Tracing

This guide walks through how to trace your code in Braintrust. Tracing is an invaluable tool for exploring the sub-components of your program which produce each top-level input and output. We currently support tracing in logging and evaluations.

Before proceeding, make sure to read the quickstart guide and setup an API key.

Logging Screenshot

Traces: the building block of logs

The core building blocks of logging are spans and traces. A span represents a unit of work, with a start and end time, and optional fields like input, output, metadata, scores, and metrics (the same fields you can log in an Experiment). Each span contains one or more children, which are usually run within their parent span (e.g. a nested function call). Common examples of spans include LLM calls, vector searches, the steps of an agent chain, and model evaluations.

Together, spans form a trace, which represents a single independent request. Each trace is visible as a row in the final table. Well-designed traces make it easy to understand the flow of your application, and to debug issues when they arise. The rest of this guide walks through how to log rich, helpful traces. The tracing API works the same way whether you are logging online (production logging) or offline (evaluations), so the examples below apply to either use-case.

Annotating your code

To log a trace, you simply wrap the code you want to trace. Braintrust will automatically capture and log information behind the scenes.

import { initLogger, traced, wrapTraced } from "braintrust";
 
const logger = initLogger({
  projectName: "My Project",
  apiKey: process.env.BRAINTRUST_API_KEY,
});
 
const preparePrompt = wrapTraced(function preparePrompt(body: string) {
  return [
    { role: "system", content: "You are a helpful assistant." },
    { role: "user", content: body },
  ];
});
 
const someLLMFunction = wrapTraced(async function someLLMFunction(
  prompt: Message[]
) {
  const result = await client.chat.completions.create({
    model: "gpt-3.5-turbo",
    messages: prompt,
  });
  return result.choices[0].message.content;
});
 
 
export async function POST(req: Request) {
  return traced(async (span) => {
    const result = await someLLMFunction(preparePrompt(req.body));
    span.log({ input: req.body, output: result });
    return result;
  });
}

Wrapping OpenAI

Braintrust includes a wrapper for the OpenAI API that automatically logs your requests. To use it, simply call wrapOpenAI/wrap_openai on your OpenAI instance. We intentionally do not monkey patch the libraries directly, so that you can use the wrapper in a granular way.

import { OpenAI } from "openai";
import { initLogger, traced, wrapOpenAI } from "braintrust";
 
const client = wrapOpenAI(new OpenAI());
const logger = initLogger({
  projectName: "My Project",
  apiKey: process.env.BRAINTRUST_API_KEY,
});
 
const someLLMFunction = wrapTraced(async function someLLMFunction(body: string) {
  const result = await client.chat.completions.create({
    model: "gpt-3.5-turbo",
    messages: [{ role: "user", content: req.body }],
  });
  return result.choices[0]?.message.content;
});
 
export async function POST(req: Request) {
  return traced(async (span) => {
    const result = await someLLMFunction(req.body);
    span.log({
      input: req.body,
      output: result,
      metadata: { user_id: req.user.id },
    });
    return result;
  });
}

Logging Result

Wrapping a custom LLM client

It is of course possible to wrap your own LLM client using braintrust. Similar to the previous example:

import { initLogger, traced } from "braintrust";
 
const logger = initLogger({
  projectName: "My Project",
  apiKey: process.env.BRAINTRUST_API_KEY,
});
 
async function invokeCustomLLM(llmInput) {
  return traced(async (span) => {
    const result = await callMyLLM(input);
    const content = result.completion;
    span.log({
      input: llmInput.body,
      output: content,
      metrics: { tokens: result.tokens },
      metadata: llmInput.params,
    });
    return content;
  });
}
 
export async function POST(req: Request) {
  return traced(async (span) => {
    const result = await invokeCustomLLM({
      body: req.body,
      params: { temperature: 0.1 },
    });
    span.log({ input: req.body, output: result });
    return result;
  });
}

Deeply nested code

Often, you want to trace functions that are deep in the call stack, without having to propagate the span object throughout. Braintrust uses async-friendly context variables to make this workflow easy:

The traced function/decorator will create a span underneath the currently-active span.
The currentSpan() / current_span() method returns the currently active span, in case you need to do additional logging.

import { currentSpan, initLogger, traced, wrapTraced } from "braintrust";
 
const logger = initLogger();
 
export const runLLM = wrapTraced(async function runLLM(input) {
  const result = await someLLMFunction(input);
  const output = result.content;
  currentSpan().log({
    metrics: {
      tokens: result.tokens,
    },
  });
  return output;
});
 
export const someLogic = wrapTraced(async function someLogic(input: string) {
  return await runLLM("You are a magical wizard. Answer the following question: " + input);
});
 
export async function POST(req) {
  return await traced(async () {
    const result = await someLogic(req.body);
    currentSpan().log({input: req.body, output: result, metadata: {user_id: req.user.id}});
    return result;
  });
}

Currently, in addition to text and structured data, Braintrust supports logging images. To log an image, simply provide an image URL or base64 encoded image as a string. The tree viewer will automatically render the image.

Image logging

The tree viewer will look at the URL or string to determine if it is an image. If you want to force the viewer to treat it as an image, then nest it in an object like

{
  "image_url": {
    "url": "https://example.com/image.jpg"
  }
}

and the viewer will render it as an image.

Tracing integrations

Vercel AI SDK

The Vercel AI SDK is an elegant tool for building AI-powered applications. You can wrap the SDK in Braintrust to automatically log your requests.

import { initLogger, traced, wrapAISDKModel } from "braintrust";
 
const logger = initLogger({
  projectName: "My Project",
  apiKey: process.env.BRAINTRUST_API_KEY,
});
 
const openai = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY,
});
 
const model = wrapAISDKModel(openai.chat("gpt-3.5-turbo"));
 
// This will automatically log the request, response, and metrics to Braintrust
const response = await model.doGenerate({
  inputFormat: "messages",
  mode: {
    type: "regular",
  },
  prompt: [
    {
      role: "user",
      content: [{ type: "text", text: "What is the capital of France?" }],
    },
  ],
});

Instructor

To use Instructor to generate structured outputs, you need to wrap the OpenAI client with both Instructor and Braintrust. It's important that you call Braintrust's wrap_openai first, because it uses low-level usage info and headers returned by the OpenAI call to log metrics to Braintrust.

import instructor
from braintrust import init_logger, load_prompt, wrap_openai
 
logger = init_logger(project="Your project name")
 
def run_prompt(text: str):
    # Replace with your project name and slug
    prompt = load_prompt("Your project name", "Your prompt name")
 
    # wrap_openai will make sure the client tracks usage of the prompt.
    client = instructor.patch(wrap_openai(OpenAI()))
 
    # Render with parameters
    return client.chat.completions.create(**prompt.build(input=text), response_model=MyResponseModel)

Langchain

To trace Langchain code in Braintrust, you can use the BraintrustTracer callback handler. The callback handler is currently only supported in Python, but if you need support for other languages, please let us know.

To use it, simply initialize a BraintrustTracer and pass it as a callback handler to langchain objects you create.

from braintrust import Eval
from braintrust.wrappers.langchain import BraintrustTracer
from langchain.chains import LLMMathChain
from langchain.chat_models import ChatOpenAI
 
from autoevals import Levenshtein
 
tracer = BraintrustTracer()
 
llm = ChatOpenAI(model="gpt-3.5-turbo", callbacks=[tracer])
llm_math = LLMMathChain.from_llm(llm, callbacks=[tracer])
 
Eval(
    "Calculator",
    data=[{"input": "1+1", "expected": "2"}],
    task=lambda input: llm_math.invoke(input),
    scores=[Levenshtein],
)

Distributed tracing

Sometimes it's useful to be able to start a trace in one process and continue it in a different one. For this purpose, Braintrust provides an export function which returns an opaque string identifier. This identifier can be passed to start_span to resume the trace elsewhere. Consider the following example of tracing across separate client and server processes.

Client code

import {
  initLogger,
  wrapTraced,
} from "braintrust";
 
const logger = initLogger({ projectName: "my-project" });
 
const bedTimeStory = wrapTraced(async function bedtimeStory(input: { summary: string; length: number }) {
  const output = await (new TextGeneratorClient()).chatCompletion({
    model: "gpt-3.5-turbo",
    messages: [
      {
        role: "system",
        content: "Come up with a bedtime story with the following summary and approximate length (in words)",
      },
      {
        role: "user",
        content: `summary: ${input.summary}\nlength: ${input.length}`,
      },
    ],
    extraHeaders: {
      request_id: span.export(),
    },
  });
});

Server code

import { traced } from "braintrust";
 
router.post("/chat/completion", async (request) => {
  return await traced(
    async (span) => {
      const output = await invokeLLM(request.body);
      span.log({
        input: request.body,
        output: output.completion,
        metrics: { tokens: output.tokens },
      });
      return output.completion;
    },
    {
      name: "text_generator_server",
      type: "llm",
      parent: request.headers.request_id,
    },
  );
});

Manually managing spans

In more complicated environments, it may not always be possible to wrap the entire duration of a span within a single block of code. In such cases, you can always pass spans around manually.

Consider this hypothetical server handler, which logs to a span incrementally over several distinct callbacks:

import {
    Span,
    init,
    startSpan,
    traced,
} from "braintrust";
 
function computeOutput(systemPrompt: string, userInput: string, parentSpan: Span) {
  return parentSpan.traced(async (span) => {
    const input = {messages: [
      { role: "system", text: systemPrompt },
      { role: "user", text: userInput },
    ]};
    const output = await computeOutputImpl(input);
    span.log({input, output});
    return output;
  });
}
 
class MyHandler {
  private liveSpans: Record<string, { span: Span; input: string }>;
 
  constructor() {
    this.liveSpans = {};
  }
 
  async onRequestStart(requestId, input, expected) {
    const span = startSpan({name: requestId, event: {input, expected}});
    this.liveSpans[requestId] = { span, input };
  }
 
  async onGetOutput(requestId, systemPrompt) {
    const { span, input } = this.liveSpans[requestId];
    const output = await computeOutput(systemPrompt, input, span);
    span.log({output});
  }
 
  async onRequestEnd(requestId, metadata) {
    { span } = this.liveSpans[requestId];
    delete this.liveSpans[requestId];
    span.log({metadata});
    span.end();
  }
}
 
experiment = init("My long-running experiment");
server.runForever(new MyHandler());

Importing and exporting spans

Spans are processed in Braintrust as a simple format, consisting of input, output, expected, metadata, scores, and metrics fields (all optional), as well as a few system-defined fields which you usually do not need to mess with, but are described below for completeness. This simple format makes it easy to import spans captured in other systems (e.g. languages other than Typescript/Python), or to export spans from Braintrust to consume in other systems.

Underlying format

The underlying span format contains a number of fields which are not exposed directly through the SDK, but are useful to understand when importing/exporting spans.

id is a unique identifier for the span, within the container (e.g. an experiment, or logs for a project). You can technically set this field yourself (to overwrite a span), but it is recommended to let Braintrust generate it automatically.
input, output, expected, scores, metadata, and metrics are optional fields which describe the span and are exposed in the Braintrust UI. When you use the Typescript or Python SDK, these fields are validated for you (e.g. scores must be a mapping from strings to numbers between 0 and 1).
span_attributes contains attributes about the span. Currently the recognized attributes are name, which is used to display the span name in the UI, and type, which displays a helpful icon. type should be one of "llm", "score", "function", "eval", "task", or "tool".
Depending on the container, e.g. an experiment, or project logs, or a dataset, fields like project_id, experiment_id, dataset_id, and log_id are set automatically, by the SDK, so the span can be later retrieved by the UI and API. You should not set these fields yourself.
span_id, root_span_id, and span_parents are used to construct the span tree and are automatically set by Braintrust. You should not set these fields yourself, but rather let the SDK create and manage them (even if importing from another system).

When importing spans, the only fields you should need to think about are input, output, expected, scores, metadata, and metrics. You can use the SDK to populate the remaining fields, which the next section covers with an example.

Here is an example of a span in the underlying format:

{
  "id": "385052b6-50a2-43b4-b52d-9afaa34f0bff",
  "input": {
    "question": "What is the origin of the customer support issue??"
  },
  "output": {
    "answer": "The customer support issue originated from a bug in the code.",
    "sources": ["http://www.example.com/faq/1234"]
  },
  "expected": {
    "answer": "Bug in the code that involved dividing by zero.",
    "sources": ["http://www.example.com/faq/1234"]
  },
  "scores": {
    "Factuality": 0.6
  },
  "metadata": {
    "pos": 1
  },
  "metrics": {
    "end": 1704872988.726753,
    "start": 1704872988.725727
    // Can also include `tokens`, etc. here
  },
  "project_id": "d709efc0-ac9f-410d-8387-345e1e5074dc",
  "experiment_id": "51047341-2cea-4a8a-a0ad-3000f4a94a96",
  "created": "2024-01-10T07:49:48.725731+00:00",
  "span_id": "70b04fd2-0177-47a9-a70b-e32ca43db131",
  "root_span_id": "68b4ef73-f898-4756-b806-3bdd2d1cf3a1",
  "span_parents": ["68b4ef73-f898-4756-b806-3bdd2d1cf3a1"],
  "span_attributes": {
    "name": "doc_included"
  }
}

Example import/export

The following example walks through how to generate spans in one program and then import them to Braintrust in a script. You can use this pattern to support tracing or running experiments in environments that use programming languages other than Typescript/Python (e.g. Kotlin, Java, Go, Ruby, Rust, C++), or codebases that cannot integrate the Braintrust SDK directly.

Generating spans

The following example runs a simple LLM app and collects logging information at each stage of the process, without using the Braintrust SDK. This could be implemented in any programming language, and you certainly do not need to collect or process information this way. All that matters is that your program generates a useful format that you can later parse and use to import the spans using the SDK.

import json
import time
 
import openai
 
client = openai.OpenAI()
 
 
def run_llm(input, **params):
    start = time.time()
    messages = [{"role": "user", "content": input}]
    result = client.chat.completions.create(
        model="gpt-3.5-turbo", messages=[{"role": "user", "content": input}], **params
    )
    end = time.time()
    return {
        "input": messages,
        "output": result.choices[0].message.dict(),
        "metadata": {"model": "gpt-3.5-turbo", "params": params},
        "metrics": {
            "start": start,
            "end": end,
            "tokens": result.usage.total_tokens,
            "prompt_tokens": result.usage.prompt_tokens,
            "completion_tokens": result.usage.completion_tokens,
        },
        "name": "OpenAI Chat Completion",
    }
 
 
PROMPT_TEMPLATE = "Answer the following question: %s"
 
 
def run_input(question, expected):
    result = run_llm(PROMPT_TEMPLATE % question, max_tokens=32)
    return {
        "input": question,
        "output": result["output"]["content"],
        # Expected is propagated here to make it easy to use it in the import
        # script, but it's not strictly needed to be here.
        "expected": expected,
        "metadata": {
            "template": PROMPT_TEMPLATE,
        },
        "children": [result],
        "name": "run_input",
    }
 
 
if __name__ == "__main__":
    for question, expected in [
        [
            "What is 1+1?",
            "2.",
        ],
        [
            "Which is larger, the sun or the moon?",
            "The sun.",
        ],
    ]:
        print(json.dumps(run_input(question, expected)))

Running this script produces output like:

{"input": "What is 1+1?", "output": "The sum of 1+1 is 2.", "expected": "2.", "metadata": {"template": "Answer the following question: %s"}, "children": [{"input": [{"role": "user", "content": "Answer the following question: What is 1+1?"}], "output": {"content": "The sum of 1+1 is 2.", "role": "assistant", "function_call": null, "tool_calls": null}, "metadata": {"model": "gpt-3.5-turbo", "params": {"max_tokens": 32}}, "metrics": {"start": 1704916642.978631, "end": 1704916643.450115, "tokens": 30, "prompt_tokens": 19, "completion_tokens": 11}, "name": "OpenAI Chat Completion"}], "name": "run_input"}
{"input": "Which is larger, the sun or the moon?", "output": "The sun is larger than the moon.", "expected": "The sun.", "metadata": {"template": "Answer the following question: %s"}, "children": [{"input": [{"role": "user", "content": "Answer the following question: Which is larger, the sun or the moon?"}], "output": {"content": "The sun is larger than the moon.", "role": "assistant", "function_call": null, "tool_calls": null}, "metadata": {"model": "gpt-3.5-turbo", "params": {"max_tokens": 32}}, "metrics": {"start": 1704916643.450675, "end": 1704916643.839096, "tokens": 30, "prompt_tokens": 22, "completion_tokens": 8}, "name": "OpenAI Chat Completion"}], "name": "run_input"}

Importing spans

The following program uses the Braintrust SDK in Python to import the spans generated by the previous script. Again, you can modify this program to fit the needs of your environment, e.g. to import spans from a different source or format.

import json
import sys
 
import braintrust
 
from autoevals import Factuality
 
 
def upload_tree(span, node, **kwargs):
    span.log(
        input=node.get("input"),
        output=node.get("output"),
        expected=node.get("expected"),
        metadata=node.get("metadata"),
        metrics=node.get("metrics"),
        **kwargs
    )
    for c in node.get("children", []):
        with span.start_span(name=c.get("name")) as span:
            upload_tree(span, c)
 
 
if __name__ == "__main__":
    # This could be another container, like a log stream initialized
    # via braintrust.init_logger()
    experiment = braintrust.init("My Support App")
 
    factuality = Factuality()
    for line in sys.stdin:
        tree = json.loads(line)
        with experiment.start_span(name="task") as span:
            upload_tree(span, tree)
            with span.start_span(name="Factuality"):
                score = factuality(input=tree["input"], output=tree["output"], expected=tree["expected"])
            span.log(
                scores={
                    "factuality": score.score,
                },
                # This will merge the metadata from the factuality score with the
                # metadata from the tree.
                metadata={"factuality": score.metadata},
            )
 
    print(experiment.summarize())

BRAINTRUST_SYNC_FLUSH: By default, the SDKs will log to the backend API in the background, asynchronously. Logging is automatically batched and retried upon encountering network errors. If you wish to have fine-grained control over when logs are flushed to the backend, you may set BRAINTRUST_SYNC_FLUSH=1. When true, flushing will only occur when you run Experiment.flush (or any of the other object flush methods). If the flush fails, the SDK will raise an exception which you can handle.
BRAINTRUST_MAX_REQUEST_SIZE: The SDK logger batches requests to save on network roundtrips. The batch size is tuned for the AWS lambda gateway, but you may adjust this if your backend has a different max payload requirement.
BRAINTRUST_DEFAULT_BATCH_SIZE: The maximum number of individual log messages that are sent to the network in one payload.
BRAINTRUST_NUM_RETRIES: The number of times the logger will attempt to retry network requests before failing.
BRAINTRUST_QUEUE_SIZE (Python only): The maximum number of elements in the logging queue. This value limits the memory usage of the logger. Logging additional elements beyond this size will block the calling thread. You may set the queue size to unlimited (and thus non-blocking) by passing BRAINTRUST_QUEUE_SIZE=0.
BRAINTRUST_QUEUE_DROP_WHEN_FULL (Python only): Useful in conjunction with BRAINTRUST_QUEUE_SIZE. Change the behavior of the queue from blocking when it reaches its max size to dropping excess elements. This can be useful for guaranteeing non-blocking execution, at the cost of possibly dropping data.
BRAINTRUST_QUEUE_DROP_EXCEEDING_MAXSIZE (Javascript only): Essentially a combination of BRAINTRUST_QUEUE_SIZE and BRAINTRUST_QUEUE_DROP_WHEN_FULL, which changes the behavior of the queue from storing an unlimited number of elements to capping out at the specified value. Additional elements are discarded.
BRAINTRUST_FAILED_PUBLISH_PAYLOADS_DIR: Sometimes errors occur when writing records to the backend. To aid in debugging errors, you may set this environment variable to a directory of choice, and Braintrust will save any payloads it failed to publish to this directory.
BRAINTRUST_ALL_PUBLISH_PAYLOADS_DIR: Analogous to BRAINTRUST_FAILED_PUBLISH_PAYLOADS_DIR, except that Braintrust will save all payloads to this directory.