Feature · Self-Learning Knowledge Base
Agents that get smarter with use
Upload docs once. Agents extract knowledge from real conversations and suggest updates back. The knowledge base grows by itself.
What it does
Most RAG implementations are static — you upload PDFs, the model retrieves chunks, conversations forget. KILN keeps a feedback loop: every conversation gets analyzed for missing or outdated knowledge, surfaced as a suggestion to the agency owner.
When a customer asks something the agent couldn't answer, the system flags the gap. When the agent answered with stale info, that gets flagged too. You review the suggestions, accept them, and the knowledge base updates without writing prompts.
Embeddings refresh automatically when you accept a suggestion. Visitor memory builds in parallel — the agent remembers individual users across sessions, learning their preferences and history without you re-feeding context.
How It Works
Three moving parts
Upload + auto-chunk
PDF, URL, FAQ, or raw text. KILN chunks intelligently, embeds via your BYOK provider, and stores vectors in pgvector with full provenance.
Review suggestions
After every conversation, an analyzer agent surfaces 'gaps' (questions the agent couldn't answer well). You accept, edit, or reject each suggestion in the dashboard.
Watch quality climb
Accuracy metrics per topic update in real-time. Agents pick up new knowledge automatically. No prompt-engineering treadmill.
Use Cases
What agencies actually build with this
Sales-enablement agent
Starts with the product wiki. Learns objection handling from real prospect calls. By month two, it knows the angles your top sales rep uses.
Internal knowledge bot
Onboarded with the company handbook. Suggests updates whenever policy changes are mentioned in slack threads. HR reviews and accepts.
Customer-support tier 1
Initial seed: docs + FAQ. After a month of real tickets, the bot has learned which workarounds the actual support team uses for common issues.
Technical details
- PDF / URL / FAQ / TEXT upload with auto-chunking and provenance tracking
- pgvector embeddings stored per sub-org with org-scoped retrieval
- Conversation analyzer flags knowledge gaps + outdated information
- Suggestion review queue with diff-style accept / edit / reject
- Persistent visitor memory across channels — agent remembers individuals
- Embedding cost attributed per sub-org via BYOK keys
Ready to build with this?
Free forever for testing. Start charging your first client by next week.