language/intent.py

import torch
import torch.nn as nn
from core.device import DEVICE


class IntentClassifier(nn.Module):
    def __init__(self, input_dim: int, intent_labels: list):
        """
        input_dim: dimension of sentence embeddings
        intent_labels: list of intent names
        """
        super().__init__()

        self.intent_labels = intent_labels
        self.num_intents = len(intent_labels)

        self.classifier = nn.Sequential(
            nn.Linear(input_dim, 512),
            nn.ReLU(),
            nn.Dropout(0.3),
            nn.Linear(512, 256),
            nn.ReLU(),
            nn.Dropout(0.3),
            nn.Linear(256, self.num_intents)
        )

        self.softmax = nn.Softmax(dim=-1)

        self.to(DEVICE)

    def forward(self, sentence_embedding: torch.Tensor) -> torch.Tensor:
        sentence_embedding = sentence_embedding.to(DEVICE)
        return self.classifier(sentence_embedding)

    def predict(self, sentence_embedding: torch.Tensor, threshold=0.4) -> dict:
        with torch.no_grad():
            logits = self.forward(sentence_embedding)
            probs = self.softmax(logits)

            confidence, idx = torch.max(probs, dim=-1)
            label = self.intent_labels[idx.item()]

            if confidence.item() < threshold:
                label = "question_general"

            return {
                "intent": label,
                "confidence": confidence.item()
            }


# ------------------------
# RULE-BASED OVERRIDE
# ------------------------

def override_intent(text: str, predicted: str) -> str:
    text = text.lower().strip()

    if "who are you" in text:
        return "self_identity"

    if text.startswith(("hi", "hello", "hey")):
        return "greeting"

    return predicted