embeddings.py

from datasets import load_dataset
from transformers import AutoTokenizer, AutoModel
import faiss
import torch
import numpy as np
import os

def log(message):
    print(f"✅ {message}")

# ✅ Load datasets
datasets = {
    "sales": load_dataset("goendalf666/sales-conversations"),
    "blended": load_dataset("blended_skill_talk"),
    "dialog": load_dataset("daily_dialog"),
    "multiwoz": load_dataset("multi_woz_v22"),
}

# ✅ Load MiniLM model for embeddings
model_name = "sentence-transformers/all-MiniLM-L6-v2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name)

def embed_text(texts):
    """Generate embeddings for a batch of texts."""
    inputs = tokenizer(texts, padding=True, truncation=True, return_tensors="pt", max_length=512)
    with torch.no_grad():
        embeddings = model(**inputs).last_hidden_state.mean(dim=1).cpu().numpy()
    return embeddings

# ✅ Batch processing function
def create_embeddings(dataset_name, dataset, batch_size=100):
    print(f"📥 Creating embeddings for {dataset_name}...")
    
    if dataset_name == "goendalf666/sales-conversations":
        texts = [" ".join(row.values()) for row in dataset["train"]]
    elif dataset_name == "blended_skill_talk":
        texts = [" ".join(row["free_messages"] + row["guided_messages"]) for row in dataset["train"]]
    elif dataset_name == "daily_dialog":
        texts = [" ".join(row["dialog"]) for row in dataset["train"]]
    elif dataset_name == "multi_woz_v22":
        texts = [" ".join(row["turns"]["utterance"]) for row in dataset["train"]]
    else:
        print(f"⚠️ Warning: Dataset {dataset_name} not handled properly!")
        texts = []

    log(f"✅ Extracted {len(texts)} texts from {dataset_name}.")

    # Process in batches
    all_embeddings = []
    for i in range(0, len(texts), batch_size):
        batch = texts[i : i + batch_size]
        batch_embeddings = embed_text(batch)
        all_embeddings.append(batch_embeddings)
        
        # ✅ Log progress
        log(f"🚀 Processed {i + len(batch)}/{len(texts)} embeddings for {dataset_name}...")

    # Convert list of numpy arrays to a single numpy array
    all_embeddings = np.vstack(all_embeddings)
    return all_embeddings

# ✅ Save embeddings to FAISS with unique filename
def save_embeddings_to_faiss(embeddings, index_name="my_embeddings"):
    index_file = f"{index_name}.faiss"
    
    # ✅ Check if previous FAISS index exists, append if needed
    if os.path.exists(index_file):
        log("🔄 Loading existing FAISS index to append...")
        index = faiss.read_index(index_file)
        index.add(np.array(embeddings).astype(np.float32))
    else:
        index = faiss.IndexFlatL2(embeddings.shape[1])
        index.add(np.array(embeddings).astype(np.float32))

    faiss.write_index(index, index_file)  # ✅ Save FAISS index
    log(f"✅ Saved FAISS index: {index_file}")

# ✅ Run embeddings process
for name, dataset in datasets.items():
    embeddings = create_embeddings(name, dataset, batch_size=100)
    save_embeddings_to_faiss(embeddings, index_name=name)
    log(f"✅ Embeddings for {name} saved to FAISS.")