Knowledge candidates, snippets and search
Tickessa separates expert facts, writing aids and derived search data. Changing an AI provider or search model must not lose reviewed original content.
Candidates from replies
A knowledge candidate is an unpublished suggestion derived from a reviewed AI reply.
- Generate an AI reply draft in a ticket.
- Review and adopt or edit it.
- Select Use as knowledge on the reviewed suggestion.
- Tickessa removes identifiable email addresses and case-specific details.
- Find the candidate under Knowledge & snippets → Knowledge candidates.
This sends no email and publishes no article. If anonymisation cannot be confirmed safely, no candidate is saved.
Review a candidate
Check whether the statement is generally correct, identifying details are removed, project and product version match, and prerequisites, limits and safe questions are complete.
Create as draft creates an internal article with pending review. Reject records rejection. Conversion still requires a later approval and publication process.
AI drafts from several resolved conversations
AI knowledge draft from resolved conversations deliberately derives recurring solutions. Resolving a ticket does not start it automatically.
- Select the project.
- Select one to eight resolved or closed conversations with both an incoming message and a solution recorded as sent.
- Choose a new article or an update to an existing published version.
- Set the intended future audience. Public is a review intention, not publication.
- Select Suggest draft … and confirm the provider-call notice.
Before AI processing, Tickessa removes known contact details, emails, ticket numbers, URLs and phone numbers. Typical salutations, signatures and appended quoted email histories are stripped from problems and answers. AI receives only these minimised problem/solution pairs and, for updates, the approved target version.
Cleaned ticket excerpts remain untrusted customer data and are checked for manipulation. Only an existing approved target article is a reviewed reference. Ticket text cannot change safety rules, approval or publication.
When no AI call occurs
Local preflight stops when projects differ, categories or product versions conflict, a supported problem or solution is missing, solutions contradict one another, or the target lacks an appropriate approved version.
Where safe to save, a Review issue candidate records the unresolved points. It cannot be adopted and provider calls remain zero. Select compatible sources or reject the warning candidate.
Review generated candidates
Each candidate shows project and source tickets, new/update intent, future audience, uncertainties, cited and uncited sources, provider/model/run/prompt/context versions and anonymisation results.
Show anonymised review basis displays the exact cleaned problem and solution excerpts, not raw tickets. Contact details, ticket numbers and reply signatures have already been removed. Technical issues such as missing product versions or conflicting solutions appear as readable explanations.
For updates, Changes compared with … shows a line diff: − is the approved version and + the suggestion. The approved content remains unchanged.
Create as draft creates internal version 1 for a new article or a new internal version of an existing article. Both remain Review pending. Correct, approve and, only when intended, publish through the normal workflow.
AI generation, candidate review and conversion cannot publish. Even a candidate intended for a future public audience is initially stored internally.
Budgets, outages and provider changes
Generation requires a budgeted Knowledge draft task. Only an explicitly approved fallback can trigger a second run for an allowed error. Daily/monthly budgets, emergency stops and call limits apply. Disabled tasks, exhausted budgets and unavailable providers leave no half-finished candidate. Manual article creation remains possible.
A provider change affects only who generates the suggestion; project boundaries, internal drafts, review and separate publication stay the same.
Reply snippets
Snippets are reusable wording, optionally project-specific or cross-project, with a shortcut. They send nothing automatically, do not replace reviewed knowledge, must contain no customer case data or secrets, and need checking in each reply's context.
Version 0.8.0 shows existing snippets with title, preview, project and shortcut. Further editorial management is developed separately.
Local search index
Original articles and snippets stay in your MariaDB installation. Search indexes and future embeddings are rebuildable derivatives.
Local search index shows availability, generation count, number of approved documents, last complete rebuild and the index/embedding model.
Administrators can select Rebuild safely to replace only derived data. Original articles, versions, sources, reply profiles and quality feedback remain intact.
Which sources may AI use?
The server requires sources to belong to the ticket's project or be explicitly cross-project, be approved and published, suit the audience, be within validity dates, match the known product version and have no unresolved review-date or conflict warning.
The interface shows the actual offered source versions. A model citing another source is rejected. Without suitable knowledge, a draft may ask questions but cannot assert an unsupported solution and always requires manual review.