@ai-refiner/refiner-js

refiner-js

The Refiner Javascript package can be used to convert and store text and metadata as vector embeddings. Embeddings are generated using OpenAI and stored as vectors in Pinecone. Stored embeddings can then be "queried" using the search method. Matched embeddings contain contextually relevant metadata.

Installation

npm install @ai-refiner/refiner-js

OpenAI and Pinecone API Keys.

You'll need API keys for OpenAI and Pinecone.

Once you have your API keys, you can either set local ENV variables in a shell:

# Required variables
export PINECONE_API_KEY="API_KEY"
export PINECONE_ENVIRONMENT_NAME="ENV_NAME"
export OPENAI_API_KEY="API_KEY"

# Optional variables
export OPENAI_ADA_200_DEFAULT_DIMENSIONS="1536"
export OPENAI_TEXT_EMBEDDING_MODEL="text-embedding-ada-002"

or you can create a .env (dotenv) config file and pass in the file path when initializing the Embeddings class:

import { Embeddings } from "@ai-refiner/refiner-js";
const embeddings = new Embeddings("/path/to/.env");

Your .env file should follow key/value format:

PINECONE_API_KEY="API_KEY"
PINECONE_ENVIRONMENT_NAME="ENV_NAME"
OPENAI_API_KEY="API_KEY"
OPENAI_ADA_200_DEFAULT_DIMENSIONS="1536"
OPENAI_TEXT_EMBEDDING_MODEL="text-embedding-ada-002"

Create Index

const indexes = new Indexes("./.env");
const index = await indexes.createIndex("new-index");
// creating index: new-index

// Index is being created... Please wait 30 seconds before trying to store embeddings.
// { success: true }

Create Embeddings

const embeddings = new Embeddings("/path/to/.env");
const payload: PayloadItem[] = [
  { id: "1", text: "hello world", metadata: { key: "value" } },
];
embeddings.create(payload, "test-index");
// {'upserted_count': 1}

Semantic Search

const embeddings = new Embeddings("/path/to/.env");
const limit = 10;
embeddings.search("hello", "test-index", limit);
// {'matches': [...]}

Loaders - getDocumentFromURL

// Get web page content from a URL and create a document.
let loaders = new Loaders();
let data = await loaders.getDocumentFromURL("https://news.yahoo.com/");
// [
//  Document {
//    pageContent:

Loaders - getDocumentFromPDF

// Get web page content from a PDF filepath or blob and create a document.
let path = "/path/to/PDF/example.pdf";
let data = await loader.getDocumentFromPDF(path);
//  Document {
//    pageContent:

Transformers

// Split the document text and create embeddings
const embeddings = new Embeddings("/path/to/.env");
let transformers = new Transformers();
let texts = await transformers.splitText(data);
let vectors = [] as any;
texts.map(async (item, index) => {
    // remove loc from metadata. It's not needed and throws a PineconeError when used.
    if ("loc" in item.metadata) delete item.metadata.loc;

    let vector = {
        id: `id-${index}`,
        text: item.pageContent,
        metadata: { ...item.metadata, pageContent: item.pageContent },
    };

    vectors.push(vector);
});
const created = await embeddings.create(vectors, "test-index");
// { upsertedCount: 251 }
...

Document Chatbot Example

// Ask the document questions. Format a Q/A style prompt.
// Stream the completion results
import {
  ChatCompletionCreateParams,
  EmbeddingCreateParams,
} from "openai/resources";
const question = "What are the headlines of todays news?";

let results = await embeddings.search(question, "test-index", 10);
const openaiClient = new RefinerOpenAIClient("OPENAI_API_KEY");

const document = results.matches.map((m) => m.metadata?.pageContent).join("\n");

const prompt = `
  Q: ${question}\n
  Using this document answer the question as a friendly chatbot that knows about the details in the document.
  You can answer questions only with the information in the documents you've been trained on.
  ${document}\n
  A:
  `;

const payload: ChatCompletionCreateParams = {
  model: "gpt-3.5-turbo",
  messages: [{ role: "user", content: "Hello, world!" }],
  stream: true,
};

const stream = await openaiClient.createCompletion(
  payload as ChatCompletionCreateParams
);
for await (const part of stream as any) {
  console.log(part.choices[0].delta);
}

CLI

You can install the CLI to create and search your vectors.

pip install refiner-cli

The --help option can be used to learn about the create and search commands.

refiner --help
refiner create --help
refiner search --help