LLM Configuration¶

Aotanami uses LLMs (Large Language Models) for intelligent diagnosis and remediation. You bring your own API keys.

Supported Providers¶

Configuration¶

Via AotanamiConfig CRD¶

apiVersion: aotanami.com/v1alpha1
kind: AotanamiConfig
metadata:
  name: default
spec:
  llm:
    provider: openrouter
    model: "anthropic/claude-sonnet-4-20250514"
    apiKeySecret: aotanami-llm
    temperature: "0.1"
    maxTokensPerRequest: 4096
  tokenBudget:
    hourlyTokenLimit: 50000
    dailyTokenLimit: 500000
    monthlyTokenLimit: 10000000
    alertThresholdPercent: 80
    enableCaching: true
    batchingEnabled: true

API Key Secret¶

kubectl create secret generic aotanami-llm \
  --namespace aotanami-system \
  --from-literal=api-key=<YOUR_API_KEY>

Cost Optimization¶

Aotanami is designed to minimize LLM API costs:

1. Local Triage First¶

Most events are handled locally without any LLM call. The correlator deduplicates, scores severity, and filters before escalation. Only novel, complex incidents reach the LLM.

2. Prompt Caching¶

Repeated analysis patterns use cached prompt templates. If the same type of issue recurs, the cached response is reused.

3. Structured Output¶

All LLM calls use JSON structured output schemas, getting machine-parseable responses on the first attempt without re-prompting.

4. Batching¶

Multiple related findings are batched into a single LLM call, reducing per-request overhead.

5. Token Budgets¶

Configure hard limits on token consumption:

When a budget is exhausted, Aotanami falls back to rule-based detection only (no LLM) until the budget resets.

6. Monitor Usage¶

Check token consumption via:

kubectl get aotanamiconfigs default -o jsonpath='{.status.tokenUsage}'

Or via the dashboard's LLM Usage view.

Recommended Models¶