TypeScript Prompt SDK Reference

`AutoblocksPromptManager`

Below are the arguments that can be passed when initializing the prompt manager:

name	required	default	description
`id`	true		The ID of the prompt.
`version.major`	true		Must be pinned to a specific major version.
`version.minor`	true		Can be one of: a specific minor version, the string `"latest"`, or a weighted list. If a weighted list, the minor version will be chosen randomly at runtime for each `exec` call according to the weights.
`apiKey`	false	`AUTOBLOCKS_API_KEY` environment variable	Your Autoblocks API key.
`refreshInterval`	false	`{ seconds: 10 }`	How often to refresh the latest prompt. Only relevant if the minor version is set to `"latest"` or `"latest"` is used in the weighted list.
`refreshTimeout`	false	`{ seconds: 30 }`	How long to wait for the latest prompt to refresh before timing out. A refresh timeout will not throw an uncaught error. An error will be logged and the background refresh process will continue to run at its configured interval.
`initTimeout`	false	`{ seconds: 30 }`	How long to wait for the prompt manager to be ready (when calling `init()`) before timing out.

import { AutoblocksPromptManager } from '@autoblocks/client/prompts';

const mgr = new AutoblocksPromptManager({
  id: 'text-summarization',
  version: {
    major: '1',
    minor: '0',
  },
});

When using a weighted list, the weights do not need to add up to 100. The weights simply represent the relative probability of choosing a given minor version. For example, the weighted list:

minor: [
  {
    version: 'latest',
    weight: 10,
  },
  {
    version: '0',
    weight: 90,
  },
]

is the same as:

minor: [
  {
    version: 'latest',
    weight: 1,
  },
  {
    version: '0',
    weight: 9,
  },
]

`exec`

Starts a prompt execution context by creating a new PromptExecutionContext instance.

const response = await mgr.exec(async ({ prompt }) => {
  ...
});

An instance of this class is created every time a new execution context is started with the exec function. It contains a frozen copy of the prompt manager's in-memory prompt at the time exec was called. This ensures the prompt is stable for the duration of an execution, even if the in-memory prompt on the manager instance is refreshed mid-execution.

`params`

An object with the prompt's parameters.

const response = await mgr.exec(async ({ prompt }) => {
  const params: ChatCompletionCreateParamsNonStreaming = {
    model: prompt.params.model,
    temperature: prompt.params.temperature,
    ...
  };
});

`renderTemplate`

The renderTemplate function accepts a template ID and parameters and returns the rendered template as a string.

name	required	description
template	true	The ID of the template to render.
params	true	The parameters to pass to the template. These values are used to replace the template parameters wrapped in double curly braces.

const response = await mgr.exec(async ({ prompt }) => {
  // Use `prompt.renderTemplate` to render a template
  console.log(prompt.renderTemplate(
    {
      template: 'util/language',
      params: {
        // Replaces "{{ language }}" with "Spanish"
        language: 'Spanish',
      },
    },
  ));
  // Logs "Always respond in Spanish."
});

`renderTool`

The renderTool function accepts a tool name and parameters and returns the rendered tool as an object in the JSON schema format that OpenAI expects.

name	required	description
tool	true	The name of the tool to render.
params	true	The parameters to pass to the tool. These values are used to replace the tool parameters wrapped in double curly braces.

const response = await mgr.exec(async ({ prompt }) => {
  // Use `prompt.renderTool` to render a tool
  console.log(prompt.renderTool(
    {
      tool: 'MyTool',
      params: {
        // Replaces "{{ language }}" with "Spanish"
        language: 'Spanish',
      },
    },
  ));
});

`track`

This function returns tracking information for current prompt execution context. It is meant to be sent as the promptTracking property on an LLM response event.

tracer.sendEvent('ai.response', {
  properties: {
    response,
  },
  promptTracking: prompt.track(),
});