This article details a resolved issue within the Kore.ai platform concerning the Knowledge Graph (KG) engine’s proximity match functionality for Few-Shot KG.
Problem Description:
The issue manifested as returning false positive matches during utterance testing, even when the proximity value in the KG engine was set to 5%. This meant that intents with a higher difference than 5% were still being identified as matches.
Replication:
- The problem consistently occurred (always) during utterance testing.
- Steps to replicate:
- Navigate to “Build” > “Natural Language” > “Threshold & Configurations.”
- Select “Knowledge Graph” and then “Proximity of Suggested Matches.”
- Set the value to 5%.
Analysis:
The investigation revealed that the bot was qualifying multiple matches despite the proximity score between the intents exceeding the 5% threshold configured in the Few-Shot KG.
Resolution:
Identified the root cause as the system not using proximity for upper thresholds matching within the Few-Shot KG.
A fix was implemented that enabled proximity for upper thresholds, ensuring matches fall within the configured range.
Impact and Availability:
- This issue potentially impacted intent identification accuracy in Few-Shot KG scenarios.
- The fix is available in Kore.ai platform version 10.1.22.
- Users on versions prior to 10.1.22 might experience the described behaviour. Upgrading to the latest version is recommended to benefit from the fix.