Sangeetha-Grantha

Metadata	Value
Status	Completed
Version	2.0.0
Last Updated	2026-07-10
Author	Sangeetha Grantha Team

Goal

Batch 3c — Upgrade the Gemini SDK google-genai 1.34.0 → 2.9.0 (major) in the Python extraction/enrichment worker. Likely breaking surface in the worker’s LLM calls; review before raising the floor.

Scope (tools/krithi-extract-enrich-worker/pyproject.toml + uv.lock)

• google-genai floor >=1.0.0 (resolved 1.34.0) → >=2.0.0 (resolved 2.9.0).

Affected code

• Worker LLM integration (Intelligence-in-Python layer) — structured output / batch calls added in TRACK-107.

Implementation Plan

1. Review the google-genai 1.x → 2.x changelog for breaking changes (client init, request/response shapes, structured-output API).
2. Bump the floor in pyproject.toml; uv lock.
3. Update worker LLM call sites and Pydantic response models as needed.
4. Run worker pytest; smoke-test an extraction+enrichment pass against a sample source.
5. Sync current-versions.md (Python tools section).
6. Commit per commit-policy.

Risks

• Major SDK jump (1.x → 2.x) — API changes expected; isolated to the Python worker.
• Pairs conceptually with TRACK-107 (AI Platform Lifecycle Uplift) — review that track’s structured-output/batch usage.

Completion (2026-07-10)

Per the upstream changelog, google-genai 2.0’s breaking changes are Interactions-API-only (“GenerateContent usage is unaffected”) — the worker’s surface (genai.Client(api_key=), models.generate_content + GenerateContentConfig(response_mime_type, response_schema), response.text, batches.create) needed no code changes. Floor raised to >=2.0.0, resolved 2.9.0. Verified: all 144 worker tests green, including the TRACK-113 W3 integration tests that drive the real 2.x SDK through respx-stubbed HTTP (client init → request → response parsing). Live-key smoke test not run — no SG_GEMINI_API_KEY available in this environment; the HTTP-stubbed round-trip is the deepest offline verification.