PyTales
The friendly helper for everything in research, teaching and administration
An AI assistant with tools for documents, literature, transcription and tables – for Windows, macOS, Linux and Android.
Research prototype tested at the chair and in the team · as of 10 November 2026
Demo with prepared answers: No AI is running and nothing leaves your browser. All outputs are examples, not real model outputs; all names and data are made up.
New chat
Demo mode – prepared answers
One click plays the example in the demo.
The path of a request – what happens on the device and what a language model on the server takes over.
Lora’s own model Lora Befehle recognises in 3.5 ms whether a setting, a tool or a deadline is meant – otherwise Lora asks.
Commands run locally at once, short tasks with a local language model, large ones via the server.
Names, places and numbers become placeholders such as [Name 1] before anything leaves the device.
The language model plans and calls tools. Anything that changes something waits for “Run”.
The placeholders turn back into names; sources are numbered.
Lora Namensschutz (“name protection”) and Lora Befehle (“commands”) are small enough for the device and were trained on invented texts only. Each is measured on a test set that never went into training.
Finds names, institutions and places
99.2 % of private details found
Target of at least 98 % reached
Found, in per cent – compared with the fixed rules alone
rules only78.2 with model99.2Understands requests on the device
0.0 % of commands executed wrongly
Target of at most 2 % reached · rules only: 19.4 %
Recognised straight away, in per cent – compared with the fixed rules alone
rules only38.1 with model84.8Example
Frau Keller unterrichtet an der Lindenschule in Pirna. (Ms Keller teaches at the Lindenschule in Pirna.)
It is based on gbert-base (German BERT); the model gives every word piece a label. Together with the app’s rules it finds almost all private details; it still replaces public names too often.
| Figure | with model | rules only | Target | Status |
|---|---|---|---|---|
| private names, institutions and places found | 99.2 %97.2–99.8 % | 78.2 % | at least 98 % | reached |
| public names wrongly replaced | 7.4 %4.0–13.5 % | 9.9 % | at most 2 % | open |
| Time per paragraph (app on a PC) | ≈ 11 ms | – | under 50 ms | reached |
| Size as ONNX int8 | 110 MB | – | – | 0.0 percentage points from the full model |
Test set: 174 invented, everyday texts (interview, parent meeting, report, e-mail, minutes, table, chat, academic text) with 261 private and 121 public details. Counting is per detail: it is found if all its core words are replaced; a public detail counts as wrongly replaced as soon as one word is replaced. The ranges below the values are Wilson 95 % intervals.
found wrongly replaced · Evaluation by the training script with rules and model: 5.8 % wrongly replaced; the app with its own rules reaches 7.4 % (above); 99.2 % are found in both cases.
What the numbers do not show: the test corpus was written before the training templates, and words from the test set do not occur in the training data – but both are made up. A test with real, approved texts is still pending, as is a comparison with other filters such as Presidio or spaCy on the same corpus and an analysis of how many texts stay completely free of gaps.
It is based on distilbert-base-german-cased. The model maps inputs such as “make it dark” or “remind me of the deadline on 15 December” to an action of the app – offline, without a request to a server. Guiding rule: a wrongly executed action is worse than a follow-up question.
| Figure | with model | rules only | Target | Status |
|---|---|---|---|---|
| commands executed wrongly | 0.0 % | 19.4 % | at most 2 % | reached |
| commands recognised straight away | 84.8 % | 38.1 % | at least 90 % | open |
| commands recognised, with a follow-up question | 97.0 % | – | – | two buttons to choose from |
| open questions correctly passed on to a language model | 100 % | – | – | – |
| Time per input | ≈ 3.5 ms | – | – | ONNX int8, 68 MB |
Test set: 386 invented requests for 178 actions, including typos and English.
What the app replaces, where the mapping is kept and what the filter does not do.
With a model on the server: the text of the messages, the text of attached files and their names – with placeholders instead of personal details. Pseudonymisation is on by default. Web search and literature databases also only receive placeholders. Images and scanned pages cannot be pseudonymised; the app asks before sending them.
With a local model nothing leaves the device, so placeholders are not needed. The app addresses language models on the server via an OpenAI-compatible interface – during testing at the chair, the university’s AI service. There are no usage statistics; error logs stay on the device and are only included if you use “Report a problem” yourself.
Details are replaced by placeholders in square brackets – [Name 1], [Einrichtung 2], [E-Mail 1], optionally with a role such as [Schüler/in 2] – not by made-up names. This way the language model sees that something was replaced and is instructed to keep the placeholders unchanged.
The same person keeps the same placeholder in all requests; with “placeholders per chat” every chat starts at [Name 1], so the server cannot link people across chats. Other spellings (MÜLLER, Mueller) get the same placeholder. Inflected German names are replaced too: “Annas Heft” becomes “[Name 1]s Heft”, “dem Müller” becomes “dem [Name 2]” – and that is how they come back.
The mapping is kept as a file in the app’s data folder on the device, is not synchronised and stays until you delete it under Settings › Privacy. It is not separately encrypted (unlike the API keys). On the device the answer gets the real details back, in created documents too; only tools that send to the internet keep the placeholders.
Fixed rules on the device: email addresses, phone numbers, addresses with postcode and town, dates of birth, student, pupil, staff and customer numbers, file references, IBAN, BIC and credit cards (with check digit), tax ID, ID card and social security numbers, number plates, IP addresses, profile links and @ handles; names after roles, greetings and signatures, whole name columns in tables, CSV and Excel, names from email addresses, file and folder names.
Plus Lora Namensschutz for names, institutions and places with a personal reference, a custom list (also from class lists) and “Never replace”. Cited authors and those in your own reference library stay as they are. Every rule can be switched off individually; on request, a local language model also reads through the text (“Thorough”).
If Lora Namensschutz is missing or switched off, the rules carry on alone. If pseudonymisation itself fails, the request is not sent. When in doubt, the filter replaces too much rather than too little.
What was replaced is shown in the preview under Settings › Privacy and in the log of every agent run; one click on “do not replace” or “also replace” applies to all chats from then on. In the vault the strict level applies: everything on, no exception for public sources, plus the local language model.
Pseudonymised data remain personal data (Recital 26 GDPR): whoever has the mapping can trace them back. The filter reduces the risk but does not by itself make a use GDPR-compliant. Responsibility remains with whoever uses Lora with real data, or their institution. Whether a server service may be used for such data, for example under a data processing agreement, is something to clarify with the data protection officer.
Lora Namensschutz (files lora-namensschutz-pytales.*, as of 09/11/2026): gbert-base by deepset (MIT licence), three epochs on 27,000 invented texts, threshold 0.9, ONNX int8. The templates were written by Claude (Anthropic) as teacher model; a script inserted invented names. Lora Befehle (files lora-befehle-pytales.*, as of 10/11/2026): distilbert-base-german-cased (Apache-2.0), five epochs on 26,360 invented requests, threshold 0.8, ONNX int8.
Licence: Lora and both models are free for everyone to use, permanently. They may only be modified or redistributed with written permission from PyTales. The base models remain under their own licences; their authors have neither reviewed nor endorsed Lora Namensschutz and Lora Befehle. No warranty – the models can make mistakes; when in doubt, a human checks the result.
History: 09/11/2026 test corpus, training and integration of Lora Namensschutz on the PC · 10/11/2026 Lora Befehle and mixed mode.
Names that slipped through or false alarms – please report them in the app under Settings › Report a problem or to dominik.toepper@tu-dresden.de, using made-up examples only.
Lora is a personal project and not an official service of TU Dresden.