"""GPT-5.2 generator. Uses OpenAI's API to generate reviews for a paper with prompts/control_review.md. Falls back through base64 upload, URL reference, and page images for PDF delivery. """ from __future__ import annotations from pathlib import Path from src.db.queries import insert_comments_batch, delete_comments, count_comments, get_paper from src.lib.openai import ( assess_with_pdf, assess_with_pdf_url, assess_with_pdf_images, assess_with_text, download_pdf, ) from src.lib.json_repair import parse_json_array PROMPTS_DIR = Path(__file__).parent.parent / "prompts" PROMPT_FILE = PROMPTS_DIR / "control_review.md" MODEL = "gpt-5.2" class ControlGPTGenerator: source = "gpt-5.2" async def generate(self, paper_id: str, force: bool = False, **kwargs) -> int: """Generate GPT-5.2 review with minimal prompt.""" if not force and count_comments(paper_id, "gpt-5.2") > 0: print(f" {paper_id}: GPT-5.2 comments already exist, skipping (use --force)") return 0 paper = get_paper(paper_id) if not paper: return 0 if force: delete_comments(paper_id, "gpt-5.2") prompt = PROMPT_FILE.read_text() pdf_url = paper.get("pdf_url") raw_output = None if pdf_url: pdf_bytes = await download_pdf(pdf_url) # Try 1: base64 upload try: raw_output = await assess_with_pdf( pdf_bytes=pdf_bytes, system_prompt="", user_prompt=prompt, model=MODEL, temperature=0, ) except Exception as e: print(f" {paper_id}: base64 failed ({type(e).__name__}), trying URL fallback...") # Try 2: Responses API with PDF URL if raw_output is None: try: raw_output = await assess_with_pdf_url( pdf_url=pdf_url, system_prompt="", user_prompt=prompt, model=MODEL, temperature=0, ) except Exception as e: print(f" {paper_id}: URL fallback failed ({type(e).__name__}), trying image fallback...") # Try 3: render pages as images if raw_output is None: raw_output = await assess_with_pdf_images( pdf_bytes=pdf_bytes, system_prompt="", user_prompt=prompt, model=MODEL, temperature=0, ) else: manuscript_text = paper.get("manuscript_text") if not manuscript_text: print(f" {paper_id}: No PDF URL or manuscript text, skipping") return 0 raw_output = await assess_with_text( text=manuscript_text, system_prompt="", user_prompt=prompt, model=MODEL, temperature=0, ) comments = parse_json_array(raw_output) if not isinstance(comments, list): raise ValueError(f"Expected JSON array, got {type(comments)}") batch = [] for comment in comments: text = comment if isinstance(comment, str) else str(comment) if len(text.strip()) < 20: continue batch.append({ "id": f"{paper_id}:gpt-5.2-{len(batch) + 1}", "paper_id": paper_id, "source": "gpt-5.2", "content": text.strip(), }) insert_comments_batch(batch) return len(batch)