[Enhancement] Configurable Confidence Threshold and Intent Extraction Improvements

[amitb0ra]

Description

Currently, the system uses a hardcoded confidence threshold to validate model responses. This lack of flexibility prevents users from tuning the app for their specific needs (precision vs. recall). Additionally, smaller models like llama-3.1-8b-instant are currently struggling to consistently extract user intent under these fixed conditions.

Current Behavior

• Hardcoded Threshold: The confidence check is locked at 75.
• Inflexible Precision: Users cannot lower the threshold for more "creative" results or raise it for mission-critical accuracy.
• Model Inconsistency: Specifically with llama-3.1-8b-instant, the intent extraction often fails to meet the hardcoded 75% bar, causing the app to skip valid actions or fail to respond to the user.

Proposed Changes

• Settings Integration: Add a new configuration field in the app settings (e.g., ID_CONFIDENCE_THRESHOLD) allowing users to set a value between 0 and 100.
• Dynamic Validation: Replace the hardcoded 75 in the codebase with the value retrieved from the user settings.
• Intent Extraction Optimization: Review the prompt template used for intent extraction to ensure better compatibility with the llama-3.1-8b-instant model.