API reference

This page covers the key APIs in the Braintrust Go SDK. For setup, see the Quickstart. For the complete reference, see pkg.go.dev.

Tracing

Tracing records what your application does as spans you can inspect in Braintrust. The recommended way to capture AI calls is auto-instrumentation: use the trace/contrib packages to instrument supported provider libraries, either at build time with Orchestrion or with runtime middleware (see Go SDK integrations). Tracing is built on OpenTelemetry, so you trace your own code with the standard OpenTelemetry API. The APIs below create the client, trace your own code, and link to your traces.

`braintrust.New`

Creates a Braintrust client and configures the OpenTelemetry pipeline that exports spans to Braintrust. Call it once on startup, passing your TracerProvider and any options.

#skip-compile

import (
	"github.com/braintrustdata/braintrust-sdk-go"
	"go.opentelemetry.io/otel"
	"go.opentelemetry.io/otel/sdk/trace"
)

tp := trace.NewTracerProvider()
otel.SetTracerProvider(tp)

client, err := braintrust.New(tp, braintrust.WithProject("My project"))
if err != nil {
	log.Fatal(err)
}

Returns: (*braintrust.Client, error). braintrust.New reads BRAINTRUST_API_KEY from the environment. Configure the rest with functional options or environment variables (see Configuration). Because tracing is built on OpenTelemetry, you trace your own application code with the standard OpenTelemetry API, and traced AI calls nest under your spans.

#skip-compile

ctx, span := otel.Tracer("my-app").Start(ctx, "process-request")
defer span.End()

`Client.Permalink`

Builds a Braintrust UI URL for a span, so you can link straight to a trace from your own logs or app.

#skip-compile

url := client.Permalink(span)

Returns: string.

`config.SpanCustomizer`

Span customizers modify each span after it ends and before the SDK exports it to Braintrust. Register them with braintrust.WithSpanCustomizers. They run in registration order on every exported span, including spans you create with the OpenTelemetry API. See Span customizers for examples and behavior.

#skip-compile

import (
	sdktrace "go.opentelemetry.io/otel/sdk/trace"

	"github.com/braintrustdata/braintrust-sdk-go"
	"github.com/braintrustdata/braintrust-sdk-go/config"
)

client, err := braintrust.New(tp,
	braintrust.WithSpanCustomizers(config.SpanCustomizer{
		OnSpanExport: func(span sdktrace.ReadOnlySpan) (sdktrace.ReadOnlySpan, error) {
			return span, nil
		},
	}),
)

config.SpanCustomizer fields:

OnSpanExport (func(span sdktrace.ReadOnlySpan) (sdktrace.ReadOnlySpan, error)): receives a completed span and returns it or a replacement that embeds sdktrace.ReadOnlySpan.

Key rules:

Don’t change the trace ID, span ID, parent trace ID, or parent span ID. Attributes can change.
If a customizer returns an error, panics, returns nil, or changes an ID, the SDK skips the remaining customizers for that span and exports it with its content stripped. Other spans in the batch export normally. See When a customizer fails.
Customizers run before the SDK converts and uploads attachments, so removing inline base64 data keeps the file out of Braintrust.
The SDK can call customizers from background goroutines, so make them safe for concurrent use.

Evaluations

An evaluation runs your task over a set of cases, scores each output, and logs the results to an experiment, which is how you measure quality and catch regressions as you change prompts or models. The recommended pattern is to define an eval once with braintrust.NewEval, then call Run with any dataset. The same definition works for local runs, bt eval <dir> runs from the command line, and remote eval runs triggered from the playground.

`braintrust.NewEval`

Creates a runnable *eval.Eval by combining a client with an eval definition. Call Run on it to execute the evaluation.

#skip-compile

import (
	"context"

	"github.com/braintrustdata/braintrust-sdk-go"
	"github.com/braintrustdata/braintrust-sdk-go/eval"
)

e := braintrust.NewEval(client, &eval.Eval[string, string]{
	Name: "classify",
	Task: eval.T(func(ctx context.Context, input string) (string, error) {
		return classify(input), nil
	}),
	Scorers: []eval.Scorer[string, string]{exactMatch},
	ProjectName: "my-project",
})

_, err := e.Run(ctx, eval.RunOpts[string, string]{
	Dataset: eval.NewDataset([]eval.Case[string, string]{
		{Input: "apple", Expected: "fruit"},
	}),
})

Returns: *eval.Eval[I, R]. Run returns (*eval.Result, error). eval.Eval[I, R] fields:

Name (string, required): eval name, used as the default experiment name and as the registration key for remote evals.
Task (eval.TaskFunc[I, R], required): the function under evaluation. Wrap a plain func(ctx, input) (output, error) with eval.T, or use eval.TaskWithHooks if the task needs to read parameters or other run metadata.
Scorers ([]eval.Scorer[I, R], required): scoring functions applied to each case. To use classifiers, use braintrust.NewEvaluator instead.
ParameterSchema (eval.ParameterSchema): declares parameters the eval accepts. Each entry becomes a configurable control in the Braintrust playground (see Parameters).
Dataset (eval.Dataset[I, R]): default dataset for the eval. A playground run supplies its own dataset and overrides this. bt eval <dir> and a direct Run call use it when no dataset is supplied at call time.
ProjectName (string): Braintrust project for this eval. Falls back to the client’s configured project.

eval.RunOpts[I, R] fields (passed to Run):

Dataset (eval.Dataset[I, R]): test cases for this run. Defaults to Eval.Dataset; required if the eval defines none.
Experiment (string): experiment name. Defaults to Eval.Name.
ProjectName (string): overrides the project for this run.
ProjectID (string): identifies the project by ID instead of by name. Takes precedence over ProjectName when set.
Tags ([]string): tags to apply to the experiment.
Metadata (eval.Metadata): metadata to attach to the experiment.
Update (bool): append to an existing experiment with the same name. Defaults to false.
Parallelism (int): number of goroutines. Defaults to 1.
Quiet (bool): suppress result output. Defaults to false.
Parameters (eval.Parameters): parameter values for this run, merged over the defaults in Eval.ParameterSchema and passed to the task through hooks.Parameters. When unset, the task gets the declared defaults.
NoSendLogs (bool): run the eval locally without sending anything to Braintrust. Cases still run and are scored, but no experiment is created and no eval spans are uploaded, so the result has no experiment ID or permalink. Defaults to false. Requires v0.17.2 or later.

`braintrust.NewEvaluator`

Creates an evaluator for input type I and result type R, bound to a client. Call Run on it with the cases, task, and scorers to execute the evaluation and log an experiment.

#skip-compile

import (
	"context"

	"github.com/braintrustdata/braintrust-sdk-go"
	"github.com/braintrustdata/braintrust-sdk-go/eval"
)

evaluator := braintrust.NewEvaluator[string, string](client)

_, err := evaluator.Run(context.Background(), eval.Opts[string, string]{
	Experiment: "answers-v1",
	Dataset: eval.NewDataset([]eval.Case[string, string]{
		{Input: "How do I reset my password?", Expected: "Use the account recovery flow."},
		{Input: "How do I export my data?", Expected: "Open Settings and choose Export."},
	}),
	Task: eval.T(answerQuestion),
	Scorers: []eval.Scorer[string, string]{
		eval.NewScorer("exact_match", func(_ context.Context, r eval.TaskResult[string, string]) (eval.Scores, error) {
			v := 0.0
			if r.Output == r.Expected {
				v = 1.0
			}
			return eval.S(v), nil
		}),
	},
})

Returns: *eval.Evaluator[I, R]. Run returns (*eval.Result, error). eval.Opts[I, R] fields:

Experiment (string, required): experiment name.
Dataset (eval.Dataset[I, R], required): the cases to run. Build an in-memory one with eval.NewDataset (see Datasets).
Task (eval.TaskFunc[I, R], required): the function under test. Wrap a plain func(ctx, input) (output, error) with eval.T.
Scorers ([]eval.Scorer[I, R]): scorers to apply to each case. Provide Scorers, Classifiers, or both.
Classifiers ([]eval.Classifier[I, R]): classifiers to apply to each case. Provide Scorers, Classifiers, or both.
ProjectName (string): project to log to. Defaults to the client’s configured project.
ProjectID (string): project ID. When set, takes precedence over ProjectName and uses the project as-is rather than creating it.
Tags ([]string): tags to apply to the experiment.
Metadata (eval.Metadata): metadata to attach to the experiment.
Update (bool): append to an existing experiment with the same name. Defaults to false.
Parallelism (int): number of goroutines. Defaults to 1.
TrialCount (int): number of times to run each case. Defaults to 1.
Quiet (bool): suppress result output. Defaults to false.
NoSendLogs (bool): run the eval locally without sending anything to Braintrust. Cases still run and are scored, but no experiment is created and no eval spans are uploaded, so the result has no experiment ID or permalink. Defaults to false. Requires v0.17.2 or later.

`eval.NewScorer`

Creates a scorer from a function. A scorer measures how good the task’s output is, returning one or more named scores per case.

#skip-compile

scorer := eval.NewScorer("exact_match", func(_ context.Context, r eval.TaskResult[string, string]) (eval.Scores, error) {
	if r.Output == r.Expected {
		return eval.S(1.0), nil
	}
	return eval.S(0.0), nil
})

Returns: eval.Scorer[I, R]. The score function receives an eval.TaskResult[I, R] (with Input, Output, Expected, and Metadata) and returns eval.Scores. Use eval.S to build a single score.

`eval.NewClassifier`

Creates a classifier from a function. Use a classifier to categorize output instead of scoring it numerically.

#skip-compile

classifier := eval.NewClassifier("topic", func(_ context.Context, r eval.TaskResult[string, string]) (eval.Classifications, error) {
	return eval.Classifications{{ID: "billing", Label: "Billing"}}, nil
})

Returns: eval.Classifier[I, R].

`eval.TaskWithHooks`

Creates a task function that receives *eval.TaskHooks, which gives access to parameters, metadata, tags, and the current spans. Use it when your task needs to read parameter values that were configured in the playground or passed via RunOpts.Parameters.

#skip-compile

task := eval.TaskWithHooks(func(ctx context.Context, input string, hooks *eval.TaskHooks) (string, error) {
	model := hooks.Parameters.String("model")
	return classify(input, model), nil
})

Returns: eval.TaskFunc[I, R]. The hooks give access to Parameters, Metadata, Tags, TrialIndex, TaskSpan, and EvalSpan. Use eval.T instead when the task doesn’t need hooks.

Workflow evaluations

A workflow evaluation is an evaluation whose task or scorers submit requests to an asynchronous provider API, such as the OpenAI Batch API, and collect the results later. The run saves its progress to a store, so your application can resume it by polling or from a webhook. With a shared store such as Redis, a different process can resume it. See Workflow evaluations for a complete OpenAI example.

Go workflow evaluations are in public preview and can change before reaching general availability.

Requires Go SDK v0.18.0 or later. The Go SDK saves run state in its own format, so start and resume a run with the Go SDK.

`braintrust.NewWorkflowEval`

Attaches a client to an *eval.WorkflowEval definition, so its runs create an experiment and log results to Braintrust. Every process that starts or resumes a run that logs to Braintrust must use a definition attached this way. Returns: *eval.WorkflowEval[I, R]. Defining an evaluation does not start a run. To start one, call Start. eval.WorkflowEval[I, R] fields:

Name (string, required): evaluation name, used as the default experiment name and to identify the run’s records in the store.
Store (eval.WorkflowEvalStore, required): persistence backend shared by every process that starts or resumes the run. See Workflow stores.
Task (eval.WorkflowTaskDefinition[I, R], required): an ordinary eval.TaskFunc[I, R] such as one built with eval.T, an eval.WorkflowTask, or an eval.WorkflowBatchTask. Task outputs are saved to the store, so they must be JSON-serializable. An ordinary task that returns eval.TaskOutput with UserData fails for that case. Use metadata instead.
Scorers ([]eval.Scorer[I, R]): ordinary scorers, which run in your process after a case’s task output is collected.
WorkflowScorers ([]eval.WorkflowScorerDefinition[I, R]): asynchronous scorers. Each is an eval.WorkflowScorer or an eval.WorkflowBatchScorer.
Classifiers ([]eval.Classifier[I, R]): classifiers, which run in your process after a case’s task output is collected.
ParameterSchema (eval.ParameterSchema): declares parameters the evaluation accepts (see Parameters). Pass runtime values to Start.
ProjectName (string): Braintrust project for this evaluation. Falls back to the client’s configured project.
MaxConcurrency (int): how many of your callbacks the SDK runs at the same time during each Start, Poll, or ProcessSubmissionResult call. Defaults to 10. Must not be negative.

Scorer and classifier names must be nonempty and unique across Scorers, WorkflowScorers, and Classifiers, and can’t be task. Keep the definition, including scorer names and order, unchanged while runs are active.

`WorkflowEval` methods

Each method returns (*eval.WorkflowResult, error) with the run’s current state.

Start(ctx, opts eval.WorkflowStartOpts[I, R]): creates the experiment, saves the dataset to the store, and submits the initial task work. It returns without waiting for the provider. With an ordinary eval.TaskFunc task, Start runs the task and the ordinary scorers and classifiers for every case before it returns. If a callback fails after the run is saved, Start still returns the run ID. Resume that run with Poll instead of starting a second one.
Poll(ctx, runID string): checks each pending poll-mode submission once, collects the completed ones, and advances the run. It also retries submissions that failed to submit. Call it on a schedule until the run completes.
ProcessSubmissionResult(ctx, ref eval.WorkflowSubmissionResult): collects one completed submission and advances its run. Call it after your application verifies a provider webhook. Repeated notifications for the same submission are harmless.
Status(ctx, runID string): reads the run’s progress without submitting, polling, or collecting.

Serialize calls that advance the same run, including across processes. The store operations are atomic, but the SDK doesn’t provide a distributed lock. Different runs can advance independently. eval.WorkflowStartOpts[I, R] fields:

Dataset (eval.Dataset[I, R], required): test cases for this run. Inputs, expected values, and metadata must be JSON-serializable.
Experiment (string): experiment name. Defaults to WorkflowEval.Name.
ProjectID (string): identifies the project by ID instead of by name.
Tags ([]string): tags to apply to the experiment.
Metadata (eval.Metadata): metadata to attach to the experiment.
TrialCount (int): number of times to run each case. Defaults to 1. A case’s TrialCount overrides this value.
Parameters (eval.Parameters): parameter values for this run, merged over the defaults in ParameterSchema.
NoSendLogs (bool): run without creating an experiment or logging results to Braintrust, so no client is required. Run state is still saved to the store. Tracing that your own callbacks do is not disabled.

eval.WorkflowSubmissionResult fields:

ExternalID (string): the provider’s ID for the completed operation, as returned by GetExternalID. Sufficient on its own unless the same ID belongs to several runs. In that case, also set RunID.
RunID (string): the run that owns the submission.
SubmissionID (string): the SDK’s submission ID. Requires RunID. Use it instead of ExternalID.

eval.WorkflowResult fields:

RunID (string): run ID to pass to Poll, Status, or ProcessSubmissionResult.
Status (eval.WorkflowRunStatus): eval.WorkflowRunWaiting ("waiting") while cases or submissions are outstanding, and eval.WorkflowRunCompleted ("completed") once every case has finished, including failed cases.
Completed (int): number of finished cases, including failed cases.
Failed (int): number of cases with a failed task, scorer, or classifier. A completed run can have failures.
PendingPoll and PendingWebhook (int): number of submitted operations not yet collected, by completion mode.
Scores (map[string]float64): average of the successful values for each score name.
ExperimentID and Permalink (string): the experiment and its URL. Empty for NoSendLogs runs.

Each trial is logged as its own experiment row. The row’s metadata.workflow_eval records run_id, case_id, and trial_index, plus batch_submission_id when an eval.WorkflowBatchTask produced the output. case_id is the dataset case’s ID when set, otherwise an ID generated when the run starts. A failed task’s error appears in the row’s error field, and failed scorers and classifiers appear in metadata.scorer_errors and metadata.classifier_errors.

`eval.WorkflowTask` and `eval.WorkflowScorer`

eval.WorkflowTask and eval.WorkflowScorer each submit one asynchronous provider operation for each case and trial. Use eval.WorkflowTask for the task and eval.WorkflowScorer for a scorer. eval.WorkflowTask and eval.WorkflowScorer fields:

Name (string, required, WorkflowScorer only): scorer name.
Submit (required): sends the request to the provider and returns a json.RawMessage handle, such as the provider’s job ID, that later callbacks receive. The handle must be valid JSON and must not contain clients or credentials. A task’s Submit receives an eval.WorkflowTaskItem[I, R], and a scorer’s receives an eval.WorkflowScorerItem[I, R].
Completion (eval.WorkflowCompletion, required): how the SDK learns the operation finished. See eval.WorkflowCompletion.
Collect (required): reads the finished operation from its handle. A task’s Collect returns an eval.WorkflowTaskResult[R], and a scorer’s returns eval.Scores. If Collect returns an error, the case stays pending. The SDK calls Collect again on the next Poll, or for webhook completion, on the next ProcessSubmissionResult call for that submission.

Every callback also receives an eval.WorkflowSubmissionContext with RunID and SubmissionID. SubmissionID is stable across retries, so use it as the provider’s idempotency key. A crash after the provider accepts a request but before the SDK saves the handle can call Submit again. eval.WorkflowTaskItem[I, R] fields: ID (unique for each trial), CaseID (shared by a case’s trials), Input, Expected, Metadata, Tags, Parameters, and TrialIndex. eval.WorkflowTaskResult[R] fields: Output, and optional Metadata and Tags. Tags replaces the case’s tags. eval.WorkflowScorerItem[I, R] embeds eval.TaskResult[I, R] (with Input, Output, Expected, and Metadata) and adds ID, CaseID, Tags, and TrialIndex.

`eval.WorkflowBatchTask` and `eval.WorkflowBatchScorer`

eval.WorkflowBatchTask and eval.WorkflowBatchScorer group many cases into one provider operation, such as one OpenAI batch. They have the same Name (scorer only) and Completion fields as eval.WorkflowTask and eval.WorkflowScorer, plus:

Batching (eval.WorkflowBatching, required): batch size limits. See below.
Submit (required): receives a slice of eval.WorkflowBatchItem values, each with a CustomID and the Item. Send CustomID as the provider’s per-request ID, such as OpenAI’s custom_id. Treat it as opaque.
Collect (required): reads the finished operation and calls its yield function once for each result with an eval.WorkflowBatchResult (CustomID, Result, and a per-item Error). Return immediately when yield returns an error.

eval.WorkflowBatching fields:

MaxSize (int, required): maximum number of cases in a batch. Must be positive.
MaxWait (time.Duration): batch scorers only. How long a partial batch can wait before it’s submitted. The SDK checks it only when the run advances, with no background timer. Defaults to 0, which waits for a full batch or for the last remaining cases. A batch task submits immediately and rejects a nonzero MaxWait.

Collection is checkpointed for each result, so a failed Collect can run again, and results already saved are ignored. A duplicate or unknown CustomID is an error. Cases with no result when Collect succeeds are recorded as failed. Failed tasks skip their scorers and classifiers, and a failed scorer keeps the case’s other scores.

`eval.WorkflowCompletion`

Selects how the SDK learns that a submission finished. Set exactly one field:

Poll: checks the provider and returns an eval.WorkflowPollResult with a Status of eval.WorkflowPending, eval.WorkflowComplete, or eval.WorkflowFailed, and an optional Error. WorkflowEval.Poll calls it.
GetExternalID: returns the provider’s ID for the submission, which the provider’s webhook later delivers. Pass that ID to ProcessSubmissionResult as ExternalID.

A submission that reports eval.WorkflowFailed stays pending, and Poll returns its error. Poll checks the submission again on every later call, so a permanent provider failure makes every later Poll return an error, and the run never completes. Start a new run to recover. To finish a run despite a failed provider batch, report eval.WorkflowComplete instead and omit the missing results from Collect, so those cases are recorded as failed.

Workflow stores

A store persists the run’s dataset, outputs, submission handles, and progress. Use the same store and key prefix for every call that starts or resumes a run.

eval.NewWorkflowEvalMemoryStore() → *eval.WorkflowEvalMemoryStore: keeps state in memory for single-process runs. State is lost when the process exits.
eval.NewWorkflowEvalRedisStore(command eval.WorkflowRedisCommand, prefix string, ttl time.Duration) → (*eval.WorkflowEvalRedisStore, error): stores state in Redis. command runs a Redis command and returns its decoded result, so the SDK doesn’t depend on a Redis client. With go-redis, wrap client.Do(ctx, args...).Result() and return nil, nil for redis.Nil. The command must support EVAL. An empty prefix defaults to braintrust-eval:, and a zero ttl defaults to seven days. Writes refresh the TTL and reads don’t, so set it longer than the full run and your webhook retry window. An expired run can’t be resumed.

For another backend, implement the eval.WorkflowEvalStore interface: Read, Write, GetOrSet, AddToSet, and GetSetSize. Read returns nil, nil for a missing key, and GetOrSet and AddToSet must be atomic.

Parameters

Parameters let you declare configurable options on an eval. When the eval runs from the Braintrust playground via a remote eval, each declared parameter becomes a control in the UI. For a local run, the task receives the declared defaults. Declare parameters with eval.ParameterSchema on the eval.Eval definition:

#skip-compile

e := braintrust.NewEval(client, &eval.Eval[string, string]{
	Name: "classify",
	ParameterSchema: eval.ParameterSchema{
		"model": {
			Type:        eval.ParameterTypeModel,
			Default:     "gpt-5-mini",
			Description: "Model to use for classification",
		},
		"threshold": {
			Type:    "number",
			Default: 0.5,
		},
	},
	Task: eval.TaskWithHooks(func(ctx context.Context, input string, hooks *eval.TaskHooks) (string, error) {
		model := hooks.Parameters.String("model")
		threshold := hooks.Parameters.Float64("threshold")
		return classify(input, model, threshold), nil
	}),
	// ...
})

eval.ParameterSchema is map[string]eval.ParameterDef. Each eval.ParameterDef has:

Type (string): "model" renders a model picker. "prompt" renders a prompt editor. "string", "number", "integer", "boolean" render plain inputs. Use the eval.ParameterTypeModel and eval.ParameterTypePrompt constants for the two dedicated controls.
Default (any): used when no value is supplied for this run. This is also what a local run sees.
Description (string): shown alongside the control in the playground.

eval.Parameters (the resolved values, available as hooks.Parameters) has typed accessors that never panic on a type mismatch:

Parameters.String(name) → string
Parameters.Int(name) → int
Parameters.Float64(name) → float64
Parameters.Bool(name) → bool
Parameters.Prompt(name) → (*prompt.Prompt, bool): returns the value for name as a renderable prompt, and whether one was found. Use this for "prompt" parameters. Returns false when the parameter is absent or cannot be converted.
Parameters.Get(name) → (any, bool)
Parameters.Has(name) → bool

Remote evals

Remote evals let the Braintrust playground trigger your Go eval code on your own infrastructure. The evalrunner package turns a Go binary into a target that the bt CLI can drive.

Go remote evals are in public preview and can change before reaching general availability.

`evalrunner.New` and `evalrunner.RegisterEval`

Create a runner and register your evals. The runner reads the bt environment variables, dispatches the right eval, and streams results back.

#skip-compile

package main

import (
	"context"

	"github.com/braintrustdata/braintrust-sdk-go/eval"
	"github.com/braintrustdata/braintrust-sdk-go/evalrunner"
)

func main() {
	r := evalrunner.New()

	evalrunner.RegisterEval(r, &eval.Eval[string, string]{
		Name: "classify",
		Task: eval.TaskWithHooks(func(ctx context.Context, input string, hooks *eval.TaskHooks) (string, error) {
			model := hooks.Parameters.String("model")
			return classify(input, model), nil
		}),
		Scorers: []eval.Scorer[string, string]{exactMatch},
		ParameterSchema: eval.ParameterSchema{
			"model": {Type: eval.ParameterTypeModel, Default: "gpt-5-mini"},
		},
		ProjectName: "my-project",
	})

	evalrunner.Main(r)
}

Then run from the command line:

# Run all evals locally
bt eval ./cmd/evals

# Run all evals without logging in or sending results to Braintrust (v0.17.2 or later)
bt eval --no-send-logs ./cmd/evals

# Start the dev server so the playground can trigger runs
bt eval --dev --language go ./cmd/evals

evalrunner.New accepts evalrunner.Option values:

evalrunner.WithLogger(l logger.Logger): custom logger. It must write to stderr, because stdout is reserved for the bt protocol. The default logger writes to stderr.
evalrunner.WithTracerProvider(tp *sdktrace.TracerProvider): shared OpenTelemetry provider. Supply one when instrumented code (LLM clients, custom spans) should appear in the same trace as eval spans. When nil, a per-run provider is created and shut down on exit.

evalrunner.RegisterEval[I, R any](r *Runner, ev *eval.Eval[I, R]) registers an eval by its Name. Registering two evals under the same name replaces the first. evalrunner.Main(r *Runner) dispatches and exits. Use evalrunner.Run(ctx, r) instead if you need to handle the error yourself. r.Mode() returns the current evalrunner.Mode: ModeList, ModeEval, ModeBatch, ModeInspect, or ModeUnknown. ModeUnknown means bt requested a mode the runner doesn’t recognize, and evalrunner.Run returns an error. Use it to skip expensive setup when bt only wants the list of registered evals:

#skip-compile

if r.Mode() != evalrunner.ModeList {
	warmCaches()
}

Prompts

Prompts saved in Braintrust carry a template, a model, and parameters. The Go SDK loads them from the API, renders their variables, and returns a provider-agnostic result you hand to any LLM client. A prompt is fetched every time. Nothing is cached.

`client.LoadPrompt`

Loads a prompt from Braintrust by slug and returns it ready to render. The client’s configured project is used when no project is specified in opts.

#skip-compile

import "github.com/braintrustdata/braintrust-sdk-go/prompt"

p, err := bt.LoadPrompt(ctx, prompt.LoadOpts{Slug: "summarizer"})
if err != nil {
	return err
}
built, err := p.Build(map[string]any{"input": article})

Returns: (*prompt.Prompt, error). prompt.LoadOpts fields:

Slug (string): prompt slug. Required unless ID is set.
Project (string): project name. Filled in from the client’s project when empty.
ProjectID (string): project ID. Takes precedence over Project when set.
ID (string): load by Braintrust prompt ID instead of slug.
Version (string): pin to a specific prompt version. Defaults to the latest.
Environment (string): load the prompt deployed to an environment (for example, "production"). Ignored when Version is set.

For lower-level access, use prompt.LoadWithClient(ctx, apiClient, opts) with an *openapi.Client from openapi.NewClient, configured with a base URL and Authorization header (v0.19.0 or later). prompt.Load, which takes an *api.API, is deprecated but remains available for existing code.

`prompt.Definition`

Declares a prompt in Go code. Use it as the Default for a prompt eval parameter:

#skip-compile

import "github.com/braintrustdata/braintrust-sdk-go/prompt"

eval.ParameterSchema{
	"summary_prompt": {
		Type: eval.ParameterTypePrompt,
		Default: prompt.Definition{
			Model: "gpt-5-mini",
			Messages: []prompt.Message{
				prompt.System("You summarize articles in one sentence."),
				prompt.User("Summarize this:\n\n{{input}}"),
			},
			Params: map[string]any{"temperature": 0},
		},
	},
}

prompt.Definition fields:

Model (string, required): model the prompt targets.
Messages ([]prompt.Message): chat messages. Mutually exclusive with Prompt.
Prompt (string): completion template. Mutually exclusive with Messages.
Params (map[string]any): model parameters such as temperature or max_tokens.
Tools ([]prompt.Tool): tool definitions offered to the model. Chat prompts only.
TemplateFormat (string): "mustache" (default) or "none".

Helper constructors for prompt.Message: prompt.System(text), prompt.User(text), prompt.Assistant(text).

`prompt.Prompt.Build`

Renders the prompt’s template variables and returns a *prompt.Built ready to send to a model.

#skip-compile

built, err := p.Build(map[string]any{"input": article})
if err != nil {
	return err
}
// built is provider-agnostic. built.Map() returns an OpenAI-shaped map you can
// send to any client:
body, err := json.Marshal(built.Map())

Returns: (*prompt.Built, error). Build fails if the prompt has no body or no model, the template is malformed, or the format is not supported ("nunjucks" is not rendered locally). For the OpenAI Go client, traceopenai.ChatCompletionParams(built) (from github.com/braintrustdata/braintrust-sdk-go/trace/contrib/openai) converts a chat prompt directly into openai.ChatCompletionNewParams. prompt.Built fields:

Model (string): model to call.
Messages ([]prompt.Message): rendered chat messages, for a chat prompt.
Prompt (string): rendered completion text, for a completion prompt.
Params (map[string]any): model parameters.
Tools ([]prompt.Tool): rendered tool definitions.
Metadata (*prompt.Metadata): prompt identity for tracing. Nil when the prompt has no Braintrust identity.

built.Map() returns the built prompt as an OpenAI-shaped map[string]any (model, messages or prompt, tools, and every model parameter at the top level) for clients this SDK has no adapter for. Build options:

prompt.WithDefaults(defaults map[string]any): default model parameters, used for any parameter the prompt does not set.
prompt.WithTemplateFormat(format string): override the template format saved on the prompt.
prompt.WithStrict(): fail when the template references a variable that was not supplied. By default a missing variable renders as an empty string.

`built.AnnotateSpan`

Records the prompt’s identity and rendered variables on a span, linking the model call back to the prompt in Braintrust. Call it before ending the span that covers the model call.

#skip-compile

built, err := p.Build(map[string]any{"input": input})
if err != nil {
	return err
}
built.AnnotateSpan(hooks.TaskSpan)

Does nothing when Metadata is nil (the prompt has no Braintrust identity) or the span is not recording.

Datasets

A dataset is the set of cases an evaluation runs against. Define cases inline in memory, or manage datasets in Braintrust through the API client.

`eval.NewDataset`

Groups cases into an in-memory dataset you pass to Evaluator.Run, as an alternative to loading one from Braintrust.

#skip-compile

dataset := eval.NewDataset([]eval.Case[string, string]{
	{Input: "How do I reset my password?", Expected: "Use the account recovery flow."},
	{Input: "How do I export my data?", Expected: "Open Settings and choose Export."},
})

Returns: eval.Dataset[I, R]. Each eval.Case[I, R] has an Input and optional Expected, Tags, Metadata, and TrialCount.

Attachments

When your traces involve binary content like images or PDFs, log it as an attachment so it appears in Braintrust instead of as an opaque blob. When you trace AI calls, Braintrust automatically converts base64 attachments in provider messages into uploaded attachments, so you rarely need the APIs below for instrumented calls. Reach for them when you’re attaching binary content to a span yourself.

`attachment.From*`

Creates an attachment from bytes, a file, or a URL.

#skip-compile

import "github.com/braintrustdata/braintrust-sdk-go/trace/attachment"

att, err := attachment.FromFile("image/png", "chart.png")

Constructors:

attachment.FromBytes(contentType string, data []byte) → *attachment.Attachment: from raw bytes.
attachment.FromFile(contentType string, path string) → (*attachment.Attachment, error): reads a file.
attachment.FromURL(url string) → (*attachment.Attachment, error): fetches a URL and uses the response content type.
attachment.FromReader(contentType string, r io.Reader) → *attachment.Attachment: from an io.Reader.

API client

The openapi package gives direct access to the Braintrust REST API. Its client, methods, and models are generated from the Braintrust OpenAPI specification. It covers projects, experiments, datasets, prompts, functions, project automations, function invocation, and other REST resources, beyond what the higher-level APIs above cover. Requires v0.19.0 or later.

`openapi.NewClientWithResponses`

Creates a REST API client for a base URL. The client doesn’t read environment variables or add authentication, so set the base URL and the Authorization header yourself.

#skip-compile

import (
	"context"
	"fmt"
	"log"
	"net/http"
	"os"

	"github.com/braintrustdata/braintrust-sdk-go/openapi"
)

ctx := context.Background()
key := os.Getenv("BRAINTRUST_API_KEY")
client, err := openapi.NewClientWithResponses("https://api.braintrust.dev",
	openapi.WithRequestEditorFn(func(_ context.Context, req *http.Request) error {
		req.Header.Set("Authorization", "Bearer "+key)
		return nil
	}),
)
if err != nil {
	log.Fatal(err)
}

limit := 5
resp, err := client.GetProjectWithResponse(ctx, &openapi.GetProjectParams{Limit: &limit})
if err != nil {
	log.Fatal(err)
}
if resp.StatusCode() != http.StatusOK || resp.JSON200 == nil {
	log.Fatalf("list projects: HTTP %d: %s", resp.StatusCode(), resp.Body)
}
for _, project := range resp.JSON200.Objects {
	fmt.Println(project.Id, project.Name)
}

Returns: (*openapi.ClientWithResponses, error). Use openapi.WithHTTPClient to supply a custom *http.Client, for example with a timeout or transport. Key rules:

An HTTP error response isn’t a Go error. Check StatusCode() and the decoded field for the expected status, such as JSON200, as well as err. Body holds the raw response bytes, including error details.
Methods ending in WithResponse read and close the response body. If you use the raw methods from openapi.NewClient, close each response body yourself.
List and fetch operations return one page. Pass the pagination parameters or fetch cursor to request the next page. The client doesn’t paginate automatically.
Optional fields use pointers or nullable.Nullable[T]. For a nullable field, the zero value omits the field, nullable.NewNullableWithValue(value) sends a value, and nullable.NewNullNullable[T]() sends JSON null. The difference matters for PATCH requests.

For a complete example, including dataset and automation create, update, and delete flows, see the API client guide in the SDK repository.

`api.NewClient` (deprecated)

The api package and Client.API() are deprecated as of v0.19.0. They remain available and keep their existing behavior, but they won’t cover new REST resources. Use the openapi package for new code.

Creates a REST API client from an API key.

#skip-compile

import "github.com/braintrustdata/braintrust-sdk-go/api"

client := api.NewClient(os.Getenv("BRAINTRUST_API_KEY"))

Returns: *api.API. Namespaces:

client.Projects(): project management.
client.Experiments(): experiment management. Methods include Create, Register, List, Get, Delete, and InsertEvents(ctx, experimentID, events), which inserts experiments.Event rows (v0.18.0 or later). To update fields of an existing row instead of replacing it, set the event’s ID to the row’s ID and IsMerge to true.
client.Datasets(): dataset management. Methods include Create, Insert, InsertEvents, Delete, Fetch, and Query.
client.Functions(): function management, including Invoke(ctx, functionID, input) to call a deployed function.
client.Prompts(): prompt management. Methods include Query(ctx, params) to search for prompts and Get(ctx, id, params) to fetch one by ID.

Configuration

Configure the client with functional options passed to braintrust.New, or with environment variables.

#skip-compile

client, err := braintrust.New(tp,
	braintrust.WithProject("My project"),
	braintrust.WithBlockingLogin(true),
)

Client options:

braintrust.WithAPIKey(apiKey string): API key. Defaults to BRAINTRUST_API_KEY.
braintrust.WithAPIURL(apiURL string): Braintrust API URL.
braintrust.WithAppURL(appURL string): Braintrust app URL, used for permalinks.
braintrust.WithOrgName(orgName string): organization name, useful when credentials can access multiple orgs.
braintrust.WithProject(projectName string): project that receives exported spans.
braintrust.WithProjectID(projectID string): project ID. Takes precedence over the project name.
braintrust.WithBlockingLogin(enabled bool): log in synchronously during New instead of in the background.
braintrust.WithExporter(exporter trace.SpanExporter): supply a custom span exporter. Intended for testing.
braintrust.WithEnableTraceConsoleLog(enabled bool): print spans to the console.
braintrust.WithFilterAISpans(enabled bool): export only AI-related spans.
braintrust.WithSpanCustomizers(customizers ...config.SpanCustomizer): register span customizers that modify spans before export. Repeated calls append to the list.
braintrust.WithEnvironment(environmentType string, name ...string): set span-origin environment provenance. Overrides auto-detection from CI and server environment variables.

Environment variables

BRAINTRUST_API_KEY (required): Braintrust API key.
BRAINTRUST_API_URL: Braintrust API URL. Defaults to https://api.braintrust.dev.
BRAINTRUST_APP_URL: Braintrust app URL, used for permalinks. Defaults to https://www.braintrust.dev.
BRAINTRUST_DEFAULT_PROJECT: project that traced spans route to. Defaults to default-go-project.
BRAINTRUST_DEFAULT_PROJECT_ID: project UUID. Takes precedence over the project name.
BRAINTRUST_ORG_NAME: organization name, useful when credentials can access multiple orgs.
BRAINTRUST_OTEL_FILTER_AI_SPANS: set to true to export only AI-related spans.
BRAINTRUST_ENABLE_TRACE_CONSOLE_LOG: set to true to print spans to the console.
BRAINTRUST_OTEL_ENABLE_BUILTIN_ADK_TRACES: set to true to export spans from Google ADK’s built-in telemetry. Defaults to false.
BRAINTRUST_AUTO_CONVERT_AI_ATTACHMENTS: scan spans for base64 attachments and upload them to object storage. Defaults to true; set to false to disable.
BRAINTRUST_DEBUG: set to true to enable SDK debug logging. Defaults to false.
BRAINTRUST_BLOCKING_LOGIN: set to true to log in synchronously at startup.
BRAINTRUST_ENVIRONMENT_TYPE: environment type for span-origin provenance (for example, ci or server). Detected automatically from common CI and server runtimes when unset.
BRAINTRUST_ENVIRONMENT_NAME: environment name for span-origin provenance (for example, github_actions or aws_lambda). Detected automatically when unset.

Use the SDK

Features

Reference

Tracing

`braintrust.New`

`Client.Permalink`

`config.SpanCustomizer`

Evaluations

`braintrust.NewEval`

`braintrust.NewEvaluator`

`eval.NewScorer`

`eval.NewClassifier`

`eval.TaskWithHooks`

Workflow evaluations

`braintrust.NewWorkflowEval`

`WorkflowEval` methods

`eval.WorkflowTask` and `eval.WorkflowScorer`

`eval.WorkflowBatchTask` and `eval.WorkflowBatchScorer`

`eval.WorkflowCompletion`

Workflow stores

Parameters

Remote evals

`evalrunner.New` and `evalrunner.RegisterEval`

Prompts

`client.LoadPrompt`

`prompt.Definition`

`prompt.Prompt.Build`

`built.AnnotateSpan`

Datasets

`eval.NewDataset`

Attachments

`attachment.From*`

API client

`openapi.NewClientWithResponses`

`api.NewClient` (deprecated)

Configuration

Environment variables

​Tracing

​braintrust.New

​Client.Permalink

​config.SpanCustomizer

​Evaluations

​braintrust.NewEval

​braintrust.NewEvaluator

​eval.NewScorer

​eval.NewClassifier

​eval.TaskWithHooks

​Workflow evaluations

​braintrust.NewWorkflowEval

​WorkflowEval methods

​eval.WorkflowTask and eval.WorkflowScorer

​eval.WorkflowBatchTask and eval.WorkflowBatchScorer

​eval.WorkflowCompletion

​Workflow stores

​Parameters

​Remote evals

​evalrunner.New and evalrunner.RegisterEval

​Prompts

​client.LoadPrompt

​prompt.Definition

​prompt.Prompt.Build

​built.AnnotateSpan

​Datasets

​eval.NewDataset

​Attachments

​attachment.From*

​API client

​openapi.NewClientWithResponses

​api.NewClient (deprecated)

​Configuration

​Environment variables

Tracing

`braintrust.New`

`Client.Permalink`

`config.SpanCustomizer`

Evaluations

`braintrust.NewEval`

`braintrust.NewEvaluator`

`eval.NewScorer`

`eval.NewClassifier`

`eval.TaskWithHooks`

Workflow evaluations

`braintrust.NewWorkflowEval`

`WorkflowEval` methods

`eval.WorkflowTask` and `eval.WorkflowScorer`

`eval.WorkflowBatchTask` and `eval.WorkflowBatchScorer`

`eval.WorkflowCompletion`

Workflow stores

Parameters

Remote evals

`evalrunner.New` and `evalrunner.RegisterEval`

Prompts

`client.LoadPrompt`

`prompt.Definition`

`prompt.Prompt.Build`

`built.AnnotateSpan`

Datasets

`eval.NewDataset`

Attachments

`attachment.From*`

API client

`openapi.NewClientWithResponses`

`api.NewClient` (deprecated)

Configuration

Environment variables