Autonomous knowledge improvement with the Knowledge Flywheel

Autonomous Self-Learning

Vindex Ai features a state-of-the-art Knowledge Flywheel that automatically identifies gaps in your AI chatbot's knowledge. Instead of manual audits to guess what customers are asking, the system tracks confusion points and brings them directly to your dashboard.

How the Flywheel Works

The flywheel runs through three simple phases: Detection, Smart Grouping, and Review.

1. Confusion & Gap Detection

Every conversation is analyzed in real-time for learning signals.
Our system monitors multiple factors to identify when a customer is confused or when the chatbot is missing information:
  • AI Uncertainty: When the chatbot has to say "I don't have information about that" or "I'm not sure."
  • Low Document Match: When the chatbot searches your uploaded PDFs but cannot find any relevant sections.
  • Customer Frustration: When a customer re-words their question several times in a row.
  • Explicit Feedback: When a customer clicks the "thumbs down" button on a chatbot reply.
When a conversation hits these triggers, a Learning Candidate is instantly flagged for your review.
We calculate a weighted Signal Score in real-time based on these triggers:
  • AI Uncertainty (+5 pts): Detected when the AI outputs uncertainty sentinels.
  • Knowledge Gap (+3 pts): Triggered when vector database semantic search returns similarity scores below the similarity threshold.
  • User Frustration (+3 pts): Identified when semantic analysis flags a user rephrasing queries within a single session.
  • Negative Feedback (x2 multiplier): A user "thumbs down" triggers a priority multiplier.
When the combined score reaches 6 or higher, a
text
LearningCandidate
is logged.

2. Smart Grouping of Questions (Clustering)

To keep your review list clean and prevent clutter:
  • Deduplication: If multiple customers ask the same missing question in slightly different ways, Vindex automatically groups them together instead of showing multiple duplicates.
  • Clustering: Related topics are clustered together, allowing you to see the exact variations of how customers are asking for the same missing information.

3. Human-in-the-Loop Review

Your pending candidates appear in your dashboard's Learning Pool. You can teach the AI in three different ways depending on what went wrong:
  • One-Click Approval: If the AI chatbot has draft-written a perfect response, simply click the Checkmark button. The system immediately learns it and adds it to the chatbot's memory.
  • Fill the Gap: If the chatbot completely lacked information, it will show you exactly what is missing. You can write the correct answer directly into your dashboard.
  • Fix Synthesis Failures: Sometimes the chatbot has the facts but fails to put them together nicely. The system will show you the exact files retrieved and let you adjust the draft before approving.

Advanced Productivity Options

Bulk Actions

Review multiple gaps at once. Use the checkboxes in the Learning Pool to:
  • Approve Selected: Teach the AI all selected answers instantly.
  • Archive Selected: Remove irrelevant questions from your list in one click.

Autonomous Auto-Learning (Auto-Approve)

For channels with very high-quality website scrapers and documentation, you can enable Auto-Learning:
  • Answers drafted by the AI that have extremely high confidence and are directly verified by your site content will be learned automatically without requiring admin approval.Candidates with a calculated Grounding Score above 0.85 are auto-approved.
  • The system continually audits these auto-learned entries and alerts you if their effectiveness drops.

Effectiveness & Impact

Once approved, the new information is instantly embedded. The system continues to monitor its real-world performance:
  • Impact Matches: Shows you how many times this new approved fact was successfully used to answer subsequent customer questions.
  • Post-Learning Failures: If users still click thumbs down on the new fact, the score drops, alerting you to refine the text.

How to Enable Self-Learning

  1. Go to your Site Settings (or Channel Settings).
  2. Toggle on Autonomous Learning.
  3. (Optional) Toggle Auto-Approve Verified Gaps for full hands-off automation.
  4. Visit your Learning tab in the dashboard to review pending candidates!
PreviousChannel SandboxNextTools Overview