Transparency
How the assistant is built, and what it will not do
Ask FEFunding is a retrieval and explanation layer over approved knowledge. It is not allowed to improvise FE funding policy, and the pipeline is built so that it structurally cannot.
What happens when you ask something
- Your question is checked for personal information, on our server, before anything leaves it. A learner identifier, a date of birth, contact details, pasted data or an ILR file stops the request there.
- The question is classified for intent and funding year. No model is involved in this step.
- Some questions are answered without a model at all. A glossary term, a diagnostic code or a return date comes straight from structured data, which is faster, free and cannot drift.
- Otherwise we retrieve from reviewed Academy content and approved official sources, filtered by funding year and publication status.
- If the evidence is thin, we stop here and say so. The refusal happens before generation, not after, because a model given thin context produces a confident answer from its own memory. That is precisely what must not happen with funding rules.
- Only then is a model asked to write the answer, from those passages and nothing else, with a hard cap on length.
- The answer is checked before you see it. An answer that cites nothing, states a figure the context never contained, drifts into another funding year, or offers a supplier menu path we have not verified is rejected and the retrieved passages are shown instead.
What it refuses, deliberately
It will not tell you
- A funding rate, threshold or band figure. These are set annually and belong in the official document.
- Anything about a funding year the Academy holds no reviewed content for.
- What a validation rule or report identifier means, unless its official catalogue has been imported.
- Which menu in a student record system to use, unless we have verified it for a stated version.
It will not accept
- Learner names, Unique Learner Numbers, dates of birth or contact details.
- ILR files, data extracts or pasted tables.
- A request to inspect, check or correct a learner record.
Do not enter learner names, Unique Learner Numbers, dates of birth or any other personal information. The Academy is a public knowledge service and does not accept learner data.
How it is tested
A gold set of 170 questions is run before any change to the prompt, the model, the index or the content. 6 of those are questions the Academy must refuse rather than answer, and they matter as much as the positive cases: a system that answers "what is the rate for band 5 this year?" from model memory is exactly the failure this Academy exists to prevent.
Retrieved content is treated as data and never as instructions, and the prompt is fenced accordingly. The current prompt version is recorded against every answer, so feedback can be traced to how it was produced.
Prompt version academy-ask-v1170 gold questions6 required refusals
Why it has limits
The Academy is free to use and privately funded. The assistant has a daily ceiling and an hourly rate limit, checked before any model call rather than after, so the cost stays predictable.
When a limit is reached, the Academy does not break. Search, the error lookup, the glossary, the return calendar, the checklist generator and every page are unaffected, because none of them involves a model. The assistant degrades to showing you the retrieved passages and their sources, which is often what you wanted anyway.
What happens to what you type
Your question is used to produce an answer. Common questions are cached so a repeated question costs nothing, and nothing you type is used to identify you.
Feedback on an answer stores what you tell us and which answer it was about. It does not store your identity, and the personal-information guard runs on the feedback box too. Read the full privacy note.
Often you do not need the assistant at all. The deterministic parts of the Academy are faster and never guess: the error lookup, search, the glossary and the return calendar.