 2"))                           # skip_ai
print(predict("Vorlesung 7 Folien"))                         # skip_ai
```

## Training Data

Trained on a synthesized dataset of student search queries generated by Claude (Anthropic) in English and German.