memory.py

"""
Vector Memory Module
Memória de longo prazo usando ChromaDB + Sentence Transformers
"""

import chromadb
from sentence_transformers import SentenceTransformer
from datetime import datetime
import hashlib

# Usar modelo leve para embeddings
EMBEDDING_MODEL = "all-MiniLM-L6-v2"  # ~80MB, rápido

# Singleton para evitar recarregar
_memory_instance = None


def get_memory():
    """Get or create memory instance."""
    global _memory_instance
    if _memory_instance is None:
        _memory_instance = VectorMemory()
    return _memory_instance


class VectorMemory:
    def __init__(self):
        print("Inicializando memória vetorial...")
        
        # Modelo de embeddings
        self.model = SentenceTransformer(EMBEDDING_MODEL)
        print(f"Modelo carregado: {EMBEDDING_MODEL}")
        
        # ChromaDB em memória (persiste enquanto o servidor estiver rodando)
        self.client = chromadb.Client()
        self.collection = self.client.get_or_create_collection(
            name="chat_memory",
            metadata={"hnsw:space": "cosine"}
        )
        
        print(f"Memória pronta. {self.collection.count()} memórias existentes.")
    
    def _generate_id(self, text: str) -> str:
        """Generate unique ID for a memory."""
        timestamp = datetime.now().isoformat()
        content = f"{timestamp}:{text}"
        return hashlib.md5(content.encode()).hexdigest()
    
    def add_memory(self, user_message: str, bot_response: str):
        """
        Add a conversation exchange to memory.
        Stores the combined context for better retrieval.
        """
        # Combinar mensagem e resposta para contexto completo
        combined = f"Usuário: {user_message}\nAssistente: {bot_response}"
        
        # Gerar embedding
        embedding = self.model.encode(combined).tolist()
        
        # Gerar ID único
        doc_id = self._generate_id(combined)
        
        # Metadados
        metadata = {
            "user_message": user_message[:500],  # Truncar se muito longo
            "bot_response": bot_response[:500],
            "timestamp": datetime.now().isoformat()
        }
        
        # Adicionar ao banco
        self.collection.add(
            ids=[doc_id],
            embeddings=[embedding],
            documents=[combined],
            metadatas=[metadata]
        )
        
        print(f"Memória adicionada. Total: {self.collection.count()}")
    
    def search_memories(self, query: str, k: int = 3) -> list[dict]:
        """
        Search for relevant memories based on the query.
        Returns list of {text, user_message, bot_response, score}
        """
        if self.collection.count() == 0:
            return []
        
        # Gerar embedding da query
        query_embedding = self.model.encode(query).tolist()
        
        # Buscar similares
        results = self.collection.query(
            query_embeddings=[query_embedding],
            n_results=min(k, self.collection.count())
        )
        
        memories = []
        if results and results['documents'] and results['documents'][0]:
            for i, doc in enumerate(results['documents'][0]):
                metadata = results['metadatas'][0][i] if results['metadatas'] else {}
                distance = results['distances'][0][i] if results['distances'] else 0
                
                memories.append({
                    "text": doc,
                    "user_message": metadata.get("user_message", ""),
                    "bot_response": metadata.get("bot_response", ""),
                    "score": 1 - distance,  # Converter distância em similaridade
                    "timestamp": metadata.get("timestamp", "")
                })
        
        return memories
    
    def clear_memories(self):
        """Clear all memories."""
        # Recriar collection
        self.client.delete_collection("chat_memory")
        self.collection = self.client.get_or_create_collection(
            name="chat_memory",
            metadata={"hnsw:space": "cosine"}
        )
        print("Memórias limpas.")
    
    def get_stats(self) -> dict:
        """Get memory statistics."""
        return {
            "total_memories": self.collection.count(),
            "model": EMBEDDING_MODEL
        }