"""
Epstein Files Dataset Loader

Loads data from two HuggingFace sources:
  1. teyler/epstein-files-20k — raw OCR text (2.1M rows, filename + text)
  2. devankit7873/EpsteinFiles-Vector-Embeddings-ChromaDB — pre-computed
     all-MiniLM-L6-v2 embeddings in ChromaDB format

Both can feed directly into the ContextualSimilarityEngine pipeline.
"""

import logging
import re
import time
from pathlib import Path
from typing import Optional

import numpy as np

logger = logging.getLogger(__name__)

# HuggingFace dataset identifiers
RAW_DATASET = "teyler/epstein-files-20k"
EMBEDDINGS_DATASET = "devankit7873/EpsteinFiles-Vector-Embeddings-ChromaDB"


def load_raw_dataset(
    max_docs: Optional[int] = None,
    min_text_length: int = 100,
    source_filter: Optional[str] = None,
) -> list[dict]:
    """
    Load raw Epstein Files from HuggingFace.

    Args:
        max_docs: Limit number of documents loaded (None = all ~2.1M).
        min_text_length: Skip documents shorter than this.
        source_filter: Filter by filename prefix, e.g. "TEXT-" or "IMAGES-".

    Returns:
        List of {"doc_id": str, "text": str, "filename": str}
    """
    from datasets import load_dataset

    t0 = time.time()
    logger.info(f"Loading {RAW_DATASET} from HuggingFace...")

    ds = load_dataset(RAW_DATASET, split="train")
    docs = []

    for i, row in enumerate(ds):
        if max_docs and len(docs) >= max_docs:
            break

        text = (row.get("text") or "").strip()
        filename = row.get("filename") or f"doc_{i}"

        if len(text) < min_text_length:
            continue

        if source_filter and not filename.startswith(source_filter):
            continue

        doc_id = Path(filename).stem
        docs.append({"doc_id": doc_id, "text": text, "filename": filename})

    elapsed = time.time() - t0
    logger.info(f"Loaded {len(docs)} documents in {elapsed:.1f}s")
    return docs


def load_raw_to_engine(
    engine,
    max_docs: Optional[int] = 500,
    min_text_length: int = 100,
    source_filter: Optional[str] = None,
    build_index: bool = True,
) -> dict:
    """
    Load raw dataset directly into a ContextualSimilarityEngine.

    Args:
        engine: ContextualSimilarityEngine instance (must be initialized).
        max_docs: Limit documents to load.
        min_text_length: Skip short documents.
        source_filter: Filter by filename prefix.
        build_index: Whether to build FAISS index after loading.

    Returns:
        Stats dict with counts and timing.
    """
    t0 = time.time()
    docs = load_raw_dataset(max_docs, min_text_length, source_filter)

    total_chunks = 0
    skipped = 0
    for doc in docs:
        try:
            chunks = engine.add_document(doc["doc_id"], doc["text"])
            total_chunks += len(chunks)
        except ValueError as e:
            logger.warning("Skipped document '%s': %s", doc["doc_id"], e)
            skipped += 1

    if build_index and total_chunks > 0:
        engine.build_index(show_progress=True)

    elapsed = time.time() - t0
    return {
        "documents_loaded": len(docs) - skipped,
        "documents_skipped": skipped,
        "total_chunks": total_chunks,
        "index_built": build_index and total_chunks > 0,
        "seconds": round(elapsed, 2),
    }


def load_chromadb_embeddings(
    download_dir: str = "./chroma_epstein",
) -> dict:
    """
    Download and load the pre-computed ChromaDB embeddings.

    Returns:
        Dict with "texts", "embeddings", "metadatas", "ids", and stats.
    """
    import chromadb
    from huggingface_hub import snapshot_download

    t0 = time.time()
    logger.info(f"Downloading {EMBEDDINGS_DATASET} from HuggingFace...")

    # This repo contains ChromaDB persistence files (not standard datasets),
    # so we use snapshot_download instead of load_dataset.
    local_path = snapshot_download(
        repo_id=EMBEDDINGS_DATASET,
        repo_type="dataset",
        local_dir=download_dir,
    )

    # Find the chroma_db directory
    chroma_dir = None
    for candidate in [
        Path(local_path) / "chroma_db",
        Path(local_path),
    ]:
        if (candidate / "chroma.sqlite3").exists():
            chroma_dir = str(candidate)
            break

    if not chroma_dir:
        raise FileNotFoundError(
            f"ChromaDB files not found in {local_path}. "
            f"Expected chroma.sqlite3 in the download."
        )

    # Open ChromaDB
    client = chromadb.PersistentClient(path=chroma_dir)
    collections = client.list_collections()
    if not collections:
        raise ValueError("No collections found in ChromaDB.")

    collection = collections[0]
    count = collection.count()
    logger.info(f"ChromaDB collection '{collection.name}': {count} vectors")

    elapsed = time.time() - t0
    return {
        "chroma_dir": chroma_dir,
        "collection_name": collection.name,
        "total_vectors": count,
        "seconds": round(elapsed, 2),
        "_collection": collection,
        "_client": client,
    }


def import_chromadb_to_engine(
    engine,
    max_chunks: Optional[int] = None,
    batch_size: int = 1000,
) -> dict:
    """
    Import pre-computed ChromaDB embeddings into the engine's FAISS index.

    Since both use all-MiniLM-L6-v2 (384-dim), we can directly import
    the vectors without re-encoding.

    Args:
        engine: ContextualSimilarityEngine (must be initialized with all-MiniLM-L6-v2).
        max_chunks: Limit vectors to import (None = all).
        batch_size: How many vectors to fetch from ChromaDB at a time.

    Returns:
        Stats dict.
    """
    t0 = time.time()
    chroma_data = load_chromadb_embeddings()
    collection = chroma_data["_collection"]
    total = chroma_data["total_vectors"]

    if max_chunks:
        total = min(total, max_chunks)

    # Fetch in batches
    all_texts = []
    all_embeddings = []
    all_sources = []

    offset = 0
    while offset < total:
        limit = min(batch_size, total - offset)
        results = collection.get(
            limit=limit,
            offset=offset,
            include=["embeddings", "documents", "metadatas"],
        )

        if not results["ids"]:
            break

        for i, doc_id in enumerate(results["ids"]):
            text = results["documents"][i] if results["documents"] is not None else ""
            embedding = results["embeddings"][i] if results["embeddings"] is not None else None
            metadata = results["metadatas"][i] if results["metadatas"] is not None else {}
            source = metadata.get("source", f"chunk_{offset + i}")

            if text and embedding is not None:
                all_texts.append(text)
                all_embeddings.append(embedding)
                all_sources.append(source)

        offset += len(results["ids"])
        logger.info(f"Fetched {offset}/{total} vectors from ChromaDB")

    # Group texts by source document and add to engine
    doc_chunks = {}
    for text, source in zip(all_texts, all_sources):
        stem = Path(source).stem if source else "unknown"
        if stem not in doc_chunks:
            doc_chunks[stem] = []
        doc_chunks[stem].append(text)

    docs_added = 0
    chunks_added = 0
    for doc_id, texts in doc_chunks.items():
        combined = "\n\n".join(texts)
        try:
            chunks = engine.add_document(doc_id, combined)
            chunks_added += len(chunks)
            docs_added += 1
        except ValueError as e:
            logger.warning("Skipped ChromaDB document '%s': %s", doc_id, e)

    if chunks_added > 0:
        engine.build_index(show_progress=True)

    elapsed = time.time() - t0
    return {
        "source": "chromadb_embeddings",
        "chromadb_vectors": len(all_embeddings),
        "documents_created": docs_added,
        "chunks_indexed": chunks_added,
        "index_built": chunks_added > 0,
        "seconds": round(elapsed, 2),
    }


def get_dataset_info() -> dict:
    """Return metadata about available datasets (no download)."""
    return {
        "raw_texts": {
            "dataset_id": RAW_DATASET,
            "url": f"https://huggingface.co/datasets/{RAW_DATASET}",
            "description": "2.1M OCR text documents from U.S. House Oversight Committee Epstein Files release",
            "columns": ["filename", "text"],
            "size_mb": 106,
        },
        "embeddings": {
            "dataset_id": EMBEDDINGS_DATASET,
            "url": f"https://huggingface.co/datasets/{EMBEDDINGS_DATASET}",
            "description": "Pre-computed all-MiniLM-L6-v2 embeddings in ChromaDB format (~100K+ chunks)",
            "model": "all-MiniLM-L6-v2",
            "vector_dim": 384,
        },
    }