Gemini

What it gives you — the investigation loop driven by Google’s Gemini API (native streamGenerateContent function calling, not an OpenAI-compatible shim) — version-gated, see below before you wire it.

Gemini 3.x does not work

Do not configure a Gemini 3.x model. RunLore’s Gemini client does not capture or replay the thought_signature that Gemini 3 requires on functionCall parts. An investigation gets through the first turn, then fails on the second turn with:

400 INVALID_ARGUMENT: Function call is missing a thought_signature

This was reproduced, not inferred — every multi-turn investigation on a Gemini 3.x model fails this way. Tracked as issue #392.

Gemini 2.5 works. Verified with gemini-2.5-flash-lite.

Minimal config

model:
  provider: gemini
  model: gemini-2.5-flash-lite      # verified working — do NOT use a gemini-3.x model
  api_key_env: GEMINI_API_KEY       # base_url defaults to https://generativelanguage.googleapis.com

Verify it locally

kubectl -n runlore create secret generic runlore-secrets \
  --from-literal=GEMINI_API_KEY='<key>'

Fire a test incident (see Alertmanager or hack/demo.sh) and watch a multi-turn investigation (more than one tool call) complete rather than fail on its second turn:

kubectl -n runlore logs deploy/runlore | grep -E 'msg=findings|msg=incident|INVALID_ARGUMENT'

Notes

base_url is optional — unset, it defaults to https://generativelanguage.googleapis.com.
Forced tool calls (verify pass, recall reranker, eval judge, semantic KB validation, kb import --model) use Gemini’s functionCallingConfig: {mode: "ANY", allowedFunctionNames: […]} — the provider-native equivalent of OpenAI’s forced tool_choice. This works correctly on Gemini 2.5; the thought-signature issue above is what breaks Gemini 3.x specifically, not forced calling in general.
effort is rejected at startup for this provider. Gemini’s thinkingConfig has replay semantics (thought signatures) the provider-agnostic message history can’t carry — the same class of problem behind the 3.x incompatibility above, just enforced up front instead of surfacing as a runtime 400.
thinking: adaptive is Anthropic-only and is likewise rejected for this provider.
RunLore relies on Gemini’s automatic implicit prefix caching (enabled on Gemini 2.5+): the request prefix (system instruction + tools + earlier turns) must stay byte-stable and append-only across the loop’s steps, which the client guarantees.
A common non-fatal ReAct pattern: Gemini sometimes concludes an investigation in prose instead of calling submit_findings. The loop nudges it once, then falls back to a synthetic inconclusive result rather than hanging — same fallback described on OpenAI-compatible.
A cleartext API key is rejected on a public http:// host — use https:// or a loopback/in-cluster address.

Reference

Configuration → model for the full key reference.

Grafana Alerting Grafana Loki