Getting started

Let’s start by creating our chains backend! Create a repository with a chains folder inside it. Our CLI will automatically deploy all chains we create in this folder.

Speaking of CLI, let’s install it!

npm install -g @relevanceai/chain

Next, let’s authenticate:

relevance login

After following the instructions, you’ll be set up and ready to go! Make sure to also add your Open AI key.

Want to skip to the end? We have a Github repo with a ChatGPT style frontend and chains folder ready-made! Click here.

Loading your PDFs

The first thing we need to do is index our PDFs in some sort of database. Ideally, we actually want this to be a vector database - where we can store embeddings for the PDF text.

What is a vector database?

Vector databases allow you to store and query “embeddings” alongisde your metadata. You can think of embeddings as numerical representations of your text, capturing the many semantic dimensions.

By querying embeddings against embeddings (vector similarity search), you’re able to find content based on meaning. This is in contrast to traditional search techniques such as keyword matching.

For example, in a vector database, the following two sentences would have similar embeddings:

1. There are many ways to skin a cat.

2. You can achieve your goal through a variety of possible methods.

This is because the meaning of the two sentences is similar, even though the words are different. Traditional keyword matching would not find these two sentences similar.

How do they help with LLMs?

A very common technique is to use vector databases to provide additional context to LLMs. We can search the vector database with the user’s input and retrieve relevant information. This can then be added into the prompt.

Loading PDFs into our built-in vector database

Good news! Relevance has a built in vector database. We call these “datasets”. We want to:

  1. Scrape PDFs from URLs
  2. Split up the PDF text into chunks
  3. Create embeddings with OpenAI for each chunk
  4. Upload the chunks into a Relevance AI dataset

We’ll be creating more tooling and content around this process, but for now you can run this Google Colab demo. Make sure to fill in your authentication details at the top of the notebook.

Planning the chain

Now let’s build the chain. We want it to:

  1. Receive a question
  2. Use the question to search our vector database for relevant text chunks from our PDFs
  3. Feed those chunks into the LLM as context, and have the LLM answer our question

Additionally, we want the chain to be able to consider the chat history when answering questions. This will allow us to have a conversation, like with ChatGPT.

To do that, we need to alter our chain to also accept a history param. The trick is, we want to use that history to re-word the user’s question.

So for example, if their question is “What’s that in fahrenheit?” - standalone, this will make no sense to the LLM. However, if we have just asked the LLM “what is the boiling point of water in celsius?”, by passing in the history, we can re-word the new question to “what is the boiling point of water in fahrenheit?“. An LLM is able to answer that successfully!

So with that in mind, our chain should look like:

  1. Receive a question and chat history
  2. If chat history exists, use an LLM to reword the question based on the history
  3. Use the question to search our vector database for relevant text chunks from our PDFs
  4. Feed those chunks into the LLM as context, and have the LLM answer our question

Defining our chain input

Let’s start by creating a new file in the chains folder called pdf-chat.js. This will be our chain!

We start by importing the defineChain method from the @relevanceai/chain package:

import { defineChain } from '@relevanceai/chain';

const chain = defineChain({
    title: 'PDF Q&A',

Next, we want to define the input parameters for this chain. As discussed, those should be question and history. We then give the params types, based on JSONSchema.

const chain = defineChain({
    title: 'PDF Q&A',
    params: {
        question: {
            type: 'string',
        history: {
            type: 'array',
            items: {
                type: 'object',
                properties: {
                    message: {
                        type: 'string',
                    role: {
                        type: 'string', // either 'user' or 'ai'

Our history will be an array of objects, with the message and who sent the message (role).

Defining our chain steps

The meat and potatoes of any chain are the “steps”. These are the transformations that will execute on running a chain. They run in our hosted Relevance AI run time.

We have a whole library of pre-built steps, which you can see here! We also also support more advanced use cases where you can self-host any code as a chain step! Speak to our team if this is you.

To build this chain, we’ll need to use the following steps as specified earlier:

  1. Modify question based on chat history
  2. Search vector database for relevant text chunks
  3. Feed chunks into LLM and get answer

We do this in the setup function in defineChain, which receives params and a step function as arguments. The step function is used to define the steps of the chain.

Note: the setup function only allows you to declare each step of a chain. No Javascript you write outside of these steps, except for your returned output will be executed.

However, you can use the code function provided in setup to run custom Javascript. You can also use the runIf function to conditionally run steps, as well as forEach to iterate over arrays. Learn more.

const chain = defineChain({
    setup({ params, step, code, runIf }) {
        // steps go here
        const { question, history } = params;

        // use an LLM to modify the question
        const { modifiedQuestion } = runIf(history, () => {
            // wrap our custom Javascript function in our code utility
            const { transformed: reducedHistory } = code(
                { history },
                ({ history }) =>
                        (acc, item) =>
                                item.role === 'ai' ? 'ASSISTANT' : 'USER'
                            }: ${item.message}\n`,

            const { answer: modifiedQuestion } = step('prompt_completion', {
                prompt: `The human is having a conversation with an assistant, where the assistant will query a knowledgebase to get the context needed to provide an answer. You must use the chat history and the follow-up input to rephrase the question such that it can be used on its own. This means the standalone question must be include context from the chat history necessary to provide an answer.

            For example, if the chat history contains information about the temperature and the user asks to convert it into another unit - provide that in the standalone question. Or if the user asks a question about a subject from the history, reference that subject with detail.

            Chat History: ${reducedHistory}

            Follow Up Input: ${question}.

            Standalone question:`,

            return { modifiedQuestion };

        const { transformed: questionTransform } = code(
            { history, modifiedQuestion, question },
            ({ history, modifiedQuestion, question }) => {
                return {
                        history && history.length
                            ? modifiedQuestion
                            : (question as string),

        const questionToUse = questionTransform.question;

        // Step 2. search vector database for relevant text chunks
        const { results: vectorSearchResults } = step('search', {
            query: questionToUse,
            dataset_id: 'knowledge_base',
            vector_field: 'embedding_vector_',
            model: 'text-embedding-ada-002',
            page_size: 3,

        const { transformed: searchTransformed } = code(
            { vectorSearchResults },
            ({ vectorSearchResults }) => {
                return {
                        (result: Record<string, any>) => ({
                            text: result.text,
                            url: result.pdf_url,

        const resultsMapped = searchTransformed.results;

        // Step 3. ask LLM, injecting vector database results as context
        const { answer } = step('prompt_completion', {
            prompt: `Use the following pieces of context to answer the question at the end in JSON. If you don't know the answer, just say that you don't know, don't try to make up an answer. Provide a reference to the "pdf_url" of the items used to give the answer. The response must be JSON using the format "{ answer, references: [url] }".


Question: ${questionToUse}

Helpful Answer JSON:`,

        // finally, let's convert the answer to JSON
        const { transformed: answerJSON } = code({ answer }, ({ answer }) => {
            let returnAnswer = {
                answer: "I couldn't answer the question. Please try again.",
                references: [],

            try {
                // try to convert the LLM's answer into JSON
                returnAnswer = JSON.parse(answer);
            } catch (err) {}

            return returnAnswer;

        return {
            answer: answerJSON,


Let’s show off how to run this bad boy client side! I’ve created a simple Nuxt.js application containing the above chain and a chat frontend. You’re free to clone this template and use it as a starting point for your own projects.

Check it out here!

To get it working with your own chain, make sure to:

relevance login
relevance deploy

The relevant code exists in the useChain composable. We run the chain by it’s ID, which is the file name of the chain.

import { Client } from '@relevance-ai/chain';

// we can import the type for the chain to get type safety (optional)
import type PdfQa from './chains/pdf-qa.ts';

const client = new Client({
    regionId: REGION_ID,
    projectId: PROJECT_ID,

const response = await client.runChain<typeof PdfQa>('pdf-qa', {
    question: clonedUserInput,
    history: chatHistory.value,

const responseAnswer = response.answer;