Knowledge Base

The AI draws from two knowledge sources when generating responses:

1. Admin-learned knowledge

Short, admin-verified Q&A pairs learned from ticket conversations. Built over time by:

• Admin replies in tickets (with enableAdminLearning on) — the bot proposes candidates
• Admins review candidates in AI → Öğrenme Kuyruğu
• Approved candidates become "admin precedent" — highest priority in AI responses

2. General knowledge files

.txt files uploaded under AI → Dosyalar. Used as factual context for AI answers.

Best practices:

• Keep files focused (one topic per file)
• Use clear headings and short paragraphs
• Remove outdated info — the bot will cite stale data
• Turkish content is fine — the AI auto-translates per user language

Priority order

1. Admin precedents (highest) — override general knowledge
2. General knowledge files — baseline context
3. Recent announcements — timestamped (older than 7 days gets a "past announcement" marker)
4. User memory — recent conversation context
5. User profile — history with this user

FAQ cache

Common questions get cached for 48 hours (configurable). Cache hits skip the AI call entirely — saves cost and latency.

Cache invalidation happens when:

• Any .txt file in knowledge/ is modified
• You manually clear via /clear-faq command
• The TTL expires

Learning filter

Not every admin reply becomes a learning candidate. The bot filters out:

• Messages with links (discord.gg, http)
• @everyone / @here mentions
• Very short replies (< 20 chars)
• Pure greetings ("tamam", "ok", "teşekkürler")
• Replies with > 3 sentences (too complex)

This ensures the knowledge base stays focused and high-signal.