Model performance
What OnAI is good at, and where it stops.
A candid account of how the assistant performs, where its knowledge comes from, and the limitations you should plan around rather than discover the hard way.
How OnAI produces an answer
OnAI is a large language model served through our inference partners. When you send a message, Passenger assembles the relevant part of the conversation — your earlier messages, OnAI's replies, file names and the project it belongs to — and asks the model to continue from there.
The model does not look anything up in a private database of facts. It predicts a helpful continuation from the patterns it learned in training, plus whatever context the workspace supplies. That is why specific, recent context matters so much: it is the part of the answer that comes from you.
How much context it holds
Long conversations are included up to a limit; once a thread grows past it, the oldest turns are trimmed first. If a detail from early in a thread keeps mattering, restate it, or start a dedicated conversation with the brief at the top.
Attached files are read as part of the request rather than stored as knowledge. A file attached to one conversation is not automatically available in another.
Where the knowledge comes from
- Publicly available text — books, articles, documentation, forums — collected by the model providers before a fixed training cut-off.
- Human review and alignment work that shapes tone, safety and instruction-following.
- Supplemental retrieval when a request explicitly asks for current information from the web.
- Your own workspace context, used only to answer the request in front of it.
Known limitations
- A knowledge cut-off: recent events, prices, policies and personnel may be out of date.
- Confident errors: a fluent answer can still be wrong, and the tone rarely signals doubt.
- Arithmetic and precise counting over long inputs can drift; verify numbers independently.
- It cannot see what you did not share: an unmentioned constraint is an unknown constraint.
- It is not a clinician, lawyer or financial adviser, and it is not a substitute for one.
How we measure quality
We track three signals: explicit feedback (the like and dislike actions on a reply), the rating left on the Feedback page, and task follow-through — whether conversations turn into work that gets done. Feedback is reviewed by the team, not only counted.
Benchmarks are useful, and they are also a snapshot. A model can lead a benchmark and still be the wrong fit for your particular work, which is why the feedback loop stays open in the product.
Using it responsibly
Check anything that carries real consequence — legal language, medical information, financial decisions, safety procedures — against a qualified source. Keep a person accountable for the final call.
Do not paste material you are not allowed to share. If you work with regulated data, confirm that your use is permitted before you rely on any AI tool, including this one.
When the model changes
We change the underlying model when a newer one clearly serves the work better. Changes are announced in the product, since a new model can shift tone and structure even when the interface stays the same.