PastpaperMaster/backend/backfill_null_ai_trio.py

"""Backfill AI trio for questions where knowledge_reminder IS NULL.

For each question, generates fields in two separate LLM calls to avoid token truncation:
  Call 1 → knowledge_reminder + ai_hint  (short, ~500 tokens output)
  Call 2 → solution                      (long, up to 4096 tokens output)

Run from the backend directory:
    uv run python backfill_null_ai_trio.py [--dry-run]
"""

from __future__ import annotations

import asyncio
import json
import sys
from app.services.supabase_client import get_supabase
from app.services.paper_processor import qwen_json_completion


KNOWLEDGE_HINT_PROMPT = """\
You are an expert tutor. Given a past-paper question, produce two short study aids in English.

Return JSON exactly:
{{
  "knowledge_reminder": "2-4 sentences summarising the key concept or formula the student must recall.",
  "ai_hint": "1-3 sentence nudge that guides WITHOUT giving the answer away."
}}

Question:
{payload}
"""

SOLUTION_PROMPT = """\
You are an expert tutor. Given a past-paper question and its reference answer, write a clear, \
step-by-step model solution in English. Show all working. Be thorough but stop when the answer \
is complete — do not pad.

Return JSON exactly:
{{
  "solution": "<full step-by-step solution as a single string, use \\n for line breaks>"
}}

Question:
{payload}
"""


def build_payload(q: dict) -> dict:
    ref = ""
    if q.get("raw_answer_text"):
        ref = q["raw_answer_text"]
    elif q.get("correct_option"):
        ref = f"Correct option: {q['correct_option']}"
    elif q.get("correct_answer"):
        ref = f"Correct answer: {q['correct_answer']}"

    return {
        "question_number": q["question_number"],
        "question_type": q["question_type"] or "long_question",
        "score": q.get("score") or "unknown",
        "question_text": q.get("question_text") or "",
        "topics": q.get("topics") or [],
        "reference_answer": ref,
    }


async def process_one(sb, q: dict, dry_run: bool) -> bool:
    payload_str = json.dumps(build_payload(q), ensure_ascii=False)
    row_id = q["id"]
    qnum = q["question_number"]

    if dry_run:
        print(f"    [dry-run] would process {qnum}")
        return True

    update: dict = {}

    # ── Call 1: knowledge_reminder + ai_hint ─────────────────────────
    try:
        r1 = await qwen_json_completion(
            system_prompt=KNOWLEDGE_HINT_PROMPT.format(payload=payload_str),
            temperature=0.3,
            max_tokens=1024,
        )
        if r1.get("knowledge_reminder"):
            update["knowledge_reminder"] = r1["knowledge_reminder"]
        if r1.get("ai_hint"):
            update["ai_hint"] = r1["ai_hint"]
    except Exception as e:
        print(f"    WARN call-1 failed for {qnum}: {e}")

    await asyncio.sleep(1)

    # ── Call 2: solution ──────────────────────────────────────────────
    try:
        r2 = await qwen_json_completion(
            system_prompt=SOLUTION_PROMPT.format(payload=payload_str),
            temperature=0.3,
            max_tokens=4096,
        )
        if r2.get("solution"):
            update["solution"] = r2["solution"]
    except Exception as e:
        print(f"    WARN call-2 failed for {qnum}: {e}")

    if not update:
        print(f"    SKIP {qnum}: both calls returned nothing")
        return False

    sb.table("paper_questions").update(update).eq("id", row_id).execute()
    return True


async def backfill(dry_run: bool = False) -> None:
    sb = get_supabase()

    papers = (
        sb.table("papers")
        .select("id")
        .eq("course_code", "COMP2211")
        .eq("source_kind", "course_library")
        .execute()
        .data
    )
    paper_ids = [p["id"] for p in papers]
    if not paper_ids:
        print("No COMP2211 course-library papers found.")
        return

    questions = (
        sb.table("paper_questions")
        .select("id, paper_id, question_number, question_type, score, question_text, topics, raw_answer_text, correct_option, correct_answer")
        .in_("paper_id", paper_ids)
        .is_("knowledge_reminder", "null")
        .order("paper_id")
        .order("display_order")
        .execute()
        .data
    )

    if not questions:
        print("No NULL questions found — all done!")
        return

    print(f"Found {len(questions)} questions with NULL knowledge_reminder.")

    # Group by paper for cleaner output
    from collections import defaultdict
    by_paper: dict[str, list] = defaultdict(list)
    for q in questions:
        by_paper[q["paper_id"]].append(q)

    total_updated = 0
    for paper_idx, (paper_id, qs) in enumerate(by_paper.items(), 1):
        print(f"\n[{paper_idx}/{len(by_paper)}] paper_id={paper_id} — {len(qs)} NULL questions")
        for q in qs:
            print(f"  Processing {q['question_number']}...", end=" ", flush=True)
            ok = await process_one(sb, q, dry_run)
            if ok:
                total_updated += 1
                print("done")
            await asyncio.sleep(1.5)

    print(f"\nDone. {total_updated}/{len(questions)} questions updated.")


if __name__ == "__main__":
    dry_run = "--dry-run" in sys.argv
    asyncio.run(backfill(dry_run=dry_run))