"""Reconstruct full conversations from this dataset by joining back to the
original source datasets on their own ids. This dataset ships NO text from
either source — only `prompt_id` (a HuggingFaceH4/ultrachat_200k row id) and
`persona_uuid` (an nvidia/Nemotron-Personas-USA row id). Run this script to
produce a local, enriched copy with the opening prompt and persona
description filled back in.

Requires: pandas, pyarrow, requests. Downloads ~730MB (ultrachat_200k
train_sft) + ~244MB (one Nemotron-Personas-USA shard) on first run.

Usage:
    python3 enrich.py [--in train.parquet] [--out enriched.parquet]
"""
import argparse
import os
import urllib.request

import pandas as pd

CACHE_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), ".enrich_cache")

ULTRACHAT_SHARDS = [
    "https://huggingface.co/api/datasets/HuggingFaceH4/ultrachat_200k/parquet/default/train_sft/0.parquet",
    "https://huggingface.co/api/datasets/HuggingFaceH4/ultrachat_200k/parquet/default/train_sft/1.parquet",
    "https://huggingface.co/api/datasets/HuggingFaceH4/ultrachat_200k/parquet/default/train_sft/2.parquet",
]
# Only shard 0 of Nemotron-Personas-USA was ever sampled from when this
# dataset was built, so that's the only shard needed to resolve every
# persona_uuid present here.
PERSONA_SHARD = "https://huggingface.co/api/datasets/nvidia/Nemotron-Personas-USA/parquet/default/train/0.parquet"


def download(url, dest):
    if os.path.exists(dest):
        return
    os.makedirs(os.path.dirname(dest), exist_ok=True)
    print(f"downloading {url} -> {dest}")
    urllib.request.urlretrieve(url, dest)


def load_ultrachat_prompts():
    frames = []
    for i, url in enumerate(ULTRACHAT_SHARDS):
        dest = os.path.join(CACHE_DIR, f"ultrachat_{i}.parquet")
        download(url, dest)
        frames.append(pd.read_parquet(dest, columns=["prompt", "prompt_id"]))
    df = pd.concat(frames, ignore_index=True)
    return dict(zip(df["prompt_id"], df["prompt"]))


def load_personas():
    dest = os.path.join(CACHE_DIR, "personas_0.parquet")
    download(PERSONA_SHARD, dest)
    df = pd.read_parquet(
        dest, columns=["uuid", "persona", "age", "sex", "occupation", "city", "state"]
    )
    out = {}
    for row in df.itertuples(index=False):
        out[row.uuid] = {
            "persona": row.persona,
            "age": int(row.age),
            "sex": row.sex,
            "occupation": row.occupation,
            "city": row.city,
            "state": row.state,
        }
    return out


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--in", dest="in_path", default="train.parquet")
    ap.add_argument("--out", dest="out_path", default="enriched.parquet")
    args = ap.parse_args()

    df = pd.read_parquet(args.in_path)
    print(f"loaded {len(df)} rows from {args.in_path}")

    prompts = load_ultrachat_prompts()
    personas = load_personas()

    missing_prompts = 0
    missing_personas = 0

    def enrich_row(row):
        nonlocal missing_prompts, missing_personas
        msgs = list(row["messages"])
        prompt_text = prompts.get(row["prompt_id"])
        if prompt_text is None:
            missing_prompts += 1
        msgs[0] = {"role": "user", "content": prompt_text}
        row["messages"] = msgs
        p = personas.get(row["persona_uuid"])
        if p is None:
            missing_personas += 1
            row["persona"] = None
        else:
            row["persona"] = p["persona"]
            row["persona_age"] = p["age"]
            row["persona_sex"] = p["sex"]
            row["persona_occupation"] = p["occupation"]
            row["persona_city"] = p["city"]
            row["persona_state"] = p["state"]
        return row

    df = df.apply(enrich_row, axis=1)

    if missing_prompts or missing_personas:
        print(f"warning: {missing_prompts} prompt_ids and {missing_personas} "
              f"persona_uuids had no match in the source datasets")

    df.to_parquet(args.out_path)
    print(f"wrote {len(df)} enriched rows to {args.out_path}")


if __name__ == "__main__":
    main()