Verifiable over persuasive
Every statement in the wiki should trace back to its source. An answer without evidence counts as a bug, not a feature.
Non-commercial research project
Wissensbau studies how language models can condense raw notes, emails and documents into a linked wiki – following Andrej Karpathy’s LLM-wiki pattern. The prototype is the testbed for our experiments.
PDF · DOCX · XLSX · PPTX · Email · CSV · Images · Audio
wiki / hardware / battery
„How long does the sensor battery last in continuous use?“
Principles
Every statement in the wiki should trace back to its source. An answer without evidence counts as a bug, not a feature.
Wissensbau is a research and learning project with no product promise. There are no ads, no subscriptions and no user profiles.
Self-hosted, with no trackers and no external fonts. Personal data can be redacted before it is processed.
Method
Raw data stays untouched in a folder – files, emails, or connected sources such as IMAP and PostgreSQL.
A language model extracts topics, entities and concepts and writes wiki pages from them.
Pages are cross-linked; relationships show up in the knowledge graph.
A chat answers from the wiki and names the sources behind each answer.
Research questions
How reliably can text, email, PDF, Office documents, spreadsheets, images and audio be turned into structured knowledge?
How well does raw material become consistent, cross-linked pages that point back to their sources?
Do chat answers demonstrably rest on the wiki? A faithfulness evaluation checks this.
Are people, projects and concepts merged correctly across documents and reflected in the graph?
Keyword search versus embeddings – including a benchmark of how each scales with corpus size.
Redaction of personal data and source trust as building blocks for responsible use.
Application areas
Not a finished product, but scenarios in which we try the method and probe its limits.
Condense lab notebooks, protocols and reading notes into one shared, searchable wiki.
Bring a project’s scattered specifications, reports and email threads together in a traceable way.
Turn knowledge locked in inboxes and file shares into something newcomers can query themselves.
Open up large, heterogeneous holdings – with cross-references that do not exist in the raw material.
Technology
Four layers, deliberately simple: raw data stays untouched, and everything generated is readable and versionable Markdown.
01
Python
Reads raw data, calls the language model and produces wiki pages, entities and links.
02
Markdown · JSON · Graph
Pages as Markdown files, knowledge graph and state as JSON – no separate database.
03
Node.js · Express · TypeScript
Authentication, pipeline control, retrieval and chat endpoints behind a lean REST interface.
04
EJS · Tailwind CSS
Server-rendered, bilingual (DE/EN) and built for small servers and small screens.
Reproducibility
The whole prototype runs in a few containers with fixed resource limits and can be reproduced with Docker Compose – even on a small server. Evaluation scripts for retrieval, faithfulness, entity resolution and PII redaction are part of the project.
FAQ
No. It is a non-commercial research project. There are no prices or sales plans, and no study results exist yet.
From the LLM-wiki pattern described by Andrej Karpathy: a language model maintains a wiki built from raw material instead of searching afresh for every question. We study how well that works in practice.
Access is limited to test accounts so the small server stays reliable. Write to us via the contact page if you are interested.
Raw data stays untouched on the server. For processing, text excerpts are sent to the configured language-model provider – or not sent outside at all when a local model is configured. The prototype’s sample data is fictional.
Yes, the source code is public on GitHub. The setup is reproducible with Docker Compose, even on small hardware.
Through evaluation scripts for retrieval, faithfulness, entity resolution and PII redaction. We only publish results once an evaluation supports them.
About the project
Wissensbau is being built as a research and learning project around LLM-assisted knowledge systems. Feedback, questions about the method and reports of errors are always welcome.
Access is limited to test accounts. Sign in to explore the pipeline, wiki and chat with sample data.