loader.py

from __future__ import annotations

import gc
import logging
import shutil
from pathlib import Path
from typing import Dict, List, Union

import faiss
import numpy as np
import torch
from datasets import Dataset
from faiss.contrib.ondisk import merge_ondisk
from huggingface_hub import snapshot_download
from sentence_transformers import SentenceTransformer
from tqdm import auto as tqdm

logger = logging.getLogger(__name__)


class OnDiskIVFHelper:
    def __init__(self, path: Union[str, Path]):
        self.path = path

    @property
    def path(self):
        return self._path

    @path.setter
    def path(self, value: Union[str, Path]):
        value = Path(value)
        value.mkdir(parents=True, exist_ok=True)
        self._path = value

    def train(self, xt: np.ndarray):
        index = faiss.index_factory(xt.shape[1], "IVF4096,Flat")
        logger.info("Training index...")
        index.train(xt)
        train_index_path = str(self.path / "faiss.index")
        logger.info(f"Write {train_index_path}")
        faiss.write_index(index, train_index_path)

    def add_with_ids(self, xb: np.ndarray, ix: np.ndarray | int):
        if isinstance(ix, int):
            ix = np.arange(ix, ix + xb.shape[0])
        block_num = ix[0]
        train_index_path = str(self.path / "faiss.index")
        index = faiss.read_index(train_index_path)
        logger.info("Adding vectors to index...")
        index.add_with_ids(xb, ix)
        block_index_path = str(self.path / f"block_{block_num}.index")
        logger.info(f"Write {block_index_path}")
        faiss.write_index(index, block_index_path)

    def merge(self):
        logger.info("Loading trained index")
        train_index_path = str(self.path / "faiss.index")
        index = faiss.read_index(train_index_path)
        block_fnames = [str(p) for p in self.path.glob("block_*.index")]
        if len(block_fnames) == 0:
            return
        merged_index_path = str(self.path / "merged_index.ivfdata")
        merge_ondisk(index, block_fnames, merged_index_path)
        populated_index_path = str(self.path / "populated.index")
        logger.info(f"Write {populated_index_path}")
        faiss.write_index(index, populated_index_path)

    def search(self, xq: np.ndarray, top_k: int = 5, nprobe: int = 16):
        populated_index_path = str(self.path / "populated.index")
        logger.info("Read " + populated_index_path)
        index = faiss.read_index(populated_index_path)
        index.nprobe = nprobe
        D, I = index.search(xq, top_k)  # noqa: E741
        return D, I

    def delete(self):
        shutil.rmtree(self.path)

    @property
    def ntotal(self) -> int:
        populated_index_path = str(self.path / "populated.index")
        logger.info("Read " + populated_index_path)
        index = faiss.read_index(populated_index_path)
        return index.ntotal

    def __len__(self) -> int:
        return self.ntotal

    @property
    def is_trained(self) -> bool:
        index_file_path = self.path / "populated.index"
        if not index_file_path.exists():
            return False
        index_file_path = str(index_file_path)
        index = faiss.read_index(index_file_path)
        return index.is_trained


class DatasetIndex:
    def __init__(self, path: Union[str, Path]):
        self._context_embedding_model = None
        self._query_embedding_model = None
        self.context_model = "facebook-dpr-ctx_encoder-multiset-base"
        self.query_model = "facebook-dpr-question_encoder-multiset-base"
        # Ensure we're using CPU
        if (
            hasattr(torch.backends, "mps")
            and torch.backends.mps.is_available()
        ):
            self.device = torch.device("mps")
        elif torch.cuda.is_available():
            self.device = torch.device("cuda")
        else:
            self.device = torch.device("cpu")
        self.index = OnDiskIVFHelper(path)

    @property
    def context_embedding_model(self) -> SentenceTransformer:
        if self._context_embedding_model is None:
            self._query_embedding_model = None
            self._context_embedding_model = SentenceTransformer(
                self.context_model, device=self.device
            )
        return self._context_embedding_model

    @property
    def query_embedding_model(self) -> SentenceTransformer:
        if self._query_embedding_model is None:
            self._context_embedding_model = None
            self._query_embedding_model = SentenceTransformer(
                self.query_model, device=self.device
            )
        return self._query_embedding_model

    @property
    def ntotal(self) -> int:
        try:
            ntotal = len(self.index)
        except Exception:
            try:
                self.index.merge()
                ntotal = len(self.index)
            except Exception:
                ntotal = 0
        return ntotal

    def create_index(
        self,
        dataset: Dataset | List[Dict] | List[str],
        /,
        *,
        column: str | None = None,
        batch_size: int = 32,
        block_size: int = 159744,
        force: bool = False,
    ):
        """Build FAISS index for the given documents"""
        n_documents = len(dataset)
        n_index = self.ntotal
        if not force and n_index == n_documents:
            logging.info("Found existing index. Skipping...")
            return
        self.index.delete()
        logging.info("Training index/Adding vectors to index...")
        # Add vectors to index
        for block_start in tqdm.trange(
            0, n_documents, block_size, desc="Blocks"
        ):
            # logging.info(f"Starting encoding of {n_documents} documents...")
            # Train index
            embeddings = self.context_embedding_model.encode(
                (
                    dataset[block_start : block_start + block_size]
                    if column is None
                    else dataset[block_start : block_start + block_size][
                        column
                    ]
                ),
                batch_size=batch_size,
                convert_to_numpy=True,
                show_progress_bar=True,
                device=self.device,
            )
            batch_end = min(block_start + block_size, n_documents)
            assert len(embeddings) == batch_end - block_start
            if block_start == 0:
                self.index.train(embeddings)
            self.index.add_with_ids(embeddings, block_start)
        self.index.merge()
        print("Number of indexed documents:", len(self.index))
        # Clear memory
        del embeddings
        gc.collect()
        if self.device.type == "cuda":
            torch.cuda.empty_cache()
        logging.info("Indexing complete!")

    def search(self, query: str, k: int = 5) -> List[Dict]:
        """Search the index for similar documents"""
        if self.index is None:
            raise ValueError("Index not built yet!")
        # Encode query
        if isinstance(query, str):
            query = [query]
        query_vector = self.query_embedding_model.encode(
            query, convert_to_numpy=True
        )
        # Search
        D, I = self.index.search(query_vector, k)  # noqa: E741
        # Format results
        results = []
        for i in range(len(query)):
            results.append([
                (idx, score) for idx, score in zip(I[i], D[i]) if idx >= 0
            ])
        return results

    @classmethod
    def from_pretrained(cls, model_id: str) -> DatasetIndex:
        """Download and load a pre-trained index"""
        snapshot_path = snapshot_download(
            repo_id=model_id,
            repo_type="dataset",
            allow_patterns="index/*",
        )
        snapshot_path = Path(snapshot_path)
        self = cls(snapshot_path / "index")
        # make sure the index is merged
        _ = self.ntotal
        return self