🏗️ Architecture: How should Vitto learn user communication style and adapt pe...

Date: 2026-03-27
Time: 19:45:57 UTC
Session: gspace1-277.5
Agent: angular
Confidence: 90%
Status: ✅ Documented

📋 Decision Context

Question:

How should Vitto learn user communication style and adapt persona parameters?

Task: S277.4: Implement Adaptive Persona Learning

🎯 Decision Made

Chosen:

Heuristic-based CommunicationSignalExtractor (no LLM) + EMA-updated dossier profile + 8 rule-based adaptation paths in DossierPersonaAdapter + data-driven language injection in system prompt

Options Considered:

1. LLM-based signal extraction (accurate but slow/expensive)
2. Manual user preference form (no automatic learning)
3. Full persona swap instead of individual trait adjustment

💭 Reasoning

LLM-based extraction would add latency and cost to every message. Regex heuristics for language/formality/verbosity/emoji are ~90% accurate and run in <1ms. Exponential moving average (alpha=0.1) prevents sudden preference flips. 8 declarative rules are maintainable and testable. Suppression logic (2 rejections = 30-day suppress) prevents nagging. Language injection in prompt section 2 (stable per session) maximizes LLM cache hits.

📁 Files Involved

• engine/context-assembly/communication-signal-extractor.service.ts
• engine/dossier-persona-adapter.service.ts
• engine/context-assembly/channel-prompt-builder.service.ts
• engine/channel-processing.service.ts
• users/schemas/user-dossier.schema.ts

🔗 Related

• Decision ID: dev_1774640757404_czgn5wm
• Session: 277
• Workspace: gspace1

Auto-generated by MCP Smart Consolidation