Tutor content
The Knowledge Base page holds everything the tutor draws on, in four tabs: Corpus, Glossary, Personas and Prompts.
The tutor needs an OpenAI API key. Corpus documents and glossary entries are searched from the vector store set by OPENAI_VECTOR_CORPUS_ID, and personas from the one set by OPENAI_VECTOR_PERSONAS_ID. Without those, the content is kept but the tutor can't search it. See Deploying.
Corpus
Reference documents the tutor can search and cite: guides, policies, scam reports. Upload them with Upload File(s). Accepted formats are Markdown, plain text, PDF, Word (.docx), JSON, JSON Lines and YAML, up to 50 MB each.
Each file is sent to OpenAI and added to the vector store. Replacing a file's content uploads the new version and retires the old one.
Glossary
ScamSafe's glossary of scam types: what each one is, its warning signs and what to do. The tutor uses it to recognize and explain scams.
On a partner deployment the glossary comes from GCA and is updated automatically. Your own edits are kept: an entry you edited keeps your version, an entry you deleted stays deleted, and new entries from GCA are added. Re-import seed discards your changes and returns every entry to GCA's current version.
Personas
Profiles of typical learners, which the tutor uses to pitch its answers at the right level. They behave like the glossary: updated from GCA on a partner deployment, with your edits kept. Re-import YAML returns to GCA's set.
Sync status
Corpus files, glossary entries and personas each show a status:
| Status | Meaning |
|---|---|
synced | In the vector store. The tutor can find it |
localPending | Saved in ScamSafe, not yet sent to OpenAI |
replacePending | Edited; the new version hasn't reached OpenAI yet |
uploaded | Not yet usable. The file reached OpenAI but couldn't be added to the vector store, so the tutor can't find it. The reason is shown on the item |
error | The upload failed. The reason is shown on the item |
Only synced means the tutor can use an item. Sync all (Sync All Pending for the corpus) retries everything that isn't synced, which is the fix after a problem such as a missing vector store ID or a lapsed OpenAI account has been resolved.
Prompts
The instructions given to the AI for each task: the tutor itself, interpreting free-text answers, explaining a recommendation, understanding a learner's reason for visiting, and summarizing long conversations.
An edit takes effect on the next request, with no restart. Reset returns a prompt to the version that ships with ScamSafe.
Edit with care:
- Free-text interpretation feeds scores. A change that makes it misread answers changes learners' results.
- The tutor prompt ends with two safety notices, on lines starting
REPORTING:andRECOVERY:. ScamSafe shows these to a learner word for word, at most once each, when the tutor flags that one applies. The tutor never sees or rewrites them. Keep each on a single line, and have any change to them reviewed: they are written for people who may have just been defrauded.