LLM Agents

A library to create LLM Agents with Node.js

Usage

``bash
$ bun install llm-agents

or the legacy way

$ npm install llm-agents
`

$3

Extends the LLMAction class to define an action with:

- name
- usage
- parameters

`ts
import { execSync } from 'child_process';

import { Action, ActionFeedback } from 'llm-agents';

type ExecuteShellCommandActionParametersNames = 'command';

export class ExecuteShellCommandAction extends LLMAction {
public name = 'executeShellCommand';
public usage = 'execute a shell command';
public parameters = [
{
name: 'command' as const,
usage: 'command to execute',
},
];

protected async executeAction(
parameters: Record
): Promise {
const { command } = parameters;

try {
const result = execSync(command);

return {
message: $ ${command}\n${result.toString()},
type: 'success',
};
} catch (error) {
return {
message: $ ${command}\n${error.message},
type: 'error',
};
}
}
}
`

Examples:

- ListFilesActions
- CopyFileActions
- CreateDirectoryActions

$3

Extends the LLMAgent class to define an Agent with:

- template content
- template formating
- actions (optionnal)

`ts
import { PromptTemplate } from 'langchain/prompts';

import { LLMAction, FileCache, LLMAgent } from 'llm-agents';

import {
CreateDirectoryAction,
CopyFileAction,
ListFilesAction,
} from './actions';

export class BackupAgent extends LLMAgent {
private source: string;
private destination: string;

protected template = new PromptTemplate({
template: You are a backup agent capable of moving files from one place to another.

You task is to backup the files containing the typescript code of a project.
You need to copy the files containing the code into a backup directory.
You can only move one file after another.

You can use the following actions:

BEGIN ACTIONS DEFINITION

{actions}

END ACTIONS DEFINITION

The last action result was:

BEGIN LAST ACTION RESULT

{feedback}

END LAST ACTION RESULT

Ensure you are copying every files in every directories from the source.
The source directory is {source} and the destination directory is {destination}.
Skip node_modules directories.

Start by a sentence summarizing the current state of your task according to the last action result.
Then, answer with the actions you want to execute.
,
inputVariables: ['source', 'destination', 'actions', 'feedback'],
});

constructor({
source,
destination,
}: {
source: string;
destination: string;
}) {
super({
actions: [
new ListFilesAction(),
new CopyFileAction(),
new CreateDirectoryAction(),
],
cacheEngine: new FileCache(),
});

this.source = source;
this.destination = destination;
}

protected async formatPrompt({
actions,
feedbackSteps,
}: {
actions: string;
feedbackSteps: string[];
}) {
return this.template.format({
source: this.source,
destination: this.destination,
actions,
feedback: feedbackSteps.join('\n\n'),
});
}
}
`

Examples:

- BackupAgent

$3

During development phase, it's advised to use the filesystem cache.

It will save both prompt and answer for each step.

The cache key is a hash of the prompt so if the prompt does not change, the cached answer will be used directly.

You can also debug your prompts and answers in the .cache/ folder. In development mode, they will be prefixed by the step number (e.g. 0-92957a2b27-answer.txt)

Tests

Unit tests can be run with Bun: bun test

See tests/unit

Integration tests consist in a benchmark of LLMs agents:

- BackupAgent: bun tests/prompt-engineering/backup-agent/run-backup-agent-test.ts`