Upgrade Path
Principle: progressive enhancement - each step independently delivers value without breaking the running pipeline.
Current status (v0.1.0-SNAPSHOT)
All five main layers + two cross-cutting layers are implemented. L2 / Kafka are off by default and unlock via config.
| Layer | Status |
|---|---|
| Domain data structures (immutable records) | Done |
| Preprocessor - fingerprint generation | Done |
| Analyzer - L1 cache + L2 vector merge + L3 LLM | Done |
| Context optimization (ContextOptimizer) | Done |
| Collector - Logback + HTTP + Kafka | Done |
| Aggregator / Notifier - surge detection + Feishu | Done |
| Cross-cutting - Micrometer metrics + feedback flywheel | Done |
Roadmap
V1.1 - Token metrics
Wire ChatResponse usage metadata into token_cost_total so the cost ladder has real data behind it.
V1.2 - L2 real enablement (PgVector)
Switch PgVectorClusterRepository.findSimilar from in-memory cosine to real vector search once Spring AI supports float[] query input.
V1.3 - Kafka real enablement
Turn on the Kafka collector for async peak-shaving in production.
V1.4 - Redis L1 cache
Replace Caffeine with Redis for cross-instance sharing and persistence.
V2.x - Knowledge base accumulation
Accumulate confirmed root causes into a searchable knowledge base, enabling zero-LLM resolution for known error patterns.
V3.x - Multi-language support
Extend beyond Java to support Python and Go stack traces.