Cortex

The problem

I have many documents: PDFs, notes, invoices and papers. Search finds words in them, but it does not tell me what I know about a subject across all of them. Most AI tools that can do this want my files on a different server. I wanted a tool that runs on one machine at home, and no data must leave that machine or the local network.

What it does

Cortex reads your documents and answers your questions about them, and all your data stays on your own computer. You use it in a browser on a laptop or a phone on your home network.

• Give each person their own account and their own private set of documents.
• Add PDF, Word, text and web page files, and sort them with tags.
• Ask questions in a chat, and get answers that quote your files and name the source, down to the page in a PDF.
• Get a wiki, a set of linked notes that Cortex writes for you, with pages about people, companies and topics.
• See a warning when new information does not agree with the wiki.
• Ask about only some files, for example the files with the tag "#tax-2025".
• See the wiki as a map of linked pages, and see how each page changed over time.
• Download the wiki for the Obsidian notes app.

Screenshot pending

Decisions & trade-offs

Two layers, not only search. I keep a search layer and also a wiki that the model writes. The search layer gives exact quotes, and the wiki keeps knowledge that grows over time. The cost is a second, slower model pass for each file.

A small model with strict limits. I let the app select the pages and apply the changes, and the model only writes section-sized edits. This lets a 9B model that runs at home update the wiki safely. The cost is that the model cannot make small edits or change the wiki structure on its own.

Wiki pages in the database, not on disk. I keep the wiki in Postgres. Each person gets a private wiki, live updates and a history of changes. The cost is that the files on disk are only an export, and changes in Obsidian do not come back yet.

Fully offline. I run all parts on one host, and Cortex does not use the cloud models of Ollama. Your files stay on your machine. The cost is speed and quality, because the results depend on what the host can run.

What's next

• A one-click revert. The app saves each version of a page, but now it only shows the history.
• Features that are not in v1: "remember this" in chat, OCR for scanned documents, audio, two-way Obsidian sync and a job that cleans the wiki.
• Known limits: you cannot edit the wiki in the app, and Cortex expects a trusted home network without HTTPS hardening.
• Cortex is not publicly available yet.