Streaming and Phases

The frontend trace model is built from structured SSE events. The most important runtime contract is that phases remain understandable and auditable.

Main phases

Phase	Meaning
routing	Question qualification and profile selection
planning	Agentic planning, clarification, and retrieval shaping
retrieval	Search, fusion, rerank, and result shaping
crag	Sufficiency evaluation and refinement
enrichment	Context and graph enrichment
complementary	Additional retrieval for deeper search modes
compression	Context reduction before generation
generation	LLM answer generation
citations	Citation validation and repair
cypher	Graph support branch activity

Metrics linkage

Phase timings are exported through lalandre_rag_service_phase_duration_seconds. The metric can also contain normalized names derived from:

• phase_timings_ms

• graph_query_phase_timings_ms

• graph_retrieval_phase_timings_ms

Why the phase chart matters

The Engine/RAG dashboard should expose these phases clearly enough for:

• runtime bottleneck detection,

• regression analysis after prompt or retrieval changes,

• explanation of user-visible latency in the chat trace.

That coverage is currently incomplete and is tracked in the dashboard audit pages.