ENEM data

Where the questions come from

The question bank comes from INEP's official exams, through the static dataset of the enem-api project, in the updated fork updated-enem-api.

The dataset is GPL-2.0 and OpENEM is AGPL-3.0, so it is never committed to this repository. The importer downloads the tarball from GitHub (about 177 MB) into a folder ignored by git (ENEM_DATASET_DIR, by default openem_backend/prisma/.enem-dataset).

The import

npm run db:import:enem
npm run db:import:enem -- --exam=2023
npm run db:import:enem -- --limit=5 --dry-run

• The script is prisma/import-enem.ts (with import-enem-lib.ts); the pure, tested mapping lives in src/common/enem-import/mapping.ts.
• Each dataset folder is one application of the exam.
• The area comes from the question's position in the booklet; the dataset label only counts when the number is outside the official range. Foreign language questions already arrive with a subject and topic, through the language field.
• All other questions arrive without a topic (null topicId) and without an explanation (null explanation).
• The images go to ENEM_MEDIA_DIR and the API serves them at /media/enem/.
• An incomplete question (no alternatives or an inconsistent answer key) is rejected and listed in the final report.
• The import is idempotent: running it again updates without duplicating and does not undo explanations or classifications made afterwards.
• It never runs inside seed.ts, because the e2e suite runs the seed on every run, and it refuses to run against the test_e2e schema.

Applications

The folder suffix sets the application type (ExamYear.kind) and the letter that goes into the question code, as in ENEM-2014T-136:

The isReapplication field is only true for REAPLICACAO.

ENEM from 1998 to 2008

Until 2008, the ENEM was a single exam of 63 interdisciplinary questions, on one day, lasting 5 hours. Each one becomes a single simulation, which the screen shows as "Prova única" (single exam).

INEP assigns no area to these questions (the dataset has discipline: null). The area the app shows comes from legacy-areas.ts and is OpENEM's classification by content, not official data: made with the help of AI, checked by sampling and flagged as areaClassified in the import. A question that is not in the map comes in without an area.

The answer key of these exams, and of the 2010 and 2011 PPL, comes from INEP's microdata, through tools/enem-fix/gabaritos-antigos.mjs.

Topic classification

Spaced review, recommendations and performance by topic only reach questions that have a topic. There are three steps, in this order:

npm run db:classify:keywords -- --dry-run --sample=20
npm run db:classify:keywords
npm run db:classify:catalog
npm run db:classify -- --limit=50 --dry-run
npm run db:classify

1. Keywords (prisma/classify-heuristic.ts, with the logic in src/common/classification/heuristic.ts): deterministic and without AI. It stays silent when two topics tie, because a wrong topic contaminates review, recommendations and performance. It writes classificationSource: 'keywords' and only touches questions without a topic.
2. Topic catalog (prisma/classify-catalog.ts, with the logic in src/common/classification/catalog.ts): prisma/classificacao-assuntos.json is a classification made once with AI and committed, keyed by externalId, with the "Subject / Topic" label and the confidence. It is not official. It only fills questions without a topic and ignores confidence below 0.6, topics that left the taxonomy and topics from another area. It runs in the bootstrap right after the keywords, and with both steps the whole bank ends up with a topic.
3. AI (prisma/classify-questions.ts): takes what is left, such as the questions of an exam imported later. It only writes decisions with a confidence of at least 0.6 (--min-confidence) and needs a configured provider; see Artificial intelligence.

Fixing the dataset

The dataset came from a PDF extraction and has defects: questions without the text of the alternatives, unreadable statements in 2017 and 2021 (fonts without a Unicode mapping), formulas turned into loose text, a Spanish answer key repeating the English one and missing questions. tools/enem-fix/ rebuilds those questions from the official booklets:

• parse.mjs reads the booklet by coordinates, and whatever does not turn into clean text is cropped as an image;
• gabarito.mjs reads the answer by its position on the page, because the linearized text gets it wrong (in 2015 the columns are misaligned; in 2009 there are four booklets side by side);
• criar.mjs builds from scratch, from provas-novas.json, the exams the dataset does not have, and --comparar validates the extractor on an exam the dataset already has.

Two safeguards keep a question from being matched wrongly: the booklet numbering is checked against the dataset, and the official answer key only replaces the dataset's one when they agree on at least 90% of the exam. Whatever does not pass is left out, as an open item in the report.

Essay prompt catalog

The Redações (essays) screen has the official ENEM essay prompts from 1998 to 2025 (regular, PPL or reapplication and the other applications), with the prompt, the supporting texts and the facsimile of the booklet page.

The catalog is static and committed: openem_frontend/src/data/redacoes.json and the figures in openem_frontend/public/redacoes/. There is no backend and no relation to the essays module. Everything is generated by tools/enem-fix/redacao/ from INEP's booklets, whose PDFs stay out of git.

These are official exam documents: the text is reproduced, never rewritten. No correcting, summarizing, modernizing the spelling or making up labels. The spelling of the prompt is the booklet's, and song lyrics go in whole. Whatever departs from the transcription is flagged in the data and shown on screen: 2008 has no prompt sentence, and text the PDF only has as an image or in an unreadable font goes in as a cropped figure. The screen reader descriptions of the figures are OpENEM's and are not part of the document.

Do not edit the JSON by hand. Change the script and run, from tools/enem-fix/:

node redacao/montar.mjs
node redacao/verificar.mjs

verificar.mjs rereads each page with a second extractor and has to end with 0 differences. Then run npm test in the frontend: src/data/redacoes.test.ts locks the catalog structure.