[7.2] Instrument Glow CI with Cloud Logging + connect to Vertex AI Evaluation Service

[jasminetay-moe]

Epic: Part 7 — AI Evaluations
Priority: P1
Role: Engineer

User Story

As an Engineer, I want to instrument Glow CI LLM calls with Cloud Logging and connect to Vertex AI Evaluation Service so that all RAG retrievals and Gemini responses are observable and Vertex AI can run LLM judge evaluations.

Context

All Glow CI LLM calls (RAG retrieval + Gemini synthesis) are logged via Cloud Logging on GCP and fed into Vertex AI Evaluation Service for automated LLM judge evaluation. This replaces any third-party observability tooling — we stay fully within the GCP stack. Sprint timing: chat must be built first so there are real LLM calls to instrument.

Acceptance Criteria

• All Glow CI LLM calls instrumented via Cloud Logging: RAG retrieval queries and Gemini synthesis responses captured
• Logs structured with correct span metadata (query, retrieved chunks, synthesised response, latency)
• Vertex AI Evaluation Service configured and connected — logs feed automated eval pipeline
• Eval metrics defined and running: at minimum groundedness, coherence, citation coverage
• Smoke test: end-to-end trace from teacher query → RAG retrieval → Gemini response visible in Cloud Logging and evaluated in Vertex AI

Dependencies

• Blocked by: Epic: Part 7 — AI Evaluations #53 (7.1) — eval requirements must be agreed first
• Blocked by: Part 2 chat build — need live LLM calls to instrument

---

📄 PRD: Part 7 — Glow CI PRD