Autonomous knowledge improvement with the Knowledge Flywheel

Autonomous Self-Learning

Vindex Ai includes a state-of-the-art Knowledge Flywheel that automatically identifies gaps in your AI's knowledge and suggests entries to fill them. Instead of manually searching for what your AI doesn't know, the system brings those gaps to you.

How it Works

The flywheel operates in three distinct phases: Detection, Deduplication, and Review.

1. Signal Detection

Every conversation is analyzed in real-time for learning signals. We use multiple weighted factors to calculate a "Signal Score":
  • AI Uncertainty (+5 pts): Detected when the AI uses phrases like "I don't have information about that" or "I'm not sure."
  • Knowledge Gaps (+3 pts): Triggered when the vector search returns low similarity scores (meaning no relevant documentation was found).
  • User Frustration (+3 pts): Identified when a user rephrases their question multiple times within a single session.
  • Explicit Feedback (x2 multiplier): A user "thumbs down" on a message significantly boosts the signal score.
When a score reaches 6 or higher, a "Learning Candidate" is flagged for your review.

2. Semantic Deduplication & Clustering

To prevent your review pool from becoming cluttered, Vindex Ai uses vector similarity to group related questions:
  • Deduplication: If a similar question has already been flagged, we increment its "Occurrence Count" and boost its "Priority Score" instead of creating a new entry.
  • Clustering: Related questions are grouped into clusters, allowing you to see the different ways users are asking about the same missing information.

3. Human-in-the-Loop Review

Learning candidates appear in your site's Learning Pool. As a human reviewer, you have two modes of operation:

Mode A: Filling Knowledge Gaps

If the AI completely lacked information, it will provide a "Knowledge Gap" report identifying exactly what is missing. You can then write the definitive answer directly in the dashboard.

Mode B: Fixing Synthesis Failures

Sometimes the AI has the right information but fails to synthesize it correctly. In these cases, the system shows you exactly which chunks were retrieved and provides a grounded draft. You simply verify or tweak the draft and approve.

4. Source Traceability

In the review form, any retrieved knowledge chunks include direct links back to their source files or tools. If a hallucination was caused by a specific scrap or tool, you can jump directly to the editor to fix the root cause.

Advanced Productivity Tools

Bulk Management

Don't review one by one. Use the multi-select checkboxes in the Learning Pool to:
  • Approve Selected: Instantly learn all selected gaps.
  • Archive Selected: Remove irrelevant gaps from your view in one click.

Autonomous Auto-Learning

For sites with high-quality documentation, you can enable Auto-Learning. When active:
  • Candidates with a high "Grounding Score" (above 0.85) are automatically learned without human intervention.
  • The system periodically audits these auto-learned entries for effectiveness.

Approving & Impact

When you click Approve & Learn, the new entry is instantly embedded and added to your site's knowledge base.
The system continues to track the Effectiveness Score of every learned entry:
  • Impact Matches: How many times this new entry was used to answer future questions.
  • Post-Learning Failures: If users still give thumbs down after learning, the effectiveness score drops, alerting you that the entry may need further refinement.

Enabling Self-Learning

To enable and manage this feature:
  1. Go to your Site Settings.
  2. Toggle on Autonomous Learning.
  3. (Optional) Toggle Auto-Approve Verified Gaps for full automation.
  4. Visit the Learning tab in your site dashboard to review pending candidates.
[!NOTE] If Self-Learning is disabled for a site, you will see a status alert in the Learning Dashboard. Be sure to enable it in settings to resume gap detection.
PreviousChannel SandboxNextTools Overview