inpi-patents

French patent data, from INPI's servers to analysis-ready tables.

INPI, the French patent office, publishes the full French patent register as WIPO ST36 XML: bibliography, legal status, and the parties on each document, including inventors with city and postcode. Getting from "the data exists" to "I have a table" involves a licence form, an FTP server with no English documentation, a cookie-authenticated API with a small quota, and an XML dialect with a few silent traps.

This package handles the whole path:

• FTP route (bulk): map the server, download the full-register snapshot (~3.5 GB, ~2.5 million documents) and the weekly update files. Resumable, size-verified, restartable.
• API route (targeted): log in to API PI Diffusion, fetch notices, run searches, harvest small sets without fighting the quota by hand.
• One parser for both: every element of the notice, with the known source defects repaired and flagged rather than silently propagated.
• Geography built in: postcode to département to NUTS-3, Corsica and overseas handled, no geocoder needed.
• A CLI and a local dashboard that drive the same steps.

Documentation: https://maliyuam.com/inpi-patent-harvester/ — from requesting access to using the tables. The same pages live in docs/; mkdocs serve browses them locally.

Sixty seconds, no credentials

pip install inpi-patents
inpi parse file tests/fixtures/FR_SYNTHETIC_MODERN.xml

The synthetic record exercises every trap the parser handles: inverted city/postcode, duplicated party renderings, the filing date hidden behind INPI-specific element names, priorities that move the earliest filing date, a foreign inventor whose five-digit code is not a French postcode.

The real thing

inpi env init                                 # write .env, fill in the codes INPI emailed you
inpi ftp probe --cred NOTICES                 # what your licence can see
inpi ftp fetch /Backlog --cred NOTICES        # the snapshot, resumable
inpi parse zips data/raw/ftp/Backlog          # archives -> parquet parts, parallel
inpi merge                                    # parts -> one parquet per table, QA numbers
inpi geo                                      # inventors_geo.parquet with département + NUTS-3
inpi ui                                       # or do all of the above from a browser

import duckdb
duckdb.sql("""
    SELECT app_year, dept, count(*) AS inventors
    FROM 'data/export/inventors_geo.parquet'
    WHERE app_year BETWEEN '2000' AND '2020' AND dept IS NOT NULL
    GROUP BY 1, 2 ORDER BY 1, 3 DESC
""").show()

What you get

On a mid-2025 snapshot: 2.52 million documents, zero parse errors, ~316,000 filings in 2000-2020 of which 99.5% carry an inventor and 84% of inventor slots resolve to a département (99.6% of French-resident ones).

Getting access

The FTP route needs a free licence from INPI: fill in the database selection form on INPI's data portal and email it to their licence desk; access codes come back by email, one login per database. The API route needs a technical account on data.inpi.fr with APIs PI Access activated. The docs walk through both, and include a bilingual request letter template: Getting access.

Code is MIT. Data is INPI's.

This repository contains code, documentation and synthetic test fixtures only. Data obtained from INPI is covered by INPI's reuse licence: access codes are personal, raw data is not redistributable, and INPI must be cited. .gitignore keeps data/, archives, XML, parquet and .env out of the repository. See Licence and redistribution.

Development

git clone https://github.com/maliyuam/inpi-patent-harvester.git && cd inpi-patent-harvester
pip install -e ".[dev,ui]"
pytest                    # no network, synthetic fixtures only
ruff check src tests
mkdocs serve              # docs at http://127.0.0.1:8000

See CONTRIBUTING.md. Not affiliated with INPI.