How local AI actually works
A plain-language look at what a language model is, and how CortexEleven turns a box of paper into answers you can search - all on your own hardware.
A language model learns the patterns within language
Trained on vast amounts of text, a large language model builds a rich statistical understanding of how words, concepts and facts relate to one another. Given a question, it draws on those learned relationships to compose a precise, relevant response. CortexEleven applies that capability to your documents, not the open internet.
Six steps, one pipeline
Every document travels the same path - from a scanned page to a searchable, cited answer. Here is what happens at each step.
Reading the page: scan, then OCR
A scan or photo is just pixels - a picture of text, not text a computer can use. Optical character recognition (OCR) looks at those pixels and works out the actual characters, word by word, scoring how confident it is on every page.
- Handwriting, tables and stamps are read, not skipped
- Low-confidence pages re-run on a stronger model
- Every page keeps its confidence score
Turning text into data the model can use
Raw text is hard to search well. CortexEleven pulls out the facts that matter - people, dates, document types, obligations - classifies each document, and links related records into a knowledge graph. The archive starts to organise itself.
- Entities and key fields extracted automatically
- Documents classified and de-duplicated
- Related records connected across departments
Interpret, then answer in plain language
The model reads across the indexed text to understand what a question is really asking, then composes an answer drawn only from your records - with a citation back to the exact source page, so every claim can be checked.
- Natural-language and keyword search across everything
- Answers grounded in your documents, never invented
- Inline citations to the source document and page
All of this runs where your data already lives
The whole pipeline - OCR, index, model and search - runs inside the boundary you choose. That is what makes it safe to point at decades of sensitive records.
Runs on your hardware
The model and index sit inside your appliance or private tenant. Nothing is sent to an external provider.
Your data trains nothing
Your records are never used to train someone else’s model. The knowledge stays yours.
Grounded, not guessed
Answers are synthesised only from your own documents, every claim traceable to a source page.
Turn your archive into a searchable knowledge hub
See CortexEleven digitise and search a sample of your own records - on your hardware, in your environment.