@libretto/anthropic

A drop-in replacement of the official Anthropic client for sending events to Libretto.

Installation

npm install @libretto/anthropic

Get Started

To send events to Libretto, you'll need to create a project. From the project you'll need two things:

1. API key: (apiKey) This is generated for the project and is used to identify the project and environment (dev, staging, prod) that the event is coming from.
2. Template Name: (promptTemplateName) This uniquely identifies a particular prompt that you are using and allows projects to have multiple prompts. This can be in any format but we recommend using a dash-separated format, e.g. my-prompt-name.

Note: Prompt template names can be auto-generated if the allowUnnamedPrompts configuration option is set (see below). However, if you rely on auto-generated names, new revisions of the same prompt will show up as different prompt templates in Libretto.

Usage

You can use the Anthropic client provided by @libretto/anthropic anywhere that you're currently using the official client.

When instantiating the client, you can/should provide any of the standard Anthropic parameters in the constructor. Libretto-specific configuration can be provided via an additional libretto argument (see below).

To allow our tools to separate the "prompt" from the "prompt parameters", use the included objectTemplate helper and pass the parameters separately as follows:

import { Anthropic, objectTemplate } from "@libretto/anthropic";

async function main() {
  const anthropic = new Anthropic({
    apiKey: "<Anthropic API Key>", // defaults to process.env.ANTHROPIC_API_KEY
    libretto: {
      apiKey: "<Libretto API Key>", // defaults to process.env.LIBRETTO_API_KEY
    },
  });

  const completion = await anthropic.chat.completions.create({
    // Instead of a chat message array, you can pass objectTemplate instead.
    messages: objectTemplate([
      { role: "user", content: "Give a hearty welcome to our new user {name}" },
    ]),
    model: "claude-3-haiku-20240307",
    libretto: {
      // Uniquely identifies this prompt within your project.
      promptTemplateName: "ts-client-test-chat",
      // The parameters to fill in the template.
      templateParams: { name: "John" },
      //optional: key/value for passing any additional information for tracing
      context: { someKey: "somevalue" },
    },
  });

  console.log(completion.choices);
}

main();

Configuration

The following options may be set in the libretto object that has been added to the Anthropic client constructor:

• promptTemplateName: A default name to associate with prompts. If provided, this is the name that will be associated with any create call that's made without a libretto.promptTemplateName parameter.
• allowUnnamedPrompts: When set to true, every prompt will be sent to Libretto even if no prompt template name as been provided (either via the promptTemplateName option here or via the libretto.promptTemplateName parameter added to create).
• redactPii: When true, certain personally identifying information (PII) will be attempted to be redacted before being sent to the Libretto backend. See the pii package for details about the types of PII being detected/redacted. false by default.

Additional Parameters

The following parameters can be specified in the libretto object that has been added to the base Anthropic create call interface:

• templateParams: The parameters to use for template strings. This is a dictionary of key-value pairs.
• chatId: The id of a "chat session" - if the chat API is being used in a conversational context, then the same chat id can be provided so that the events are grouped together, in order. If not provided, this will be left blank.
• templateChat: The chat template to record for chat requests. This is a list of dictionaries with the following keys:
  • role: The role of the speaker. Either "system", "user" or "ai".
  • content: The content of the message. This can be a string or a template string with {} placeholders.
• templateText: The text template to record for non-chat completion requests. This is a string or a template string with {} placeholders.
• chainId: A UUID that groups related events together in a chain. For example, if you have a multi-step workflow where one LLM call's output feeds into another LLM call, you can use the same chainId to track the full sequence of events.
• feedbackKey: The optional key used to send feedback on the prompt, for use with sendFeedback() later. This is normally auto-generated, and the value is returned in the Anthropic response.
• context: This optional key/value map lets you send additional information along with your request such as internal tracing IDs, user IDs etc.

Maintaining Chat History

In scenarios where you are in the context of a chat dialog, you will want to maintain the history of the chat through each turn based request. Libretto has a reserved role that you can use for where you want to insert the full chat history. In your messages provided to objectTemplate you simply insert the following:

{
  role: "chat_history",
  content: "{chat_history}"
}

Finally, you have control over what gets inserted there in your request sent to Anthropic. In your templateParams you provide the aforementioned messages:

  templateParams: {
    chat_history: [
      {
        role: "user",
        content: "First user message",
      },
      {
        role: "assistant",
        content: "First response from model",
      },
    ],

Sending Feedback

Sometimes the answer provided by the LLM is not ideal, and your users may be able to help you find better responses. There are a few common cases:

• You might use the LLM to suggest the title of a news article, but let the user edit it. If they change the title, you can send feedback to Libretto that the answer was not ideal.
• You might provide a chatbot that answers questions, and the user can rate the answers with a thumbs up (good) or thumbs down (bad).

You can send this feedback to Libretto by calling sendFeedback(). This will send feedback about a prompt that was previously called, and let you review this feedback in the Libretto dashboard. You can use the feedback to develop new tests and improve your prompts.

import crypto from "crypto";
import { Anthropic, sendFeedback } from "@libretto/anthropic";

async function main() {
  const anthropic = new Anthropic();

  // Must be unique for each call to Anthropic
  const completion = await anthropic.chat.completions.create({
    // ...
  });

  // Maybe the user didn't like the answer, so ask them for a better one.
  const betterAnswer = await askUserForBetterResult(completion.choices[0].text);

  // If the user provided a better answer, send feedback to Libretto
  if (betterAnswer !== completion.choices[0].text) {
    // feedback key is automatically injected into Anthropic response object.
    const feedbackKey = completion.libretto?.feedbackKey;
    await sendFeedback({
      apiKey: "<Libretto API Key>", // defaults to process.env.LIBRETTO_API_KEY
      feedbackKey,
      // Better answer from the user
      betterResponse: betterAnswer,
      // Rating of existing answer, from 0 to 1
      rating: 0.2,
    });
  }
}

Note that feedback can include either rating, betterResponse, or both.

Parameters:

• rating - a value from 0 (meaning the result was completely wrong) to 1 (meaning the result was correct)
• betterResponse - the better response from the user