AI generations

At its heart, Agentflow is a tool to leverage advanced AI models to generate outputs as part of your workflows. Agentflow provides actions for generating text and structured data, and can be equipped with tools that bring external web services or local files and databases directly into your workflows.

Text generation

Text generation uses large language models (LLMs) to help you write content and code, plan meetings and write agendas, and even help you research and reason over complex problems.

Based on the following research paper, create a concise and detailed
scientific abstract that summarises the research findings:

{research}

<GenText as="abstract" model="anthropic:claude-3-5-sonnet-20240620" />

The <GenText /> action takes all the context immediately prior to the action (including the results of previous actions) and provides it as input to the action. In addition, the <GenText /> action accepts the following attributes:

attribute	type	description	required
as	string	Unique variable name.	✅
model	string	AI provider and model name. See AI providers.	✅
tools	array	An array of tools the LLM can use . See using tools.
stream	boolean	Whether to stream the response. Defaults true
options	object	LLM generation options. Docs TODO.

The result of the generation will be stored in the workflow's state as a text type using the name given in the as attribute. The value can then be referenced using that name in subsequent expressions.

Structured data

Object generation uses the same LLMs as text generation, but constrains and validates the output against a given schema, enabling the generation of structured data and objects.

Scrape the following web page and create a list of conference speakers:

{url}

<GenObject
  as="speakers"
  model="anthropic:claude-3-5-sonnet-20240620"
  output="array"
  schema={
    $.z.object({
    name: $.z.string().describe('Full name of the speaker'),
    company: $.z.string().describe('Comany name'),
    twitter: $.z.string().describe('LinkedIn social profile').optional(),
    twitter: $.z.string().describe('Twitter social profile').optional(),
    })
  } />

The <GenObject /> action works the same way as <GenText /> in that it uses the immediately preceding context as input. In addition, it accepts the following attributes:

attribute	type	description	required
as	string	Unique variable name.	✅
model	string	AI provider and model name. See AI providers.	✅
output	string	One of object, array, enum or no-schema. Defaults object
enum	array	Array of possible values to generate when output is enum.	✅ *
schema	ZodType	Zod schema that describes the shape of the object to generate.	✅ *
schemaName	string	Optional name of the output for additional LLM guidance.
schemaDescription	string	Optional description of the output for additional LLM guidance.
tools	array	An array of tools the LLM can use . See using tools.
options	object	LLM generation options. Docs TODO.

ℹ️ Output types *

When the output type is object or array then the schema attribute is required. When the output type is enum then a list of possible values is required in the enum attribute.

When the output type is no-schema, the LLM is free to determine its own schema, and neither the schema nor enum attributes are required.

ℹ️ Zod schemas

The schema attribute accepts an expression that returns a zod schema, using the $.z action helper. Ensuring the schema is property described helps guide the LLM to generate an accurate output.

The result of the generation will be stored in the workflow's state as a JSON type using the name given in the as attribute. Because the value is structured data, subsequent expressions can access, traverse and iterate over the value.

<Loop
  as="messages"
  until={$.index === speakers.filter(s => !!s.twitter).length}
  provide={{ speaker: speakers.filter(s => !!s.twitter)[$.index] }}>

  Write a tweet to the conference speaker thanking them for their
  presentation. Use a disturbing amount of emojis.

  Name: {speaker.name}
  Company: {speaker.company}
  Twitter: {speaker.twitter}

  <GenText as="tweet" model="anthropic:claude-3-5-sonnet-20240620" />
</Loop>

AI providers

Under the hood, Agentflow's AI generation actions are powered by Vercel's AI SDK, and so Agentflow is compatible with any AI providers supported by Vercel's SDK. Between Vercel and the open source community, all of the popular AI providers (such as OpenAI, Anthropic, Google and more) are fully supported, as well as tools such as Ollama for fully-local open source AI models. View the docs for full list of supported AI providers.

The Agentflow project initialiser installs the OpenAI provider by default. Additional providers can be installed using your package manager.

npm

npm install @ai-sdk/anthropic

yarn

yarn add @ai-sdk/anthropic

bun

bun add @ai-sdk/anthropic

Then ensure all providers are configured in your project's agentflow.config.js. The following configuration requires your .env file to contain OPENAI_API_KEY and ANTHROPIC_API_KEY environment variables.

import { defineConfig } from '@agentflow/core'
import { openai } from '@ai-sdk/openai'
import { anthropic } from '@ai-sdk/anthropic'

export default defineConfig({
  providers: {
    openai,
    anthropic,
  }
})

If a customised provider setup is needed, you can create new instances of providers with your own settings. This is particularly useful for using OpenAI compatible providers in your workflows.

import { defineConfig } from '@agentflow/core'
import { createOpenAI } from '@ai-sdk/openai'

export default defineConfig({
  providers: {
    openai: createOpenAI({
      apiKey: process.env.COMPANY_OAI_KEY,
      compatibility: 'strict',
    }),
    together: createOpenAI({
      apiKey: process.env.TOGETHER_API_KEY,
      baseURL: 'https://api.together.xyz/v1',
      compatibility: 'compatible',
    }),
  }
})

To use a specific model with the <GenText /> and <GenObject /> actions, set the model attribute to a string with the format providerId:modelId. Always refer to the documentation of the provider for the available model IDs.

Write the outline of a short survival novel set in a cyberpunk dystopian world.

<GenText
  as="outline"
  model="together:meta-llama/Meta-Llama-3.1-405B-Instruct-Turbo" />

Using tools

TODO