IPfactory
Swiss patent search engine · 76 million patent families, covering 174 million patent documents · updated 13 Oct 2026, EPO week 41
Entire infrastructure in Switzerland · your queries and results are never stored · Learn more
Swiss patent search engine · 76 million patent families, covering 174 million patent documents · updated 13 Oct 2026, EPO week 41
Entire infrastructure in Switzerland · your queries and results are never stored · Learn more
Where IPfactory runs, what it keeps, and what it never sees. Written to be checked, not to reassure.
The whole service runs in Switzerland. The web application and the semantic engine run on a server rented from Infomaniak (Public Cloud, Switzerland). The language model behind the Agent, and the speech-to-text behind the dictation button, run on Infomaniak AI Services, also in Switzerland. No call leaves the country while you search.
Nothing of what you search. Queries, filters, results, agent conversations and dictated audio are processed and discarded. An agent run is held in the server's memory for up to two hours so that your browser can reconnect if the connection drops, then it is gone; it is never written to disk.
The usage journal keeps numbers: how many searches, agent runs and assistant calls each account made per day, and how many visitors opened each page per day and per campaign link. For security it also keeps sign-in events (time, account, network address). It holds no query, no result and no visitor address.
Reports we host for you (landscape pages under /r/…) stay on the Swiss server, behind a link
that cannot be guessed, excluded from search engines; we count how often each is opened, nothing else. We remove a report on
request.
The semantic engine is IPfactory's own: it runs on our server and is calibrated on the documents that patent examiners actually cite, which is what the citation-likelihood zones measure.
The Agent is driven by an open-weights language model (Kimi K2.6) executed by Infomaniak AI Services in Switzerland, under terms that exclude any retention of your text and any training on it. The model's publisher never receives anything: the model runs on Infomaniak's machines, not the publisher's. The same holds for the dictation (Whisper). We may change the model behind the Agent; this page will say so.
When you connect Claude, ChatGPT, Cursor or your own agent to IPfactory, your question lives in your assistant. Our server receives the search queries the assistant formulates, answers them and keeps none of them; the journal counts the calls per tool per day, under daily quotas per account.
Connecting claude.ai, ChatGPT or Claude Code is a delegated sign-in (OAuth): you sign in on this site and allow the assistant; it receives a token that works only for the search tools, under your account's quotas, and never sees your password. Tokens expire after eight hours and are renewed silently for up to ninety days; you can disconnect an assistant at any time from the Connect page, and changing your password disconnects them all. Access keys and tokens are stored only as hashes.
Passwords are stored as salted hashes, never in clear; sessions are signed tokens. Your name, e-mail address, how you arrived and your usage counts (numbers only) are kept in our customer relationship system (Odoo, hosted in the European Union) so that we can follow up with you; no search content ever goes there. When paid plans open, payments will be handled by Stripe: card details never reach our servers.
Nobody at IPfactory reads searches, because there are none to read: the design above, not a policy, is what protects them. Two people hold keys to the server. Backups stay on the same Swiss infrastructure (Infomaniak Public Cloud) and contain the patent corpus, the account records and the usage journal.
Bibliographic data, abstracts and families come from the European Patent Office (DOCDB); legal status from the EPO's INPADOC data; drawings from the EPO's Open Patent Services; some fields from Google's public patent datasets. Abstracts that were not available in English were machine-translated once, at corpus build time — these are published patent texts, not user data.
IPfactory ranks published patent documents by semantic similarity and calibrated citation likelihood. It does not assess patentability, validity or freedom to operate. For a legal opinion, consult a patent attorney — our results are meant to make that conversation shorter and better informed.