ai-fun-chain

Break complex tasks up into small pieces that are easier for LLMs to handle into a chain. Link them with non-ai functions to augment, parse, or validate responses. Function calling supported.

- Aggregates token usage throughout chain
- Automatic retries
- Fully configurable, at both chain and link level

Currently only works with Open AI. May support other models in the future.

Install

Set OPENAI_API_KEY in your environment variables.

npm install ai-fun-chain

Usage

``
import { FunctionLink, ModelLink, executeChain } from "ai-fun-chain";
`

ai-fun-chain works by defining a chain of links. Each link is either a ModelLink or a FunctionLink. ModelLinks are configuration objects that define the request made to the LLM. FunctionLinks are just regular functions. ai-fun-chain makes it easy to pass results from previous links into prompts for later links.

`
const result = await executeChain([
getRandomletter, // Function Link
getRandomName, // Model Link
getGender, // Model Link (function calling)
addKunOrChanToName, // Function Link
]);
`

$3

ModelLink is an extension of Open AI's CreateChatCompletionRequest, so you can pass in any of the properties defined in the Open AI docs.

You can pass in a normal messages property, or you can use templates. templates is an array of MessageTemplate. A MessageTemplate is similar to the normal message you would pass to Open AI, but instead of content it has contentBlocks.

contentBlocks are useful for programtically including blocks of text in the prompt. Each content block is composed of a template and an include property. If include is true, the template will be included in the prompt. If include is false, the template will not be included in the prompt. This is useful for conditionally including text in the prompt based on the result of a previous link.

The modelLink also has extra parameters to control cleanup functions. autoNewLineContent defaults to true. If set to false, it will use a space. You do not need to add leading/trailing spaces between contentBlocks. removeDoubleSpaces defaults to true. If set to true, it will remove double spaces.

`

const nameShouldBeFourLetters = true;

const getRandomName: ModelLink = {
name: "getRandomName",
temparature: 0,
autoNewLineContent: false, // defaults to true. If set to false, it will use a space. You do not need to add leading/trailing spaces between contentBlocks.
removeDoubleSpaces: true, // defaults to true.
contentBlocks: [
{
template: Come up with one first name that start with the letter {{getRandomletter}},
include: true,
},
{
template: The name should be four letters long,
include: nameShouldBeFourLetters,
},
],
};

`

Full Example

`

// Define links
const getRandomletter: FunctionLink = () => {
const letters = "abcdefghijklmnopqrstuvwxyz";
const randomLetter = letters[Math.floor(Math.random() * letters.length)];
return randomLetter;
};

const getRandomName: ModelLink = {
name: "getRandomName",
model: "gpt-3.5-turbo-0613",
temperature: 0.9,
template: [
{
content: Come up with one first name that start with the letter {{getRandomletter}},
include: true,
},
],
};

const getGender: ModelLink = {
name: "getGender",
retries: 2,
model: "gpt-3.5-turbo-0613",
temperature: 0,
template: [
{
content: What is the gender of {{getRandomName}},
include: true,
},
],
functions: [
{
name: "getGender",
description: "Gets the gender of a name",
parameters: {
type: "object",
properties: {
gender: {
type: "string",
description: "Either boy or girl",
enum: ["boy", "girl"],
},
name: {
type: "string",
description: "The name",
},
},
},
},
],
function_call: {
name: "getGender",
},
};

const addKunOrChanToName = ({
gender,
name,
}: {
gender: "boy" | "girl";
name: string;
}) => {
if (gender === "boy") {
return ${name}-kun;
} else {
return ${name}-chan;
}
};

// Execute the chain
const example = async () => {
const result = await executeChain([
getRandomletter, // Function Link
getRandomName, // Model Link
getGender, // Model Link (function calling)
addKunOrChanToName, // Function Link
]);
console.log(result);
return result;
};
example()

`

Output

executeChain outputs the following object, so you can easily access the final result and aggregated usage across each link, as well as the results of each link.

`

{
finalResult: 'Ivan-chan',
totalTokens: 108,
totalPromptTokens: 90,
totalCompletionTokens: 18,
chatCompletionResponses: {
getRandomName: {
id: 'chatcmpl-7TNbk54yIUba6wvtPoT0z5HTNSbMC',
object: 'chat.completion',
created: 1687236616,
model: 'gpt-3.5-turbo-0613',
choices: [Array],
usage: [Object]
},
getGender: {
id: 'chatcmpl-7TNbkKZ9n5wmprfpXQL8YNW35CDv0',
object: 'chat.completion',
created: 1687236616,
model: 'gpt-3.5-turbo-0613',
choices: [Array],
usage: [Object]
}
},
linkResults: {
getRandomletter: 'i',
getRandomName: 'Ivan',
getGender: { name: 'Ivan' },
addKunOrChanToName: 'Ivan-chan'
}
}

`

Pass results from previous links to a template in a later link

For prompt templates, you can just refer to the name of the previous link in the template. For example, if you want to use the result of the getRandomName link in the template for the getGender link, you can just use the name of the link in the template.

`

{{getRandomName}}

`

If one of your model links outputs JSON through function calling, you can reference a specific value with dot notation

`

{{getGender.gender}}

`

If you are passing the result of a function call directly into a Function Link, you simply define the arguments in that function to match the expected result of the previous link

`

const addKunOrChanToName = ({ gender, name }: {
gender: "boy" | "girl";
name: string;
}) => {
...
}

`

Retries

You can pass retries in at each link, or you can pass in a global retry quantity in the executeChain config.

`

const result = await executeChain(
[
getRandomName, // ModelLink
determineKunOrChan, // FunctionLink
addKunOrChanToName, // FunctionLink
writeStoryAboutName, // ModelLink
],
{
retries: 2,
}
);

`

`

``