How SmartKMS uses AI
Last updated: November 16, 2026
SmartKMS uses AI to help you find, read and ask about documents. Every AI model runs on the university's own server, and anything written by AI is labelled. AI never makes decisions about people, and a person can always review what it did to your documents.
Where SmartKMS uses AI
| Feature | What the AI does | Model | Who checks it |
|---|---|---|---|
| Q&A Assistant | Writes answers from the documents you're allowed to open, marks which ones it used, suggests follow-up questions and names your conversations. | qwen2.5:3b | You: check the sources, open Why this answer?, rate the answer. |
| AI Search and the Q&A's document search | Finds documents by meaning, not only by matching words. | nomic-embed-text | You: each result says how it was found. |
| Summaries and key points | Summarizes each document when it's uploaded or edited. | qwen2.5:3b | Labelled AI-generated in the document viewer. |
| Topic tags and folder suggestions | Suggests tags and a folder for a new document. | qwen2.5:3b, nomic-embed-text | The uploader, who can change them. |
| Reading images and scans | Reads the text in images and scanned pages so they can be searched. | llava | The uploader. |
| Upload screening | Checks new and edited documents for exposed personal data, explicit images and graphic violence before they're shared. | Text: qwen2.5:3b. Images: NudeNet and llava | Admins and staff. The uploader sees How this was decided and can ask for a person to review it. |
| Chat safety check | Stops messages that are harmful, using a word filter and a language model. | Word filter and qwen2.5:3b | Administrators see stopped and flagged messages in the Activity Log. |
The model names above are the ones this server is set up to use.
What the AI sees, and where it runs
- On the university's server. The models run locally through Ollama. Your questions and documents are not sent to an outside AI service.
- Only what you can open. When you ask the Q&A Assistant a question, it is given only documents you're allowed to open, and only the parts that relate to your question.
- Not trained on your data. The models are not retrained on SmartKMS documents or conversations. Your thumbs up and down are used in two ways only: as examples the next time you ask something similar, and to move rated documents up or down in your organization's results.
- Kept as the Terms say. Conversations stay in your history until you delete them (or after 24 months without use). See Data Retention in the Terms.
How to see what the AI did
- AI-generated labels mark answers and summaries written by AI.
- Why this answer? under each Q&A answer shows whether it came from documents or from the model's general knowledge, which model wrote it, every document it was given with how well it matched, which ones it cited and how that was decided, and whether earlier ratings shaped it.
- Small numbers in an answer show which document each statement comes from. Click one to open that document.
- How this was decided, in a document's Approval section, lists each automatic check, what it found and whether a person made the final decision.
- Each AI Search result says how it was found: by meaning, title, tag, author or the words in its text.
What the AI can get wrong
- Answers can be wrong, out of date or incomplete, or credit the wrong document. The models are small enough to run on the university's own hardware, which makes them less accurate than large online services.
- For long documents the Q&A Assistant reads the parts that seem most relevant, so it can miss something said elsewhere in the document.
- When no document matches closely, an answer comes from the model's general knowledge. Why this answer? says so.
- Screening can flag a legitimate document or miss one that should be flagged.
- Summaries and tags can misread a document, especially scans and images.
Check important answers against the documents themselves before relying on them.
People stay in charge
- No decisions about people. SmartKMS is not used to admit, grade, discipline, hire or assess anyone, and its answers must not be the only basis for such a decision (Terms, section 8.1).
- Uploads. A rejected upload can be sent for Request Reapproval, which an administrator or staff member reviews. If a check can't run, the document waits for a person instead of being approved. A person's approval or rejection always overrides the checks.
- Answers. You can stop an answer while it's being written, rate it, and delete the conversation.
If you disagree with an AI result
- An answer: rate it with a thumbs down, and check its sources.
- A rejected document: use Request Reapproval in the document viewer.
- Anything else, including questions about your data: contact your SmartKMS administrator. Your rights are in the Terms, Your Data Subject Rights.
About the EU AI Act
The EU Artificial Intelligence Act (Regulation (EU) 2024/1689) is European Union law. SmartKMS is used in the Philippines and isn't subject to it, but it follows the Act's transparency principles as good practice:
- People know when they're dealing with AI (Article 50): the Q&A Assistant says it's an AI the first time you open it and in its header, and AI-written content is labelled.
- Clear information about the AI (in the spirit of Article 13): this page describes what each model does, what it sees and its limits.
- Human oversight (in the spirit of Article 14): people review automated screening decisions and can override them.
- Explanations of individual results: Why this answer? and How this was decided.
SmartKMS is not used for admission, grading, assessing students or monitoring exams, the uses the Act treats as high-risk in education (Annex III, point 3). This page describes how SmartKMS works. It is not a legal assessment or a claim of compliance.