#!/usr/bin/env python3
"""Pull lawyer feedback bundles from the private HF dataset and digest them.

Feedback Mode (the Vercel UI toggle) posts judgment bundles to /api/feedback_bundle,
which the Space pushes durably to the private dataset vg15o2/themis-feedback
(one JSON per share under feedback/YYYY-MM-DD/). This script is the owner side:
download everything, print a digest, dump a flat JSONL, and optionally emit
qrels-candidate rows for the eval set.

Grade map (placement -> graded relevance): top5=3, top10=2, after10=1, irrelevant=0.

Usage:
  python phase1/eval/fetch_feedback.py              # digest + feedback_dump.jsonl
  python phase1/eval/fetch_feedback.py --to-qrels   # + feedback_qrels_candidates.tsv
Needs the HF write token (read suffices) in ~/.git-credentials or HF_TOKEN env.
"""
import glob, json, os, socket, subprocess, sys

_o = socket.getaddrinfo
socket.getaddrinfo = lambda h, p, f=0, *a, **k: _o(h, p, socket.AF_INET, *a, **k)  # IPv6-first DNS hangs on this box

HERE = os.path.dirname(os.path.abspath(__file__))
GRADE = {"top5": 3, "top10": 2, "after10": 1}

def token():
    t = os.environ.get("HF_TOKEN")
    if t: return t
    return subprocess.run(["bash", "-c",
        "grep -m1 'huggingface.co' ~/.git-credentials | sed -E 's#https://[^:]*:([^@]+)@.*#\\1#'"],
        capture_output=True, text=True).stdout.strip()

def main():
    from huggingface_hub import snapshot_download
    local = snapshot_download("vg15o2/themis-feedback", repo_type="dataset", token=token())
    files = sorted(glob.glob(os.path.join(local, "feedback", "**", "*.json"), recursive=True))
    bundles = []
    for f in files:                                   # daily arrays (current) or single objects (legacy)
        data = json.load(open(f, encoding="utf-8"))
        bundles.extend(data if isinstance(data, list) else [data])
    print(f"[feedback] {len(bundles)} bundle(s) across {len(files)} file(s)\n")

    dump = os.path.join(HERE, "feedback_dump.jsonl")
    with open(dump, "w", encoding="utf-8") as f:
        for b in bundles: f.write(json.dumps(b, ensure_ascii=False) + "\n")

    for b in bundles:
        judged = [r for r in b.get("results", []) if r.get("placement") or r.get("irrelevant") or (r.get("comment") or "").strip()]
        print(f"— {b.get('server_ts', b.get('ts',''))[:16]}  {b.get('name','?')}  [{b.get('mode','auto')}]")
        print(f"  Q: {b.get('q','')[:100]}")
        for r in judged:
            mark = "IRRELEVANT" if r.get("irrelevant") else (r.get("placement") or "")
            c = (r.get("comment") or "").strip()
            print(f"    #{r.get('rank_shown','?'):>3} {mark:<10} {r.get('case_name','')[:52]}" + (f'  "{c[:70]}"' if c else ""))
        if (b.get("missing_case") or "").strip(): print(f"    MISSING: {b['missing_case'][:100]}")
        if (b.get("additional") or "").strip():   print(f"    NOTE: {b['additional'][:140]}")
        print()
    print(f"[feedback] dump -> {dump}")

    if "--to-qrels" in sys.argv:
        out = os.path.join(HERE, "feedback_qrels_candidates.tsv")
        with open(out, "w", encoding="utf-8") as f:
            f.write("# qid(query text)\tdoc_id\tgrade\tsource\n")
            for b in bundles:
                for r in b.get("results", []):
                    if r.get("irrelevant"): g = 0
                    elif r.get("placement") in GRADE: g = GRADE[r["placement"]]
                    else: continue
                    f.write(f"{b.get('q','')}\t{r['doc_id']}\t{g}\t{b.get('name','?')}\n")
        print(f"[feedback] qrels candidates -> {out} (review before merging into the eval set)")

if __name__ == "__main__":
    main()