queryclf 0.1.0

Ultra-fast, lightweight ONNX query difficulty classifier for pre-routing LLM cascades

QueryClassifier

Goal: A pip-installable Python library that classifies a text query into low / medium / hard complexity, for use as a pre-routing step in an LLM cascade (cheap model → medium model → frontier model).

Approach: Fine-tuned answerdotai/ModernBERT-large quantized into int8 ONNX format (query-classifier-onnx). Uses a FastEmbed-inspired architecture powered by pure ONNXRuntime + Rust tokenizers — zero PyTorch, zero transformers library dependencies, and zero heavy GPU requirements.

---

⚡ Quickstart

Installation

pip install queryclf

Dependencies installed are lightweight (~38MB total): onnxruntime, tokenizers, numpy, huggingface-hub, and python-dotenv.

---

⚙️ Configuration (.env)

Configure model paths dynamically in .env (refer to .env.example):

# Hugging Face Repository ID (Default: prvn-ramesh/query-classifier-onnx)
QUERY_CLASSIFIER_HF_REPO=prvn-ramesh/query-classifier-onnx

# Optional local model directory override
QUERY_CLASSIFIER_MODEL_PATH=

Resolution Priority:

1. Explicit Python Argument (hf_repo="..." or model_path="...")
2. Environment Variable QUERY_CLASSIFIER_HF_REPO
3. Environment Variable QUERY_CLASSIFIER_MODEL_PATH
4. Default Hugging Face Repository (prvn-ramesh/query-classifier-onnx)

---

🐍 Python API Usage

import asyncio
from query_classifier import QueryClassifier

# Initializes classifier with built-in LRU cache and CPU thread tuning
classifier = QueryClassifier(
    cache_size=1000,             # Max items in LRU query cache (default: 1000)
    intra_op_num_threads=4,      # Intra-node CPU parallelism
)

# 1. Single query prediction
result = classifier.predict("Can you summarize this document?")

print(result.label)        # Output: "low"
print(result.confidence)   # Output: 0.5779
print(result.latency_ms)   # Output: 29.8 ms
print(result.scores)       # Output: {'low': 0.5779, 'medium': 0.4201, 'hard': 0.0020}

# 2. Batch query prediction
queries = [
    "What is the capital of France?",
    "Calculate the integral of x^2 * sin(x) dx",
    "Write a lock-free multi-threaded queue in C++ using atomics"
]

batch_results = classifier.predict_batch(queries)
for q, res in zip(queries, batch_results):
    print(f"[{res.label.upper()}] ({res.latency_ms:.1f}ms) -> {q}")

# 3. Async prediction (for FastAPI / AsyncIO frameworks)
async def main():
    async_res = await classifier.predict_async("Explain quantum computing")
    print(f"Async prediction: {async_res.label} ({async_res.confidence:.2f})")

asyncio.run(main())

# 4. Cache statistics & management
print(classifier.cache_info())  # e.g., {'hits': 1, 'misses': 4, 'maxsize': 1000, 'currsize': 4}
classifier.clear_cache()

---

💻 CLI Command Usage

# Single query classification
queryclf "What is 2 + 2?"

# JSON output mode
queryclf "Write a Rust lock-free SPMC queue" --json

# Batch mode from file
queryclf --file queries.txt

---

📊 Benchmark & Accuracy Results

The underlying model prvn-ramesh/query-classifier-onnx is evaluated on an independent held-out test set of 301 unseen queries (balanced across low, medium, and hard buckets). For full details, see the Hugging Face Model README.

Accuracy & F1-Score Breakdown

• Overall Accuracy: 88.04%
• Macro F1-Score: 0.8797

Class	Precision	Recall	F1-Score	Support
low	100.00%	81.00%	0.8950	100
medium	82.18%	83.00%	0.8259	100
hard	84.87%	100.00%	0.9182	101
Overall	88.04% Acc	—	0.8797 Macro-F1	301 total

Latency & Throughput Profile (CPU Execution)

Metric	Measurement
Median Latency (p50)	67.68 ms
Mean Latency	81.44 ms
p90 Latency	134.69 ms
p99 Latency	161.74 ms
Throughput	12.3 queries/sec

---

🛠️ Project Roadmap Status

• Phase 0 — Environment & Scaffold: Package directory layout (src/query_classifier, tests/, pyproject.toml).
• Phase 2 — Labeled Dataset: dataset.csv with query samples.
• Phase 4 & 5 — Fine-Tuned ModernBERT & ONNX int8 Quantization: Quantized model published on Hugging Face at prvn-ramesh/query-classifier-onnx.
• Phase 6 — Inference Engine: Pure ONNXRuntime + Tokenizers inference engine (onnx_engine.py, classifier.py, cli.py).