Model providers

The trace analyzer can use the same supported model provider as your agents. Choose it independently from the cloud hosting PIG. For example, an analyzer on GKE with Cloud Storage can use OpenAI or Azure OpenAI.

Support for Anthropic’s Claude models is coming soon.

Supported endpoints

Provider value	Responses API base URL	Authentication
openai	https://api.openai.com/v1	api_key
azure_openai	https://RESOURCE.openai.azure.com/openai/v1 or https://RESOURCE.services.ai.azure.com/openai/v1	api_key
aws_bedrock	https://bedrock-mantle.REGION.api.aws/v1 or the same host with /openai/v1	api_key or aws_sigv4

Replace RESOURCE and REGION with your provider settings. Use a model or deployment name available to your account through that endpoint. Azure uses the model deployment name.

URLs must use HTTPS without embedded credentials, an explicit port, query parameters, or a fragment. Ordinary Bedrock Runtime URLs and arbitrary OpenAI-compatible endpoints are not accepted. Deploying on Google Cloud does not add Vertex AI model support.

Configure API key authentication

Set these fields under spec.analysis and deliver the key through your secret manager:

model:
  provider: openai
  authentication: api_key
  baseURL: https://api.openai.com/v1
  name: REPLACE_MODEL_NAME
  apiKeySecretRef:
    name: pig-credentials
    key: model-api-key

For Bedrock Signature Version 4, use provider: aws_bedrock and authentication: aws_sigv4, and omit apiKeySecretRef. Grant the analyzer’s AWS identity access to the chosen model. Storage permissions alone do not authorize inference.

Plan access, quota, and data handling

The analyzer sends session content and relevant instruction context to your selected model endpoint. Choose a region and provider policy that meet your requirements, as described in Trust and data model.

Reserve quota for analysis alongside your agents’ traffic. Monitor rate limits, latency, and analysis completion time before expanding collection. Changing cluster size does not increase model quota.

The model configuration takes effect with the rest of the analysis settings: activation time, quiet window, and a complete repository identity. Use a real session to verify successful analysis.