Gonyai: Ultra-Lightweight Konkani Intent Router (ONNX)

Model: Elfibolt/gonyai-intent-classifier-onnx
Size: 2.00 MB (ONNX)
Latency: < 1 millisecond (CPU)
Language: Goan Konkani (Devanagari script)
Classes (4): WEATHER, NEWS, CONTEXT, DIRECT

🎯 Overview & Purpose

In the Gonyai Konkani AI Assistant, this classifier acts as a high-speed router in front of the main LLM.

Before generating an answer, it classifies user queries in under 1 millisecond on CPU to determine if external information needs to be retrieved and injected inside a <संदर्भ> ... </संदर्भ> (Context) tag:

WEATHER $\rightarrow$ Fetches real-time weather/rain forecast and injects into <संदर्भ> tag.
NEWS $\rightarrow$ Fetches live news / breaking event search and injects into <संदर्भ> tag.
CONTEXT $\rightarrow$ Retrieves text from uploaded user documents, bills, reports, or RAG store and injects into <संदर्भ> tag.
DIRECT $\rightarrow$ Standard conversations, poetry, advice, recipes (bypasses retrieval, no <संदर्भ> tag).

📊 Benchmark Evaluation Across 70 Adversarial Prompts

The model was rigorously stress-tested across 70 tricky, indirect, and adversarial Konkani test cases designed with zero explicit keywords:

Adversarial Suite	Test Scenario	Score
Suite 1 (Extended)	Indirect questions & hard negatives	20 / 20 (100.0%)
Suite 2 (Hardcore)	Cross-domain traps (flood insurance, sowing seeds)	25 / 25 (100.0%)
Suite 3 (Nightmare)	Ultra-indirect idioms (drying papads, humidity, dew point)	22 / 25 (88.0%)
Total Adversarial Score	70 challenging edge cases	67 / 70 (95.7%)

Note: In boundary cases, the model safely defaults to DIRECT, preventing false API triggers.

🚀 Quickstart Usage

Installation

pip install onnxruntime numpy scikit-learn huggingface_hub

Python Inference

import json
import numpy as np
import onnxruntime as ort
from huggingface_hub import hf_hub_download
from sklearn.feature_extraction.text import HashingVectorizer

REPO_ID = "Elfibolt/gonyai-intent-classifier-onnx"

# 1. Download model and config
model_path = hf_hub_download(repo_id=REPO_ID, filename="model.onnx")
config_path = hf_hub_download(repo_id=REPO_ID, filename="config.json")

with open(config_path, "r", encoding="utf-8") as f:
    config = json.load(f)

# 2. Setup Vectorizer & ONNX Runtime Session
vectorizer = HashingVectorizer(
    n_features=config["n_features"],
    ngram_range=(config["ngram_min"], config["ngram_max"]),
    analyzer="char_wb",
    alternate_sign=False
)
session = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])
input_name = session.get_inputs()[0].name
output_name = session.get_outputs()[0].name
id2label = {int(k): v for k, v in config["id2label"].items()}

def route_query(query: str, confidence_threshold=0.55):
    feat = vectorizer.transform([query]).toarray().astype(np.float32)
    logits = session.run([output_name], {input_name: feat})[0][0]
    probs = np.exp(logits - np.max(logits))
    probs = probs / probs.sum()
    
    top_idx = int(np.argmax(probs))
    intent = id2label[top_idx]
    confidence = float(probs[top_idx])
    
    if confidence < confidence_threshold:
        intent = "DIRECT"
        
    needs_context = intent in ["WEATHER", "NEWS", "CONTEXT"]
    return intent, confidence, needs_context

# 3. Test Query
query = "आयज दनपरां म्हापश्यां भायली स्थिती पळयल्यार गाडी स्लिप जावपाची शक्यताय आसा काय?"
intent, conf, needs_ctx = route_query(query)
print(f"Query: {query}")
print(f"Intent: {intent} ({conf*100:.1f}%), Needs <संदर्भ>: {needs_ctx}")

🏷️ Intent Classes

WEATHER: Inquiries about live weather, rain forecasts, travel conditions, marine waves, temperature, humidity.
NEWS: Breaking events, civic utility outages, sports outcomes, policy decisions, local incident verification.
CONTEXT: Document grounding queries referencing uploaded bills, lab reports, contracts, RTI documents, gazettes.
DIRECT: Creative writing, general advice, recipes, science explanations, poetry, and conversational chat (no external retrieval needed).

📜 License

Apache-2.0

Downloads last month: -