Follow-up context fails for implicit reference queries (e.g., "now give formula")

[ujandey]

Problem

The system fails to resolve implicit follow-up queries that refer to the previously discussed concept.

Reproduction

Input sequence:

1. "Explain Gauss law"
2. "Explain simply"
3. "Explain even simpler"
4. "Now give formula"

Current Behavior

The system asks:
"What formula are you referring to?"

Expected Behavior

The system should infer that "formula" refers to Gauss law and return:

• The mathematical expression of Gauss law
• Optionally a brief explanation

Root Cause (Suspected)

• Follow-up detection relies too heavily on semantic similarity
• Weak handling of implicit references ("this", "that", "formula", "result")
• Missing concept memory in active topic block

Proposed Fix

• Add concept memory to topic block (e.g., current_topic = "Gauss law")

• Improve reference detection:

  • Detect keywords like "formula", "equation", "result"
  • Map them to current topic

• Override similarity threshold when reference_score is high

Priority

High — affects core conversational continuity