Multimodal RAG Platform  ·  View 13 of 16

Observability & Evaluation Matrix

Six signal types across five stages. A matrix rather than a flow, because it makes the gaps visible.

Editable source SVG draw.io All views
Emit
Emit
Collect
Collect
Store
Store
Analyse
Analyse
Act
Act
Service health
Service health
RED metrics
RED metrics
OTel Collector
OTel Collector
Metric Store
Metric Store
SLO dashboards
SLO dashboards
Page on burn rate
Page on burn rate
Traces
Traces
Span per stage
Span per stage
OTel Collector
OTel Collector
Trace Store
Trace Store
Latency breakdown
Latency breakdown
Tune slow stage
Tune slow stage
Retrieval quality
Retrieval quality
Hit rank logged
Hit rank logged
Eval Collector
Eval Collector
Eval Store
Eval Store
Recall@k, nDCG
Recall@k, nDCG
Re-chunk or re-embed
Re-chunk or re-embed
Answer quality
Answer quality
Citation coverage
Citation coverage
LLM Trace
LLM Trace
Trace Store
Trace Store
Groundedness, refusal rate
Groundedness, refusal rate
Prompt or model change
Prompt or model change
Cost
Cost
Tokens per request
Tokens per request
Gateway Meter
Gateway Meter
Cost Ledger
Cost Ledger
Cost per query, per tenant
Cost per query, per tenant
Route or cache tune
Route or cache tune
Safety & privacy
Safety & privacy
Guard verdicts
Guard verdicts
Audit Stream
Audit Stream
Audit Log
write-once
Audit Log...
Injection, PII rate
Injection, PII rate
Policy update
Policy update
Observability & Evaluation Matrix
Observability & Evaluation Matrix
Rows without an Act cell are telemetry nobody owns. Every row here has one.
Rows without an Act cell are telemetry nobody owns. Every row here has one.
v 1.0 · owner SRE & Quality
v 1.0 · owner SRE & Quality
Text is not SVG - cannot display

Beyond service health

  • Retrieval quality — recall@k and nDCG from logged hit ranks
  • Answer quality — groundedness, citation coverage, refusal rate
  • Cost — tokens and dollars per query and per tenant

Every row ends in Act

  • Low recall triggers re-chunking or re-embedding
  • Low groundedness triggers a prompt or model change
  • Cost drift triggers routing or cache tuning

Safety signals

  • Guard verdicts emitted on every request
  • Injection and PII rates tracked as first-class metrics
  • Write-once audit log feeds the SIEM