JS/TS library based on the Stanford DSP paper. Create and compose efficient prompts using prompt signatures. Reasoning + Function Calling, RAG and more.
๐ต ๐ฆ ๐ฅ โค๏ธ ๐๐ผ
LLMClient is an easy to use library build around "Prompt Signatures" from the Stanford DSP
paper. Prompt signatures like question:string -> answer:string
are automatically compiled into complete and efficient prompts. Build powerful workflows using components like RAG, ReAcT, Chain of Thought, Function calling, Agents, etc all built on prompt signatures and easy to compose together to build whatever you want.
"Write a simple search query that will help answer a complex question."
context?:string[] "may contain relevant facts", question -> query
Efficient type-safe prompts are auto-generated from a simple signature. A prompt signature is made of a "description" inputField:type -> outputField:type"
. The idea behind prompt signatures is based off work done in the "Demonstrate-Search-Predict" paper.
Provider | Best Models | Tested |
---|---|---|
OpenAI | GPT: 4, 3.5/4-Turbo | ๐ข 100% |
Azure OpenAI | GPT: 4, 3.5/4-Turbo | ๐ข 100% |
Together | Several OSS Models | ๐ข 100% |
Cohere | Command, Command Nightly | ๐ก 100% |
Anthropic | Claude 2, Claude 3 | ๐ก 50% |
Google Vertex | Palm, Bison | ๐ก 50% |
Google Gemini | Gemini 1.0 | ๐ก 50% |
Hugging Face | OSS Model | ๐ก 50% |
Groq | Lama2-70B, Mixtral-8x7b | ๐ก 50% |
import { AI, ChainOfThought, OpenAIArgs } from 'llmclient';
const textToSummarize = `
The technological singularityโor simply the singularity[1]โis a hypothetical future point in time at which technological growth becomes uncontrollable and irreversible, resulting in unforeseeable changes to human civilization.[2][3] ...`;
const ai = AI('openai', { apiKey: process.env.OPENAI_APIKEY } as OpenAIArgs);
const gen = new ChainOfThought(
ai,
`textToSummarize -> shortSummary "summarize in 5 to 10 words"`
);
const res = await gen.forward({ textToSummarize });
console.log('>', res);
Work in progress.
Use the tsx
command to run the examples it makes node run typescript code. It also support using a .env
file to pass the AI API Keys as opposed to putting them in the commandline.
OPENAI_APIKEY=openai_key npm run tsx ./src/examples/marketing.ts
Example | Description |
---|---|
customer-support.ts | Extract valuable details from customer communications |
food-search.ts | Use multiple APIs are used to find dinning options |
marketing.ts | Generate short effective marketing sms messages |
fibonacci.ts | Use the JS code interpreter to compute fibonacci |
summarize.ts | Generate a short summary of a large block of text |
chain-of-thought.ts | Use chain-of-thought prompting to answer questions |
rag.ts | Use multi-hop retrieval to answer questions |
react.ts | Use function calling and reasoning to answer questions |
Often you need the LLM to reason through a task and fetch and update external data related to this task. This is where reasoning meets function (API) calling. It's built-in so you get all of the magic automatically. Just define the functions you wish to you, a schema for the response object and thats it.
There are even some useful built-in functions like a Code Interpreter
that the LLM can use to write and execute JS code.
We support providers like OpenAI that offer multiple parallel function calling and the standard single function calling.
Function | Description |
---|---|
Code Interpreter | Used by the LLM to execute JS code in a sandboxed env. |
Embeddings Adapter | Wrapper to fetch and pass embedding to your function |
Large language models (LLMs) are getting really powerful and have reached a point where they can work as the backend for your entire product. However there is still a lot of manage a lot of complexity to manage from using the right prompts, models, etc. Our goal is to package all this complexity into a well maintained easy to use library that can work with all the LLMs out there. Additionally we are using the latest research to add useful new capabilities like DSP to the library.
// Pick a LLM
const ai = new OpenAI({ apiKey: process.env.OPENAI_APIKEY } as OpenAIArgs);
// Can be sub classed to build you own memory backends
const mem = new Memory();
const cot = new ChainOfThought(ai, `question:string -> answer:string`, { mem });
const res = await cot.forward({ question: 'Are we in a simulation?' });
const res = await ai.chat([
{ role: "system", content: "Help the customer with his questions" }
{ role: "user", content: "I'm looking for a Macbook Pro M2 With 96GB RAM?" }
]);
// define one or more functions and a function handler
const functions = [
{
name: 'getCurrentWeather',
description: 'get the current weather for a location',
parameters: {
type: 'object',
properties: {
location: {
type: 'string',
description: 'location to get weather for'
},
units: {
type: 'string',
enum: ['imperial', 'metric'],
default: 'imperial',
description: 'units to use'
}
},
required: ['location']
},
func: async (args: Readonly<{ location: string; units: string }>) => {
return `The weather in ${args.location} is 72 degrees`;
}
}
];
const cot = new ReAct(ai, `question:string -> answer:string`, { functions });
const ai = new OpenAI({ apiKey: process.env.OPENAI_APIKEY } as OpenAIArgs);
ai.setOptions({ debug: true });
We're happy to help reach out if you have questions or join the Discord twitter/dosco
Improve the function naming and description be very clear on what the function does. Also ensure the function parameter's also have good descriptions. The descriptions don't have to be very long but need to be clear.
You can pass a configuration object as the second parameter when creating a new LLM object
const apiKey = process.env.OPENAI_APIKEY;
const conf = OpenAIBestConfig();
const ai = new OpenAI({ apiKey, conf } as OpenAIArgs);
const conf = OpenAIDefaultConfig(); // or OpenAIBestOptions()
conf.maxTokens = 2000;
const conf = OpenAIDefaultConfig(); // or OpenAIBestOptions()
conf.model = OpenAIModel.GPT4Turbo;
Generated using TypeDoc