WFGY/ProblemMap/ops/live_monitoring_rag.md

Live monitoring & alerting — RAG services

Goal: list of recommended metrics, alert rules and dashboard panels to keep RAG pipelines observable and actionable.

---

Core metrics to collect (recommended names)

Service-level

• rag_e2e_latency_seconds (histogram) — E2E latency (request in → answer out)
• rag_error_count_total — errors per endpoint
• rag_request_count_total — total requests

Retrieval-level

• retriever_qps_total
• retriever_retrieved_docs_count (per request)
• retriever_empty_result_count_total — unexpected empty sets

Vectorstore

• vectorstore_index_load_time_seconds
• vectorstore_memory_bytes
• vectorstore_indexed_docs_total

Accuracy/provenance

• rag_citation_hit_rate (CHR gauge over sliding window)
• rag_precision_shipped (periodic batch scorer push)
• rag_under_refusal_count_total

Infrastructure

• llm_api_rate_limited_total
• llm_api_error_total
• queue_backlog_count (if using background queues)

---

Suggested PromQL alerts (examples)

Tune thresholds to your workload.

A) Latency breach (interactive)

alert: RAGHighP95Latency
expr: histogram_quantile(0.95, sum(rate(rag_e2e_latency_seconds_bucket[5m])) by (le,instance)) > 2
for: 5m
labels:
  severity: page
annotations:
  summary: "RAG p95 > 2s ({{ $labels.instance }})"

B) Error spike

alert: RAGErrorSpike
expr: increase(rag_error_count_total[5m]) > 50
for: 2m
labels: { severity: page }

C) Retriever empty results

alert: RetrieverEmptyResults
expr: increase(retriever_empty_result_count_total[5m]) > 1
for: 5m
labels: { severity: ticket }

D) CHR drop

alert: CHRDrop
expr: rag_citation_hit_rate < 0.6
for: 10m
labels: { severity: ticket }

E) LLM auth failure

alert: LLMAuthFail
expr: increase(llm_api_error_total{code="401"}[5m]) > 0
for: 1m

---

Dashboard panels (recommended)

1. E2E latency (p50/p95/p99) trend.
2. Requests per second and error rate.
3. Retriever QPS, avg retrieved docs, empty results.
4. CHR & Precision (batch scorer push).
5. Vectorstore memory & disk IO.
6. LLM provider error & rate-limit metrics.

---

Incident play (fast actions)

1. If CHR drop → run diagnostic retrieval for 10 golden queries (retrieved ids + cosine scores).
2. If retriever empty → check vectorstore health and index partitions. Restart index shard if needed.
3. If E2E latency spike with LLM errors → throttle traffic, put a hard rate limit and rollback deploy if needed.
4. If LLM auth failure → rotate key & redeploy secrets.

---

How to integrate scoring metrics

• Periodic scorer job should push rag_citation_hit_rate and rag_precision_shipped as a short-timeseries gauge (per 5–15m window).
• Use batching: run score_eval.py (see ProblemMap/eval/README.md) nightly and push summary metrics via a small exporter.

---

Troubleshooting queries (prometheus examples)

• Check p95 per instance:

  histogram_quantile(0.95, sum(rate(rag_e2e_latency_seconds_bucket[5m])) by (le,instance))
  

• CHR trend:

  avg_over_time(rag_citation_hit_rate[30m])
  

---

Links

• Deployment checklist → deployment_checklist.md
• Debug playbook → debug_playbook.md
• Eval & scoring → ../eval/README.md

---

🔗 Quick-Start Downloads (60 sec)

Tool	Link	3-Step Setup
WFGY 1.0 PDF	Engine Paper	1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>”
TXT OS (plain-text OS)	TXTOS.txt	1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly

---

Explore More

Module	Description	Link
WFGY Core	Canonical framework entry point	View
Problem Map	Diagnostic map and navigation hub	View
Tension Universe Experiments	MVP experiment field	View
Recognition	Where WFGY is referenced or adopted	View
AI Guide	Anti-hallucination reading protocol for tools	View

If this repository helps, starring it improves discovery for other builders.