Engine/RAG Dashboard

Purpose and audience

• Purpose: provide one operational overview for the online query runtime and the async indexing pipeline.

• Audience: engineers and operators diagnosing latency, failure rate, queue pressure, and ingestion backlog.

• Source of truth: monitoring/grafana/dashboards/engine-rag.json

• Live dashboard: Grafana Engine/RAG dashboard for RagLogic AI

Provisioned panel inventory

• Overview and backend health

• RAG runtime health and collection availability

• Query rate by mode and by outcome

• Query latency by mode and phase latency p95

• Provider and retrieval error breakdown

• Queue depth

• Errors per minute

• Traffic

• Pipeline progress and counters

• Relation sync and extraction status

• Worker throughput and latency

Current panel-to-signal mapping

Expected signal	Current coverage	Notes
Gateway backend health	Present	lalandre_api_gateway_backend_health
Gateway traffic	Present	Query and search rates are graphed
Gateway proxy errors	Present	Proxy error rate is graphed
RAG backend health	Present	lalandre_rag_service_backend_health including collections
RAG query modes	Present	lalandre_rag_service_query_requests_total broken down by mode
RAG outcomes	Present	lalandre_rag_service_query_requests_total broken down by outcome
Query latency by mode	Present	lalandre_rag_service_query_duration_seconds_bucket
RAG phase durations	Present	lalandre_rag_service_phase_duration_seconds_bucket
Provider errors	Present	lalandre_rag_service_provider_errors_total
Retrieval errors	Present	lalandre_rag_retrieval_errors_total
Queue depth	Present	Redis key size panels
Worker throughput	Present	Chunking, embedding, extraction job rate panels
Worker latency	Present	Histogram quantiles for worker durations
Extraction backlog and relation sync	Present	Extraction status and graph sync panels

Validation notes

The dashboard now exposes the core runtime signals that were previously missing:

• canonical query modes such as rag, llm_only, summarize, and compare,

• response outcomes including grounded, weakly_grounded, clarify, and hard_block,

• request latency by mode,

• phase latency via lalandre_rag_service_phase_duration_seconds,

• provider failures, retrieval failures, and rag-service backend health.

This validates the dashboard for the current runtime metrics surface described in the codebase and in the Sphinx operations pages.

Completion checklist

• [x] Provisioned dashboard JSON is versioned in the repo.

• [x] Existing panels are inventoried.

• [x] Runtime metrics have been compared with panel coverage.

• [x] Query mode coverage is complete.

• [x] Query outcome coverage is complete.

• [x] Phase timing coverage is complete.

• [x] Provider and retrieval error coverage is complete.

• [x] RAG backend health coverage is complete.