@dexaai/dexter NPM

Dexter

Dexter is a set of mature LLM tools used in production at Dexa, with a focus on real-world RAG (Retrieval Augmented Generation).

If you're a TypeScript AI engineer, check it out! 😊

Features

production-quality RAG
extremely fast and minimal
handles caching, throttling, and batching for ingesting large datasets
optional hybrid search w/ SPLADE embeddings
minimal TS package w/ full typing
uses fetch everywhere
supports Node.js 18+, Deno, Cloudflare Workers, Vercel edge functions, etc
full docs

Install

npm install @dexaai/dexter

This package requires node >= 18 or an environment with fetch support.

This package exports ESM. If your project uses CommonJS, consider switching to ESM or use the dynamic import() function.

Usage

This is a basic example using OpenAI's text-embedding-ada-002 embedding model and a Pinecone datastore to index and query a set of documents.

import 'dotenv/config';
import { EmbeddingModel } from '@dexaai/dexter/model';
import { PineconeDatastore } from '@dexaai/dexter/datastore/pinecone';

async function example() {
  const embeddingModel = new EmbeddingModel({
    params: { model: 'text-embedding-ada-002' },
  });

  const store = new PineconeDatastore({
    contentKey: 'content',
    embeddingModel,
  });

  await store.upsert([
    { id: '1', metadata: { content: 'cat' } },
    { id: '2', metadata: { content: 'dog' } },
    { id: '3', metadata: { content: 'whale' } },
    { id: '4', metadata: { content: 'shark' } },
    { id: '5', metadata: { content: 'computer' } },
    { id: '6', metadata: { content: 'laptop' } },
    { id: '7', metadata: { content: 'phone' } },
    { id: '8', metadata: { content: 'tablet' } },
  ]);

  const result = await store.query({ query: 'dolphin' });
  console.log(result);
}

Docs

See the docs for a full usage guide and API reference.

Examples

To run the included examples, clone this repo, run pnpm install, set up your .env file, and then run an example file using tsx.

Environment variables required to run the examples:

OPENAI_API_KEY - OpenAI API key
PINECONE_API_KEY - Pinecone API key
PINECONE_BASE_URL - Pinecone index's base URL
- You should be able to use a free-tier "starter" index for most of the examples, but you'll need to upgrade to a paid index to run the any of the hybrid search examples
- Note that Pinecone's free starter index doesn't support namespaces, deleteAll, or hybrid search :sigh:
SPLADE_SERVICE_URL - optional; only used for the chatbot hybrid search example

Basic

npx tsx examples/basic.ts

source

Caching

npx tsx examples/caching.ts

source

Redis Caching

This example requires a valid REDIS_URL env var.

npx tsx examples/caching-redis.ts

source

AI Function

This example shows how to use createAIFunction to handle function and tool_calls with the OpenAI chat completions API and Zod.

npx tsx examples/ai-function.ts

source

AI Runner

This example shows how to use createAIRunner to easily invoke a chain of OpenAI chat completion calls, resolving tool / function calls, retrying when necessary, and optionally validating the resulting output via Zod.

Note that createAIRunner takes in a functions array of AIFunction objects created by createAIFunction, as the two utility functions are meant to used together.

npx tsx examples/ai-runner.ts

source

Chatbot

This is a more involved example of a chatbot using RAG. It indexes 100 transcript chunks from the Huberman Lab Podcast into a hybrid Pinecone datastore using OpenAI ada-002 embeddings for the dense vectors and a HuggingFace SPLADE model for the sparse embeddings.

You'll need the following environment variables to run this example:

OPENAI_API_KEY
PINECONE_API_KEY
PINECONE_BASE_URL
- Note: Pinecone's free starter indexes don't seem to support namespaces or hybrid search, so unfortunately you'll need to upgrade to a paid plan to run this example. See Pinecone's hybrid docs for details on setting up a hybrid index, and make sure it is using the dotproduct metric.
SPLADE_SERVICE_URL
- Here is an example of how to run a SPLADE REST API, which can be deployed to Modal or any other GPU-enabled hosting provider.