Knowledge quality
Write sources a student could trust
A clean document is an act of care. Headings, dates, and conditions help the assistant answer the question that was asked, and help you notice when a fact has gone stale.
Organize around real visitor questions
Prepare a focused source for each kind of question: courses and fees, entry requirements, how to apply, intakes, scholarships, and student services. Give every source an owner and a date when someone will read it again.
Write each section so it still makes sense alone. Name the programme, campus, who it is for, the intake, and the date it applies. A retrieved passage may be the only thing the assistant sees. “The course above” or “same as last year” leaves a student guessing.
- Use a heading a person could search for, such as “BSc Data Science — 2027 intake — fees”.
- Say the currency, whether a fee is annual or total, the conditions, and the date it applies.
- Keep policy separate from estimates, exceptions, and decisions a person has to make.
- Settle contradictions before you upload. More files do not make a kinder answer if they disagree.
- Leave out passwords, internal keys, and anything you would not want repeated to a visitor.
Choose a clean format for the material
| Format | Best practice |
|---|---|
| Use selectable text. Preserve reading order and table headers. Check columns and multi-page tables after export. Convert scans before upload because OCR is off. | |
| DOCX | Use actual heading styles, simple tables, and plain paragraphs. Remove comments, tracked changes, and obsolete appendices from the training copy. |
| CSV | Use one record per row, descriptive unique headers, UTF-8 encoding, and properly quoted commas/newlines. Export spreadsheet workbooks to CSV first. |
| Markdown / TXT | Use short related paragraphs and meaningful headings. Include units and context in the section instead of relying on its position in the file. |
| HTML / public pages | Choose a substantive content page. Avoid account portals, navigation-only pages, and screens whose meaningful content is unavailable without client-side interaction. |
A CSV that explains its own values
Keep the programme, campus, intake, how the fee is counted, and the conditions on every row. Empty or repeated headers, and extra values that are not quoted, can fail. A single enormous cell can be too long to extract. Split it into records a person could read. The numbers below are a shape to copy, not your institute’s policy.
programme,campus,intake,fee_inr,fee_basis,eligibility
"BSc Data Science","Chennai","2027",120000,"per academic year","Class 12; confirm subject requirements with admissions"A self-contained policy section
Scholarship policy — 2027 intake
Applies to: BSc Data Science, Chennai campus Effective date: 1 January 2027
Students may apply after submitting verified academic records. The scholarship committee reviews eligibility and available funding. A chatbot response is not an award confirmation. Contact admissions for an individual decision.
Check knowledge after every update
- 1
Create an answer checklist
For each source, write the questions it should answer and the facts you expect to hear back.
- 2
Test conditions and exceptions
Ask about another intake, a programme you do not offer, missing documents, and eligibility that is not a simple yes.
- 3
Inspect references
If an answer cites a source, read the passage. A citation is a pointer, not proof.
- 4
Test a fresh visitor journey
Check a factual answer, the contact form, and a request to speak with someone. Compare them with the originals you approved.
- 5
Record the review
Keep the version, the owner, the review date, and open questions outside the product. Update the source that is wrong. Do not stack a second copy beside it.