| """Genera docs/qa_ingest_<doc_id>.md a partir de ingest_report.json, figures.jsonl e |
| informal.jsonl (docs/PLAN_RAG.md Fase 2, tarea 10). La rúbrica manual de las 15 |
| descripciones de figura se completa a mano en el .md generado. |
| |
| SEC-2 (docs/AUDIT_2026-09-23.md, docs/PLAN_CORRECCIONES.md Fase 2): este reporte |
| se publica (git, HF). Si el documento tiene `redistributable=false`, se redacta |
| todo contenido literal derivado de su texto — fórmulas, transcripciones de |
| huecos y descripciones de figura — y sólo quedan conteos y metadatos |
| estructurales; mismo criterio que ya aplica `scripts/publish_hf.py` a las |
| tablas `informal`/`figures` (no las sube si el documento no es redistribuible). |
| |
| Uso: python scripts/qa_ingest_report.py --doc-id edd3436ca148 |
| """ |
|
|
| from __future__ import annotations |
|
|
| import argparse |
| import json |
| from collections import Counter |
| from pathlib import Path |
|
|
| QA_GATE = (8, 16, 17, 22, 24, 37, 99) |
| _REDACTED_NOTE = "Contenido derivado del documento redactado (`redistributable=false`); disponible localmente en `data/parsed/` (gitignored), nunca en git/HF." |
|
|
|
|
| def main() -> None: |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--doc-id", default="edd3436ca148") |
| args = ap.parse_args() |
| parsed = Path("data/parsed") / args.doc_id |
| report = json.loads((parsed / "ingest_report.json").read_text(encoding="utf-8")) |
| chunks = [json.loads(l) for l in open("data/corpus/informal.jsonl", encoding="utf-8") if l.strip()] |
| figures = [json.loads(l) for l in open("data/corpus/figures.jsonl", encoding="utf-8") if l.strip()] |
| cards = [c for c in chunks if c["unit_type"] == "problem_card"] |
| by_n = {c["item_number"]: c for c in cards} |
| doc_path = parsed / "document.json" |
| redistributable = json.loads(doc_path.read_text(encoding="utf-8")).get("redistributable", False) if doc_path.exists() else False |
|
|
| diffs = report.get("formula_diffs", []) |
| accepted = [d for d in diffs if "after" in d] |
| rejected = [d for d in diffs if "rejected" in d] |
| areas = Counter(a for c in cards for a in c["math_area"]) |
| fig_by_id = {f["figure_id"]: f for f in figures} |
|
|
| lines = [ |
| f"# QA de ingesta — `{args.doc_id}` (problemas resueltos olimpiadas.pdf)", |
| "", |
| "Generado por `scripts/qa_ingest_report.py`; criterios en `docs/PLAN_RAG.md` Fase 2, tarea 10.", |
| "", |
| "## Resumen", |
| "", |
| f"- Páginas: {report['n_pages']} · cards: {report['n_cards']} · **emparejadas problema↔solución: {report['n_paired']}/100** · chunks: {report['n_chunks']} (cards + figuras)", |
| f"- Figuras recortadas: {report['n_figures_detected']} · con descripción VLM: {sum(1 for f in figures if f['description'])} · avisos red de seguridad vectorial: {len(report['figure_warnings'])}", |
| f"- Huecos del layout transcritos por VLM: {len(report.get('gaps', []))} regiones", |
| f"- QA de fórmulas: {len(accepted)} correcciones aplicadas, {len(rejected)} propuestas rechazadas por las guardas", |
| f"- Llamadas al VLM/LLM en esta corrida (sin caché): {report['llm_calls']}", |
| f"- Áreas (cards): {dict(areas.most_common())}", |
| "", |
| "## Gate 1 — emparejamiento problema↔solución (objetivo 100/100)", |
| "", |
| f"Resultado: **{report['n_paired']}/100**. Avisos: {report['warnings'] or 'ninguno'}.", |
| "", |
| "Notas del split por LLM:" if redistributable else f"Notas del split por LLM: {len(report.get('split_notes', []))}. {_REDACTED_NOTE}", |
| *([f"- {n}" for n in report.get("split_notes", [])] if redistributable else []), |
| "", |
| "## Gate 2 — figuras de los problemas 8, 16, 17, 22, 24, 37, 99", |
| "", |
| ] |
| if redistributable: |
| lines += ["| Problema | Figuras | detected_by | Descripción (inicio) |", "|---|---|---|---|"] |
| for n in QA_GATE: |
| for fid in by_n[n]["figure_ids"]: |
| f = fig_by_id.get(fid) |
| if f: |
| lines.append(f"| {n} | `{fid}` | {f['detected_by']} | {f['description'][:110]}… |") |
| else: |
| lines += [f"_{_REDACTED_NOTE} (las descripciones se generan a partir de las figuras del documento)._", "", "| Problema | Figuras | detected_by |", "|---|---|---|"] |
| for n in QA_GATE: |
| for fid in by_n[n]["figure_ids"]: |
| f = fig_by_id.get(fid) |
| if f: |
| lines.append(f"| {n} | `{fid}` | {f['detected_by']} |") |
| lines += [ |
| "", |
| "## Rúbrica manual — 15 descripciones (sí/no: ¿nombra todos los puntos?, ¿relaciones correctas?, ¿valores dados?)", |
| "", |
| "| # | Figura | Puntos | Relaciones | Valores | Observación |", |
| "|---|---|---|---|---|---|", |
| *[f"| {i+1} | `{f['figure_id']}` | | | | |" for i, f in enumerate(figures[:15])], |
| "", |
| "## Huecos del layout de MinerU recuperados por VLM", |
| "", |
| ] |
| if redistributable: |
| lines += [ |
| "| Página | Capa de texto (evidencia) | Transcripción (inicio) |", |
| "|---|---|---|", |
| *[f"| {g['page']} | {g['text_layer'][:70]} | {g['transcription'][:90].replace(chr(10), ' ')} |" for g in report.get("gaps", [])], |
| ] |
| else: |
| gap_pages = ", ".join(str(g["page"]) for g in report.get("gaps", [])) |
| lines.append(f"{len(report.get('gaps', []))} regiones recuperadas (páginas: {gap_pages or '—'}). {_REDACTED_NOTE}") |
| lines += [ |
| "", |
| "## QA de fórmulas (6b) — correcciones aplicadas", |
| "", |
| ] |
| if redistributable: |
| lines += ["| Página | Antes | Después |", "|---|---|---|", *[f"| {d['page']} | `{d['before'][:60]}` | `{d['after'][:60]}` |" for d in accepted]] |
| else: |
| pages = Counter(d["page"] for d in accepted) |
| lines.append(f"{len(accepted)} correcciones en {len(pages)} páginas (páginas: {', '.join(str(p) for p in sorted(pages))}). {_REDACTED_NOTE}") |
| lines += [ |
| "", |
| f"Rechazadas por las guardas (cambio sólo cosmético o absorción de texto vecino): {len(rejected)} — detalle en `data/parsed/{args.doc_id}/ingest_report.json` (local, gitignored).", |
| "", |
| "## Gold de OCR (10 páginas transcritas a mano)", |
| "", |
| "Pendiente: requiere transcripción manual humana (no se fabrica). Sustituto disponible hoy: las 5 correcciones reales detectadas por el VLM en 2 páginas de muestra (docs/DECISIONS.md #19).", |
| "", |
| ] |
| out = Path("docs") / f"qa_ingest_{args.doc_id}.md" |
| out.write_text("\n".join(lines), encoding="utf-8") |
| print(out) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
