AI Powered Features Within Langfuse (Beta)

Langfuse offers AI-powered features within the product to simplify and speed up your workflows. These features are currently in beta.

Opt-in

All AI-powered features require an opt-in; they are not enabled by default. The opt-in is managed at the organization level and can be changed by an admin or owner. Admins and owners can also separately enable or disable AI feature tracing for product and service improvement. This does not disable the AI features themselves.

The setting applies to all users and projects within the organization.

Available AI Features

Non-exhaustive list of available AI features in Langfuse:

• Langfuse Assistant for asking questions about traces, observations, and metrics in plain language. Available on Cloud and on self-hosted deployments with an instance-wide Langfuse AI model.
• Ask AI in the filter search bar. Available on Cloud and on self-hosted deployments with the same Langfuse AI model.
• AI-powered generation of evaluator names and descriptions. Available on Langfuse Cloud only.

Data Privacy

On Langfuse Cloud, AI features use models hosted on AWS Bedrock in your data region:

• Amazon Bedrock has a zero data retention policy, no logging, provider isolation, and does not train on your data. See Bedrock data protection for more details.
• Data is never shared with third parties, such as the model developer.
• AWS Bedrock is fully HIPAA, SOC2 and ISO 27001 compliant. See Bedrock security and compliance and our security overview for more details.

On self-hosted deployments, AI features use the Langfuse AI model you configure. You are responsible for that provider's data handling. See self-hosted Assistant setup.

Self-hosted deployments can also record Assistant turns and Ask AI completions as traces by setting LANGFUSE_AI_FEATURES_PROJECT_ID.

Nothing is recorded while the variable is unset. When set, traces are written through Langfuse's internal ingestion path and add no outbound requests. Use a dedicated project because these traces contain the prompts, tool calls, and model output of every run. See Trace the Assistant in your own instance.

---

---