dofjson

Prototype client for the JSON open data service exposed by sidof.segob.gob.mx, the Secretaría de Gobernación's system for Mexico's official gazette (DOF, Diario Oficial de la Federación).

The service's public docs only show sample responses, but its real, unauthenticated endpoints were found under https://sidof.segob.gob.mx/dof/sidof/:

Note that this service reports a missing day as 200 OK with empty legal provision lists, not as an error, and that some of the dates it lists as unpublished were in fact published — see the days SIDOF loses.

This is an experimental package for evaluating whether this service is a viable alternative (or complement) to dof2md's PDF download + Markdown conversion pipeline — legal provisions already come with structured HTML content, which may be easier to work with than OCR'd PDFs.

On top of the raw endpoints, the client offers legal-provision-scoped downloads that resolve a legal provision's page span (infer_paginas) and fetch it in whichever form you want:

• download_nota_imagenes(codNota) — the legal provision's scanned page image(s).
• download_nota_pdf(codNota) — the legal provision as its own PDF: the whole edition PDF (there is no per-legal-provision PDF endpoint) sliced to just the legal provision's pages, using pypdf.

Usage

pip install -e "packages/dofjson[test]"
dofjson 2026-07-16 --endpoint notas --outdir output

Building a local archive of daily indexes (--archivo)

dofjson --archivo downloads the daily legal provisions index incrementally, day by day, over a whole date range (by default from January 2, 1917 to today). For each date it does exactly what dofjson YYYY-MM-DD --endpoint notas does — get_notas(date) filtered with quita_notas_sin_titulo — and saves one JSON per day. It does not download each legal provision's content or scanned images: only the index.

dofjson --archivo                                      # 1917-01-02 -> today
dofjson --archivo --desde 1980-01-01 --hasta 1980-12-31
dofjson --archivo --pausa 1.0                          # slower (kinder to the server)

Output goes to notas-archivo/ (configurable with --outdir), a local, never-committed directory (it is in the repo's .gitignore), with the same per-day filenames the plain command produces:

notas-archivo/
  .completados                 # registry of finished days (for resuming)
  2026/
    15072026-notas.json        # index for 2026-07-15 (get_notas, filtered)
    16072026-notas.json
  1980/
    02011980-notas.json

The mode is resumable and idempotent: the .completados registry records the finished days, so each run only fetches what is missing. Days that fail with network errors are not marked and get retried on the next run; days with no edition (holidays, weekends) are marked so they are not retried forever. "Today" is never marked, so late additions are picked up by a later run. You can interrupt with Ctrl-C and resume at any time.

The full range is ~40,000 days: a long download, meant to be run in parts. Start with a bounded range via --desde/--hasta if you only need an era.

The days SIDOF loses, and where they are recovered from (--respaldo)

SIDOF does not report a day it is missing as an error. It answers 200 OK with every legal provision list empty — which is also how it reports a Sunday — and lists the date under FechasSinPublicacion in GET /diarios/{year}. Most of those dates are genuine: weekends and holidays. Some are not.

On 8 March 1999 the DOF published the decree amending articles 16, 19, 22 and 123 of the Constitution. SIDOF has no trace of it: the day is empty, the legal provision's codNota returns {"Nota": []}, and its codDiario 404s. Sampling four years (1999, 2006, 2010, 2020) turned up eight such days — dates SIDOF calls unpublished that were published.

www.dof.gob.mx, the DOF's own website, is a separate system on a separate database, and it has them. So an empty answer is no longer taken at face value: on a weekday, the day is put to the website before being written off.

dofjson --archivo                        # habiles (default): re-check Mon-Fri
dofjson --archivo --respaldo todos       # also weekends (~10,000 more requests)
dofjson --archivo --respaldo nunca       # trust SIDOF alone

[1999-03-08] SIDOF no la tiene; recuperada de dof.gob.mx

The same applies to a single date, so a lost day is reachable directly:

dofjson 1999-03-08 --endpoint notas      # -> "fuente": "dof.gob.mx", 22 legal provisions

The text of those legal provisions is recoverable too, not just their titles. dofweb.get_nota(codNota) reads a legal provision's page on the website and returns it in the shape client.get_nota() uses, with the legal provision's HTML in cadenaContenido — the same string SIDOF would have served, so nota2md converts it to Markdown by the ordinary HTML path:

from dofjson import dofweb

dofweb.get_nota(4997808)["Nota"]["cadenaContenido"]   # DOF 03-03-1999

On a legal provision both sources have, the recovered HTML differs from SIDOF's only in escaping its accents as entities, and the Markdown built from either is identical. The page carries no codDiario, codEdicion or pagina, so the image and PDF paths stay SIDOF-only; a legal provision the site has no HTML for comes back with existeHtml "N", and an unknown codNota with "Nota": [], as SIDOF answers.

Which source a day came from is recorded, never inferred. Every saved day carries a fuente key ("sidof" or "dof.gob.mx"), and the registry stores it next to the date, so provenance can be audited after the fact:

1999-03-06	sin-edicion
1999-03-08	dof.gob.mx
1999-03-09	sidof

In the compact --titulos dataset the marker rides along on the legal provisions it applies to: a legal provision carries fuente only when its day did not come from SIDOF, since repeating "sidof" on all ~1.2 million rows would cost more than it says.

What the fallback carries, and what it does not

The website's daily index lists the substantive gazette — PE, PJ, PL, OA, OD — and leaves out the three bulk-announcement groups, which are reachable on the site only through its POST search form:

On days both sources have, the recovered set of codNota matches SIDOF's exactly once those three are excluded (checked on days sampled from 1999 through 2026). A recovered day is therefore complete with respect to what the gazette enacted and short of what it announced, and says so in notasIncompletas rather than passing for a whole day.

The website's per-legal-provision index starts in January 1999; before that it holds only scanned images, so an older day returns an edition with no index. Those come back in edicionesSinIndice as {"codEdicion", "codDiario"}: no titles to list, but proof the gazette was published, which is what keeps the day off the empty pile. Every day confirmed lost from SIDOF so far is 1999 or later, inside the range where titles can actually be recovered.

On a page served for the wrong date. The index prints the date it is actually serving, and it has been seen answering with a different day's page. Since the parser stamps each legal provision with the date that was asked for, taking such a page at face value would file real legal provisions under the wrong day. Every page carrying content is therefore checked against what it claims to be, and a mismatch raises dofweb.PaginaDeOtroDia. --archivo treats that like a network error and leaves the day to retry — believing it would corrupt the day, and calling the day empty would bury it for good. Editions the gazette never ran carry no date and no content, which is not a mix-up and is not treated as one.

On TLS. www.dof.gob.mx serves its leaf certificate without the intermediate that signs it, so verification fails with "unable to get local issuer certificate" on any client that does not chase the issuer itself. The missing GoDaddy intermediate — and its root — ship in dofjson/certs/dof-gob-mx-chain.pem. The system trust store is tried first and the bundled chain only on failure. Certificate verification is never disabled.

Building a compact titulo dataset from the release (--titulos)

dofjson --titulos builds a small codNota + titulo + fecha dataset out of every legal provision ever published, sourced from the notas-archivo GitHub release (one notas-YYYY.tgz per year, 1917 to last year, plus one notas-YYYY-MM.tgz per month of the current year). Each asset is downloaded straight into memory, its daily JSON indexes are read without ever writing them to disk, and only codNota/titulo/fecha are kept from every legal provision (titulo is Spanish for "title", fecha for "date") — codNota to fetch that legal provision's full content later, titulo for exploratory analysis of the titles themselves, fecha to place each title in time. The result is a single gzip-compressed JSONL file (~1.2 million legal provisions fit in a few tens of MB): small enough to move to a Colab GPU runtime for experiments.

dofjson --titulos                    # -> titulos/titulos.jsonl.gz
dofjson --titulos --outdir /content  # e.g. from a Colab notebook

import gzip, json
with gzip.open("titulos/titulos.jsonl.gz", "rt", encoding="utf-8") as f:
    notas = [json.loads(line) for line in f]
# notas[0] == {"codNota": 4434476, "titulo": "CIRCULAR nº. 164, ...", "fecha": "23-03-1917"}

Or use the function directly:

from pathlib import Path
from dofjson.titulos import download_legal_provisions_titles

download_legal_provisions_titles(Path("titulos.jsonl.gz"))

Pass cache_dir to keep the downloaded .tgz assets on disk (e.g. a Colab-persisted folder) so rebuilding the dataset later only fetches assets not already there:

download_legal_provisions_titles(Path("titulos.jsonl.gz"), cache_dir=Path("cache"))

download_dof_assets does just the download/cache part, independently of building the titles dataset:

from dofjson.titulos import download_dof_assets

download_dof_assets(Path("cache"))  # one notas-YYYY[-MM].tgz per asset

Development

pytest packages/dofjson