ADR-003: LLM-Based Intent Classification

Date: July 8, 2025 Status: Proposed Deciders: Principal Architect, Chief of Staff, CTO

Context

During Claude Code integration, we discovered that Piper Morgan’s regex-based intent classifier is too rigid for natural conversation. Users must use specific command patterns rather than natural language.

Current limitations:

• Cannot understand context (“show that again”)
• No conversation memory
• Cannot resolve references (“that summary”)
• Forces command-like interaction patterns

Decision

We will enhance the intent classifier to use LLM-based classification with conversation memory, enabling natural conversational interactions while maintaining high accuracy.

Consequences

Positive

• Natural conversational flow
• Context-aware understanding
• Better user experience
• Synergy with MCP dynamic capabilities
• Enables multi-turn interactions

Negative

• Additional LLM API calls (cost)
• Potential latency increase
• More complex implementation
• Risk of intent hallucination

Neutral

• Requires conversation state management
• Changes interaction patterns
• New testing strategies needed

Alternatives Considered

1. Enhance Regex Patterns: Would not solve context/reference issues
2. Rule-Based Context: Too brittle, exponential complexity
3. Hybrid Forever: Maintains two systems indefinitely

Implementation Plan

1. Build hybrid system for testing
2. Implement conversation memory
3. Migrate to full LLM classification
4. Optimize for latency and cost

Related Decisions

• ADR-002: Claude Code Integration (revealed this need)
• ADR-001: MCP Integration Pilot (potential synergies)

---

Last Updated: July 09, 2025

Revision Log

• July 09, 2025: Added vertical resize feature to chat window for improved usability