app.py

import gradio as gr
import torch
from sentence_transformers import SentenceTransformer, util

# -----------------------------
# Load model (FAST)
# -----------------------------
model = SentenceTransformer("all-MiniLM-L6-v2")

# -----------------------------
# Intent anchors
# -----------------------------
INTENT_LABELS = {
    "chat": [
        "casual conversation",
        "small talk",
        "talk to me",
        "just chatting"
    ],
    "search": [
        "find information",
        "look up facts",
        "search the web"
    ],
    "image_generation": [
        "generate an image",
        "create artwork",
        "draw a picture"
    ],
    "code": [
        "write code",
        "debug program",
        "programming help"
    ],
    "research": [
        "deep research",
        "academic analysis",
        "investigate topic"
    ],
    "study": [
        "help me learn",
        "teach me something",
        "study help"
    ],
    "project": [
        "build a project",
        "create an application",
        "long term development"
    ],
    "action": [
        "set a reminder",
        "schedule something",
        "perform an action"
    ]
}

# -----------------------------
# Precompute anchor embeddings
# -----------------------------
anchor_embeddings = {
    intent: model.encode(phrases, convert_to_tensor=True)
    for intent, phrases in INTENT_LABELS.items()
}

# -----------------------------
# Classifier
# -----------------------------
def classify(text: str):
    if not text.strip():
        return {"error": "Empty input"}

    query_emb = model.encode(text, convert_to_tensor=True)

    scores = {}
    for intent, emb in anchor_embeddings.items():
        score = util.cos_sim(query_emb, emb).max().item()
        scores[intent] = score

    top_intent = max(scores, key=scores.get)

    return {
        "text": text,
        "top_intent": top_intent,
        "scores": scores
    }

# -----------------------------
# Gradio UI
# -----------------------------
demo = gr.Interface(
    fn=classify,
    inputs=gr.Textbox(label="User Input"),
    outputs=gr.JSON(),
    title="⚡ Ultra-Fast Intent Router",
    description="Sub-second semantic routing using MiniLM (no training)"
)

app = demo