tf_data_pipeline.py

import os
import numpy as np
import faiss
import tensorflow as tf
import h5py
import math
import random
import gc
from tqdm import tqdm
import json
from pathlib import Path
from typing import Union, Optional, Dict, List, Tuple, Generator
from transformers import AutoTokenizer
from typing import List, Tuple, Generator
from transformers import AutoTokenizer
import random

from logger_config import config_logger
logger = config_logger(__name__)

class TFDataPipeline:
    def __init__(
        self,
        config,
        tokenizer,
        encoder,
        response_pool: List[str],
        query_embeddings_cache: dict,
        max_length: int = 512,
        neg_samples: int = 10,
        index_type: str = 'IndexFlatIP',
        faiss_index_file_path: str = 'new_iteration/data_prep_iterative_models/faiss_indices/faiss_index_production.index',
        nlist: int = 100,
        max_retries: int = 3
    ):
        self.config = config
        self.tokenizer = tokenizer
        self.encoder = encoder
        self.faiss_index_file_path = faiss_index_file_path
        self.response_pool = response_pool
        self.max_length = max_length
        self.neg_samples = neg_samples
        self.query_embeddings_cache = query_embeddings_cache # In-memory cache for embeddings
        self.index_type = index_type
        self.nlist = nlist
        self.embedding_batch_size = 16 if len(response_pool) < 100 else 64
        self.search_batch_size = 16 if len(response_pool) < 100 else 64
        self.max_batch_size = 16 if len(response_pool) < 100 else 64
        self.max_retries = max_retries
        
        # Build text -> domain map for O(1) domain lookups (hard negative sampling)
        self._text_domain_map = {}
        self.build_text_to_domain_map()
        
        # Initialize FAISS index
        if os.path.exists(faiss_index_file_path):
            logger.info(f"Loading existing FAISS index from {faiss_index_file_path}...")
            self.index = faiss.read_index(faiss_index_file_path)
            self.validate_faiss_index()
            logger.info("FAISS index loaded and validated successfully.")
        else:
            dimension = self.encoder.config.embedding_dim
            self.index = faiss.IndexFlatIP(dimension)
            logger.info(f"Initialized FAISS IndexFlatIP with dimension {dimension}.")
            
        if not self.index.is_trained:
            # Train the index if it's not trained. IndexFlatIP doesn't need training, but others do (Future switch to IndexIVFFlat)
            dimension = self.query_embeddings_cache[next(iter(self.query_embeddings_cache))].shape[0]
            self.index.train(np.array(list(self.query_embeddings_cache.values())).astype(np.float32))
            self.index.add(np.array(list(self.query_embeddings_cache.values())).astype(np.float32))
    
    def save_embeddings_cache_hdf5(self, cache_file_path: str):
        """Save embeddings cache to HDF5 file."""
        with h5py.File(cache_file_path, 'w') as hf:
            for query, emb in self.query_embeddings_cache.items():
                hf.create_dataset(query, data=emb)
        logger.info(f"Embeddings cache saved to {cache_file_path}.")
    
    def load_embeddings_cache_hdf5(self, cache_file_path: str):
        """Load embeddings cache from HDF5 file."""
        with h5py.File(cache_file_path, 'r') as hf:
            for query in hf.keys():
                self.query_embeddings_cache[query] = hf[query][:]
        logger.info(f"Embeddings cache loaded from {cache_file_path}.")
    
    def save_faiss_index(self, faiss_index_file_path: str):
        faiss.write_index(self.index, faiss_index_file_path)
        logger.info(f"FAISS index saved to {faiss_index_file_path}")
    
    def load_faiss_index(self, faiss_index_file_path: str):
        """Load FAISS index from specified file path."""
        if os.path.exists(faiss_index_file_path):
            self.index = faiss.read_index(faiss_index_file_path)
            logger.info(f"FAISS index loaded from {faiss_index_file_path}.")
        else:
            logger.error(f"FAISS index file not found at {faiss_index_file_path}.")
            raise FileNotFoundError(f"FAISS index file not found at {faiss_index_file_path}.")
    
    def validate_faiss_index(self):
        """Validates FAISS index dimensionality."""
        expected_dim = self.encoder.config.embedding_dim
        if self.index.d != expected_dim:
            logger.error(f"FAISS index dimension {self.index.d} does not match encoder embedding dimension {expected_dim}.")
            raise ValueError("FAISS index dimensionality mismatch.")
        logger.info("FAISS index dimension validated successfully.")
    
    def save_tokenizer(self, tokenizer_dir: str):
        self.tokenizer.save_pretrained(tokenizer_dir)
        logger.info(f"Tokenizer saved to {tokenizer_dir}")
    
    def load_tokenizer(self, tokenizer_dir: str):
        self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_dir)
        logger.info(f"Tokenizer loaded from {tokenizer_dir}")
    
    @staticmethod
    def load_json_training_data(data_path: Union[str, Path], debug_samples: Optional[int] = None) -> List[dict]:
        """
        Load training data from a JSON file.
        Args:
            data_path (Union[str, Path]): Path to the JSON file containing dialogues.
            debug_samples (Optional[int]): Number of samples to load for debugging.
        
        Returns:
            List[dict]: List of dialogue dictionaries.
        """
        logger.info(f"Loading training data from {data_path}...")
        data_path = Path(data_path)
        if not data_path.exists():
            logger.error(f"Data file {data_path} does not exist.")
            return []
        
        with open(data_path, 'r', encoding='utf-8') as f:
            dialogues = json.load(f)
        
        if debug_samples is not None:
            dialogues = dialogues[:debug_samples]
            logger.info(f"Debug mode: Limited to {debug_samples} dialogues")
        
        logger.info(f"Loaded {len(dialogues)} dialogues.")
        return dialogues
    
    def collect_responses_with_domain(self, dialogues: List[dict]) -> List[Dict[str, str]]:
        """
        Extract unique assistant responses and their domains from dialogues.
        Returns List[Dict[str: "domain", str: text"]]
        """
        response_set = set()  # Store (domain, text) unique tuples
        results = []
        
        for dialogue in tqdm(dialogues, desc="Processing Dialogues", unit="dialogue"):
            domain = dialogue.get('domain', 'other')
            turns = dialogue.get('turns', [])
            for turn in turns:
                speaker = turn.get('speaker')
                text = turn.get('text', '').strip()
                if speaker == 'assistant' and text:
                    if len(text) <= self.max_length:
                        # Use tuple as set key to ensure uniqueness
                        key = (domain, text)
                        if key not in response_set:
                            response_set.add(key)
                            results.append({
                                "domain": domain,
                                "text": text
                            })
        
        logger.info(f"Collected {len(results)} unique assistant responses from dialogues.")
        return results
    
    def _extract_pairs_from_dialogue(self, dialogue: dict) -> List[Tuple[str, str]]:
        """Extract query-response pairs from a dialogue."""
        pairs = []
        turns = dialogue.get('turns', [])
        
        for i in range(len(turns) - 1):
            current_turn = turns[i]
            next_turn = turns[i+1]
            
            if (current_turn.get('speaker') == 'user' and
                next_turn.get('speaker') == 'assistant' and
                'text' in current_turn and 
                'text' in next_turn):
                
                query = current_turn['text'].strip()
                positive = next_turn['text'].strip()
                pairs.append((query, positive))
                
        return pairs
    
    def compute_and_index_response_embeddings(self):
        """
        Compute embeddings for the response pool and add them to the FAISS index.
        self.response_pool: List[Dict[str, str]] with keys "domain" and "text".
        """
        logger.info("Computing embeddings for the response pool...")
        
        # Extract the assistant text
        texts = [resp["text"] for resp in self.response_pool]
        logger.debug(f"Total texts to embed: {len(texts)}")
        
        batch_size = getattr(self, 'embedding_batch_size', 64)
        embeddings = []
        
        with tqdm(total=len(texts), desc="Computing Embeddings", unit="response") as pbar:
            for i in range(0, len(texts), batch_size):
                batch_texts = texts[i:i+batch_size]
                encodings = self.tokenizer(
                    batch_texts,
                    padding=True,
                    truncation=True,
                    max_length=self.max_length,
                    return_tensors='tf'
                )
                batch_embeds = self.encoder(encodings['input_ids'], training=False).numpy()
                
                embeddings.append(batch_embeds)
                pbar.update(len(batch_texts))
        
        # Combine embeddings and add to FAISS
        all_embeddings = np.vstack(embeddings).astype(np.float32)
        logger.info(f"Adding {len(all_embeddings)} response embeddings to FAISS index...")
        self.index.add(all_embeddings)
        
        # Store in memory
        self.response_embeddings = all_embeddings
        logger.info(f"FAISS index now has {self.index.ntotal} vectors.")
    
    def _find_hard_negatives(self, queries: List[str], positives: List[str], batch_size: int = 128) -> List[List[str]]:
        """
        Find hard negatives for a batch of queries using FAISS search.
        Fallback: in-domain negatives, then random negatives when needed.
        """
        retry_count = 0
        total_responses = len(self.response_pool)
        
        while retry_count < self.max_retries:
            try:
                # Build query embeddings from the cache
                query_embeddings = []
                for i in range(0, len(queries), batch_size):
                    sub_queries = queries[i : i + batch_size]
                    sub_embeds = [self.query_embeddings_cache[q] for q in sub_queries]
                    sub_embeds = np.vstack(sub_embeds).astype(np.float32)
                    faiss.normalize_L2(sub_embeds)  # If not already normalized
                    query_embeddings.append(sub_embeds)
                    
                query_embeddings = np.vstack(query_embeddings)
                query_embeddings = np.ascontiguousarray(query_embeddings)
                
                # FAISS search for nearest neighbors (hard negatives)
                distances, indices = self.index.search(query_embeddings, self.neg_samples)
                
                all_negatives = []
                # Extract domain from the positive assistant response
                for query_indices, query_text, pos_text in zip(indices, queries, positives):
                    negative_list = []
                    
                    # Build a 'seen' set with the positive
                    seen = {pos_text.strip()}
                    
                    domain_of_positive = self._detect_domain_for_text(pos_text)
                    
                    # Collect hard negatives (from config self.neg_samples)
                    for idx in query_indices:
                        if 0 <= idx < total_responses:
                            candidate_dict = self.response_pool[idx]  # e.g. {domain, text}
                            candidate_text = candidate_dict["text"].strip()
                            if candidate_text and candidate_text not in seen:
                                seen.add(candidate_text)
                                negative_list.append(candidate_text)
                                if len(negative_list) >= self.neg_samples:
                                    break
                    
                    # Fall back to random domain-based
                    if len(negative_list) < self.neg_samples:
                        needed = self.neg_samples - len(negative_list)
                        
                        random_negatives = self._get_random_negatives(needed, seen, domain=domain_of_positive)
                        negative_list.extend(random_negatives)
                        
                    all_negatives.append(negative_list)
                    
                return all_negatives
            
            except KeyError as ke:
                retry_count += 1
                logger.warning(f"Hard negative search attempt {retry_count} failed due to missing embeddings: {ke}")
                if retry_count == self.max_retries:
                    logger.error("Max retries reached for hard negative search due to missing embeddings.")
                    return self._fallback_negatives(queries, positives, reason="key_error")
                gc.collect()
                if tf.config.list_physical_devices('GPU'):
                    tf.keras.backend.clear_session()
            
            except Exception as e:
                retry_count += 1
                logger.warning(f"Hard negative search attempt {retry_count} failed: {e}")
                if retry_count == self.max_retries:
                    logger.error("Max retries reached for hard negative search.")
                    return self._fallback_negatives(queries, positives, reason="generic_error")
                gc.collect()
                if tf.config.list_physical_devices('GPU'):
                    tf.keras.backend.clear_session()
                    
    def _detect_domain_for_text(self, text: str) -> Optional[str]:
        """
        Domain detection for related negatives.
        """
        stripped_text = text.strip()
        return self._text_domain_map.get(stripped_text, None)
    
    def _get_random_negatives(self, needed: int, seen: set, domain: Optional[str] = None) -> List[str]:
        """
        Return a list of negative texts from the same domain. Fall back to any domain.
        """
        # Filter response_pool for domain
        if domain:
            domain_texts = [r["text"] for r in self.response_pool if r["domain"] == domain]
            # fallback to entire set if insufficient domain_texts
            if len(domain_texts) < needed * 2:
                domain_texts = [r["text"] for r in self.response_pool]
        else:
            domain_texts = [r["text"] for r in self.response_pool]
        
        negatives = []
        tries = 0
        max_tries = needed * 10
        while len(negatives) < needed and tries < max_tries:
            tries += 1
            candidate = random.choice(domain_texts).strip()
            if candidate and candidate not in seen:
                negatives.append(candidate)
                seen.add(candidate)
        
        if len(negatives) < needed:
            logger.warning(f"Could not find enough domain-based random negatives; needed {needed}, got {len(negatives)}.")
            
        return negatives
    
    def _fallback_negatives(self, queries: List[str], positives: List[str], reason: str) -> List[List[str]]:
        """
        Called if FAISS fails or embeddings are missing.
        We use entirely random negatives for each query, ignoring FAISS,
        but still attempt domain-based selection if possible.
        """
        logger.error(f"Falling back to random negatives due to: {reason}")
        all_negatives = []
        
        for pos_text in positives:
            # Build a 'seen' set with the positive assistant response
            seen = {pos_text.strip()}
            
            # Detect domain of the positive
            domain_of_positive = self._detect_domain_for_text(pos_text)
            
            # Use domain-based negatives when available
            negs = self._get_random_negatives(self.neg_samples, seen, domain=domain_of_positive)
            all_negatives.append(negs)
            
        return all_negatives
    
    def build_text_to_domain_map(self):
        """
        Build O(1) lookup dict: text -> domain for hard negative sampling.
        """
        self._text_domain_map = {}
        
        for item in self.response_pool:
            stripped_text = item["text"].strip()
            domain = item["domain"]
            
            if stripped_text in self._text_domain_map:
                #existing_domain = self._text_domain_map[stripped_text]
                #if existing_domain != domain:
                    # Collision detected. Using first found domain for now.
                    # This happens often with low-signal responses. "ok", "yes", etc.
                    # logger.warning(
                    #     f"Collision detected: text '{stripped_text}' found with domains "
                    #     f"'{existing_domain}' and '{domain}'. Keeping the first."
                    # )
                # By default, keep the first domain or overwrite. We'll skip overwriting:
                continue
            else:
                # Insert into the dict
                self._text_domain_map[stripped_text] = domain
        
        logger.info(f"Built text -> domain map with {len(self._text_domain_map)} unique text entries.")
        
    def encode_query(
        self, 
        query: str, 
        context: Optional[List[Tuple[str, str]]] = None
    ) -> np.ndarray:
        """
        Encode a user query (and optional conversation context) into an embedding vector.
        
        Args:
            query: The user query.
            context: Optional conversation history as a list of (user_text, assistant_text).
        Returns:
            np.ndarray of shape [embedding_dim], typically L2-normalized already.
        """
        # Prepare context: concat user/assistant pairs
        if context:
            # Take the last N turns
            relevant_history = context[-self.config.max_context_turns:]
            context_str_parts = []
            for (u_text, a_text) in relevant_history:
                context_str_parts.append(
                    f"{self.tokenizer.additional_special_tokens[self.tokenizer.additional_special_tokens.index('<USER>')]} {u_text} "
                    f"{self.tokenizer.additional_special_tokens[self.tokenizer.additional_special_tokens.index('<ASSISTANT>')]} {a_text}"
                )
            context_str = " ".join(context_str_parts)

            # Append the new query
            full_query = (
                f"{context_str} "
                f"{self.tokenizer.additional_special_tokens[self.tokenizer.additional_special_tokens.index('<USER>')]} {query}"
            )
        else:
            # Single user turn
            full_query = (
                f"{self.tokenizer.additional_special_tokens[self.tokenizer.additional_special_tokens.index('<USER>')]} {query}"
            )

        # Tokenize
        encodings = self.tokenizer(
            [full_query],
            padding='max_length',
            truncation=True,
            max_length=self.max_length,
            return_tensors='np'  # to keep it compatible with FAISS
        )
        input_ids = encodings['input_ids']
        
        # Debug out-of-vocab IDs
        max_id = np.max(input_ids)
        vocab_size = len(self.tokenizer)
        if max_id >= vocab_size:
            logger.error(f"Token ID {max_id} exceeds tokenizer vocab size {vocab_size}.")
            raise ValueError("Token ID exceeds vocabulary size.")
        
        # Get embeddings from the model. These are already L2-normalized by the model's final layer.
        embeddings = self.encoder(input_ids, training=False).numpy()
        
        return embeddings[0]
    
    def encode_responses(
        self, 
        responses: List[str], 
        context: Optional[List[Tuple[str, str]]] = None
    ) -> np.ndarray:
        """
        Encode multiple response texts into embedding vectors.
        Args:
            responses: List of assistant responses.
            context: Optional conversation context (last N turns).
        Returns:
            np.ndarray of shape [num_responses, embedding_dim].
        """
        # Incorporate context into response encoding. Note: Undecided on benefit of this
        if context:
            relevant_history = context[-self.config.max_context_turns:]
            prepared = []
            for resp in responses:
                context_str_parts = []
                for (u_text, a_text) in relevant_history:
                    context_str_parts.append(
                        f"{self.tokenizer.additional_special_tokens[self.tokenizer.additional_special_tokens.index('<USER>')]} {u_text} "
                        f"{self.tokenizer.additional_special_tokens[self.tokenizer.additional_special_tokens.index('<ASSISTANT>')]} {a_text}"
                    )
                context_str = " ".join(context_str_parts)
                
                # Treat resp as an assistant turn
                full_resp = (
                    f"{context_str} "
                    f"{self.tokenizer.additional_special_tokens[self.tokenizer.additional_special_tokens.index('<ASSISTANT>')]} {resp}"
                )
                prepared.append(full_resp)
        else:
            # Single response from the assistant
            prepared = [
                f"{self.tokenizer.additional_special_tokens[self.tokenizer.additional_special_tokens.index('<ASSISTANT>')]} {r}"
                for r in responses
            ]
        
        # Tokenize
        encodings = self.tokenizer(
            prepared,
            padding='max_length',
            truncation=True,
            max_length=self.max_length,
            return_tensors='np'
        )
        input_ids = encodings['input_ids']
        
        # Debug for out-of-vocab
        max_id = np.max(input_ids)
        vocab_size = len(self.tokenizer)
        if max_id >= vocab_size:
            logger.error(f"Token ID {max_id} exceeds tokenizer vocab size {vocab_size}.")
            raise ValueError("Token ID exceeds vocabulary size.")
        
        # Get embeddings from the model. These are already L2-normalized by the model's final layer.
        embeddings = self.encoder(input_ids, training=False).numpy()
        
        return embeddings.astype('float32')
    
    def prepare_and_save_data(self, dialogues: List[dict], tf_record_path: str, batch_size: int = 32):
        """
        Batch-Process dialogues and save to TFRecord file.
        """
        logger.info(f"Preparing and saving data to {tf_record_path}...")
        
        num_dialogues = len(dialogues)
        num_batches = math.ceil(num_dialogues / batch_size)
        
        with tf.io.TFRecordWriter(tf_record_path) as writer:
            with tqdm(total=num_batches, desc="Preparing Data Batches", unit="batch") as pbar:
                for i in range(num_batches):
                    start_idx = i * batch_size
                    end_idx = min(start_idx + batch_size, num_dialogues)
                    batch_dialogues = dialogues[start_idx:end_idx]
                    
                    # Extract query-positive pairs for the batch
                    queries = []
                    positives = []
                    for dialogue in batch_dialogues:
                        pairs = self._extract_pairs_from_dialogue(dialogue)
                        for query, positive in pairs:
                            if len(query) <= self.max_length and len(positive) <= self.max_length:
                                queries.append(query)
                                positives.append(positive)
                    
                    if not queries:
                        pbar.update(1)
                        continue
                    
                    # Compute and cache query embeddings
                    try:
                        self._compute_embeddings(queries)
                    except Exception as e:
                        logger.error(f"Error computing embeddings: {e}")
                        pbar.update(1)
                        continue
                    
                    # Find hard negatives
                    try:
                        hard_negatives = self._find_hard_negatives(queries, positives)
                    except Exception as e:
                        logger.error(f"Error finding hard negatives: {e}")
                        pbar.update(1)
                        continue  # Skip to the next batch
                    
                    # Tokenize and encode all queries, positives, and negatives in the batch
                    try:
                        encoded_queries = self.tokenizer.batch_encode_plus(
                            queries,
                            max_length=self.config.max_context_token_limit,
                            truncation=True,
                            padding='max_length',
                            return_tensors='tf'
                        )
                        encoded_positives = self.tokenizer.batch_encode_plus(
                            positives,
                            max_length=self.config.max_context_token_limit,
                            truncation=True,
                            padding='max_length',
                            return_tensors='tf'
                        )
                    except Exception as e:
                        logger.error(f"Error during tokenization: {e}")
                        pbar.update(1)
                        continue  # Skip to the next batch
                    
                    # Flatten hard_negatives. Maintain alignment.
                    # hard_negatives is List of Lists. Each sublist corresponds to a query.
                    try:
                        flattened_negatives = [neg for sublist in hard_negatives for neg in sublist]
                        encoded_negatives = self.tokenizer.batch_encode_plus(
                            flattened_negatives,
                            max_length=self.config.max_context_token_limit,
                            truncation=True,
                            padding='max_length',
                            return_tensors='tf'
                        )
                        
                        # Reshape to [num_queries, num_negatives, max_length]
                        num_negatives = self.config.neg_samples
                        reshaped_negatives = encoded_negatives['input_ids'].numpy().reshape(-1, num_negatives, self.config.max_context_token_limit)
                    except Exception as e:
                        logger.error(f"Error during negatives tokenization: {e}")
                        pbar.update(1)
                        continue
                    
                    # Serialize and write to TFRecord
                    for j in range(len(queries)):
                        try:
                            q_id = encoded_queries['input_ids'][j].numpy()
                            p_id = encoded_positives['input_ids'][j].numpy()
                            n_id = reshaped_negatives[j]
                            
                            feature = {
                                'query_ids': tf.train.Feature(int64_list=tf.train.Int64List(value=q_id)),
                                'positive_ids': tf.train.Feature(int64_list=tf.train.Int64List(value=p_id)),
                                'negative_ids': tf.train.Feature(int64_list=tf.train.Int64List(value=n_id.flatten())),
                            }
                            example = tf.train.Example(features=tf.train.Features(feature=feature))
                            writer.write(example.SerializeToString())
                        except Exception as e:
                            logger.error(f"Error serializing example {j} in batch {i}: {e}")
                            continue  # Skip to the next example
                    
                    # Update progress bar
                    pbar.update(1)
        
        logger.info(f"Data preparation complete. TFRecord saved.")
        
    def _compute_embeddings(self, queries: List[str]) -> None:
        """
        Compute embeddings for new queries and update the cache.
        """
        new_queries = [q for q in queries if q not in self.query_embeddings_cache]
        if not new_queries:
            return
        
        # Compute embeddings
        new_embeddings = []
        for i in range(0, len(new_queries), self.embedding_batch_size):
            batch_queries = new_queries[i:i + self.embedding_batch_size]
            encoded = self.tokenizer(
                batch_queries,
                padding=True,
                truncation=True,
                max_length=self.max_length,
                return_tensors='tf'
            )
            batch_embeddings = self.encoder(encoded['input_ids'], training=False).numpy()
            faiss.normalize_L2(batch_embeddings)
            new_embeddings.extend(batch_embeddings)
        
        # Update the cache
        for query, emb in zip(new_queries, new_embeddings):
            self.query_embeddings_cache[query] = emb
    
    def data_generator(self, dialogues: List[dict]) -> Generator[Tuple[str, str, List[str]], None, None]:
        """
        Generate training examples: (query, positive, [hard_negatives]).
        """
        total_dialogues = len(dialogues)
        logger.debug(f"Total dialogues to process: {total_dialogues}")
        
        with tqdm(total=total_dialogues, desc="Processing Dialogues", unit="dialogue") as pbar:
            for dialogue in dialogues:
                pairs = self._extract_pairs_from_dialogue(dialogue)
                for query, positive in pairs:
                    # Ensure embeddings are computed, find hard negatives, etc.
                    self._compute_embeddings([query])
                    hard_negatives = self._find_hard_negatives([query], [positive])[0]
                    yield (query, positive, hard_negatives)
                pbar.update(1)

    def get_tf_dataset(self, dialogues: List[dict], batch_size: int) -> tf.data.Dataset:
        """
        Creates a tf.data.Dataset for streaming training.
        yields (input_ids_query, input_ids_positive, input_ids_negatives).
        """
        # 1) Start with a generator dataset
        dataset = tf.data.Dataset.from_generator(
            lambda: self.data_generator(dialogues),
            output_signature=(
                tf.TensorSpec(shape=(), dtype=tf.string),                   # Query (single string)
                tf.TensorSpec(shape=(), dtype=tf.string),                   # Positive (single string)
                tf.TensorSpec(shape=(self.neg_samples,), dtype=tf.string)   # Hard Negatives (list of strings)
            )
        )
        
        # Batch the raw strings, then map through a tokenize step
        # Note 'Distilbert Tokenizer threw an error when using tf.data.AUTOTUNE.
        dataset = dataset.batch(batch_size, drop_remainder=True)
        dataset = dataset.map(
            lambda q, p, n: self._tokenize_triple(q, p, n),
            num_parallel_calls=1 #tf.data.AUTOTUNE
        )

        dataset = dataset.prefetch(tf.data.AUTOTUNE)
        return dataset
    
    def _tokenize_triple(
        self, 
        q: tf.Tensor, 
        p: tf.Tensor, 
        n: tf.Tensor
    ) -> Tuple[tf.Tensor, tf.Tensor, tf.Tensor]:
        """
        Wraps a Python function. Convert tf.Tensors of strings -> Python lists of strings -> HF tokenizer -> Tensors of IDs.
        q is shape [batch_size], p is shape [batch_size], n is shape [batch_size, neg_samples] (list of negatives).
        """
        # Use tf.py_function, limit parallelism
        q_ids, p_ids, n_ids = tf.py_function(
            func=self._tokenize_triple_py,
            inp=[q, p, n, tf.constant(self.max_length), tf.constant(self.neg_samples)],
            Tout=[tf.int32, tf.int32, tf.int32]
        )

        # Set shape info for the output tensors
        q_ids.set_shape([None, self.max_length])                    # [batch_size, max_length]
        p_ids.set_shape([None, self.max_length])                    # [batch_size, max_length]
        n_ids.set_shape([None, self.neg_samples, self.max_length])  # [batch_size, neg_samples, max_length]

        return q_ids, p_ids, n_ids

    def _tokenize_triple_py(
        self, 
        q: tf.Tensor, 
        p: tf.Tensor, 
        n: tf.Tensor,
        max_len: tf.Tensor,
        neg_samples: tf.Tensor
    ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
        """
        Decodes tf.string Tensor to Python List[str], then tokenize.
        Reshapes negatives to [batch_size, neg_samples, max_length].
        Returns np.array(int32) for (q_ids, p_ids, n_ids).
        
        q: shape [batch_size], p: shape [batch_size]
        n: shape [batch_size, neg_samples]
        max_len: int
        neg_samples: int
        """
        max_len = int(max_len.numpy())
        neg_samples = int(neg_samples.numpy())

        # Convert Tensors -> Python List[str]
        q_list = [q_i.decode("utf-8") for q_i in q.numpy()]  # shape [batch_size]
        p_list = [p_i.decode("utf-8") for p_i in p.numpy()]  # shape [batch_size]

        # Shape [batch_size, neg_samples], decode each row
        n_list = []
        for row in n.numpy():
            # row is shape [neg_samples], each is a tf.string
            decoded = [neg.decode("utf-8") for neg in row]
            n_list.append(decoded)

        # Tokenize queries & positives
        q_enc = self.tokenizer(
            q_list,
            padding="max_length",
            truncation=True,
            max_length=max_len,
            return_tensors="np"
        )
        p_enc = self.tokenizer(
            p_list,
            padding="max_length",
            truncation=True,
            max_length=max_len,
            return_tensors="np"
        )

        # Tokenize negatives
        # Flatten [batch_size, neg_samples] -> List
        flattened_negatives = [neg for row in n_list for neg in row]
        if len(flattened_negatives) == 0:
            # No negatives: return a zero array
            n_ids = np.zeros((len(q_list), neg_samples, max_len), dtype=np.int32)
        else:
            n_enc = self.tokenizer(
                flattened_negatives,
                padding="max_length",
                truncation=True,
                max_length=max_len,
                return_tensors="np"
            )
            # Shape [batch_size * neg_samples, max_len]
            n_input_ids = n_enc["input_ids"]

            # Reshape to [batch_size, neg_samples, max_len]
            batch_size = len(q_list)
            n_ids_list = []
            for i in range(batch_size):
                start_idx = i * neg_samples
                end_idx = start_idx + neg_samples
                row_negs = n_input_ids[start_idx:end_idx]

                # Pad with zeros if not enough negatives
                if row_negs.shape[0] < neg_samples:
                    deficit = neg_samples - row_negs.shape[0]
                    pad_arr = np.zeros((deficit, max_len), dtype=np.int32)
                    row_negs = np.concatenate([row_negs, pad_arr], axis=0)

                n_ids_list.append(row_negs)

            # Stack shape [batch_size, neg_samples, max_len]
            n_ids = np.stack(n_ids_list, axis=0)

        # Return np.int32 arrays
        q_ids = q_enc["input_ids"].astype(np.int32)  # shape [batch_size, max_len]
        p_ids = p_enc["input_ids"].astype(np.int32)  # shape [batch_size, max_len]
        n_ids = n_ids.astype(np.int32)               # shape [batch_size, neg_samples, max_len]

        return q_ids, p_ids, n_ids