Skip to main content
Google Cloud Discovery Engine is a managed service for building AI-powered search and answer generation applications. Braintrust traces Discovery Engine calls, including generative answers, conversational search, grounding checks, and document ranking.

​
Setup

Install the Braintrust and google-cloud-discoveryengine packages, set your API key, and authenticate with Google Cloud. Requires Braintrust v0.41.0+ and google-cloud-discoveryengine v0.20.3 or later.
1

Install packages

2

Set environment variables

.env
3

Authenticate with Google Cloud

The Discovery Engine clients load Application Default Credentials. For local development, authenticate with the gcloud CLI:
In production, configure Application Default Credentials for your runtime, such as an attached service account. Your credentials need permission to call Discovery Engine in the target Google Cloud project.

​
Auto-instrumentation

To trace Discovery Engine calls without modifying your application code, call braintrust.auto_instrument() before creating your Discovery Engine clients.

​
Manual instrumentation

To trace a specific Discovery Engine client instance manually, use wrap_google_discoveryengine(). Use this when you want to instrument a particular client rather than patching all Discovery Engine clients globally.

​
What Braintrust traces

Braintrust instruments synchronous and asynchronous google-cloud-discoveryengine v1 clients and creates spans per call:
  • Answer query spans (google_discoveryengine.answer_query), with the query and optional preamble as input, answer text, citations, references, and grounding details as output, and time-to-first-token for streaming.
  • Streaming answer query spans (google_discoveryengine.stream_answer_query), with the same structure as non-streaming answer queries, with incremental output aggregated across chunks.
  • Conversational search spans (google_discoveryengine.converse_conversation), with the query as input and the summary text with citations and references as output.
  • Grounding check spans (google_discoveryengine.check_grounding), with the answer candidate and grounding facts as input and support score, cited chunks, and claims as output.
  • Ranking spans (google_discoveryengine.rank), with query and document records as input, up to 100 ranked results with scores as output, and ranking configuration and model as metadata when specified.
  • Provider metadata and the model and serving configuration supplied with generation requests.
  • Errors captured on every call.
Asynchronous mode answer_query calls (where asynchronous_mode=True in the request) are not traced and pass through unmodified. This is separate from using an asynchronous Python client, which is supported.

​
Resources