Picture by Editor | ChatGPT
# Introduction
Prepared for a sensible walkthrough with little to no code concerned, relying on the method you select? This tutorial reveals tie collectively two formidable instruments — OpenAI‘s GPT fashions and the Airtable cloud-based database — to prototype a easy, toy-sized retrieval-augmented era (RAG) system. The system accepts question-based prompts and makes use of textual content knowledge saved in Airtable because the information base to provide grounded solutions. Should you’re unfamiliar with RAG techniques, or need a refresher, don’t miss this article collection on understanding RAG.
# The Elements
To apply this tutorial your self, you may want:
- An Airtable account with a base created in your workspace.
- An OpenAI API key (ideally a paid plan for flexibility in mannequin alternative).
- A Pipedream account — an orchestration and automation app that allows experimentation beneath a free tier (with limits on every day runs).
# The Retrieval-Augmented Era Recipe
The method to construct our RAG system isn’t purely linear, and a few steps will be taken in numerous methods. Relying in your stage of programming information, chances are you’ll go for a code-free or practically code-free method, or create the workflow programmatically.
In essence, we’ll create an orchestration workflow consisting of three elements, utilizing Pipedream:
- Set off: just like an online service request, this factor initiates an motion movement that passes by the subsequent components within the workflow. As soon as deployed, that is the place you specify the request, i.e., the person immediate for our prototype RAG system.
- Airtable block: establishes a connection to our Airtable base and particular desk to make use of its knowledge because the RAG system’s information base. We’ll add some textual content knowledge to it shortly inside Airtable.
- OpenAI block: connects to OpenAI’s GPT-based language fashions utilizing an API key and passes the person immediate alongside the context (retrieved Airtable knowledge) to the mannequin to acquire a response.
However first, we have to create a brand new desk in our Airtable base containing textual content knowledge. For this instance, I created an empty desk with three fields (ID: one-line textual content, Supply: one-line textual content, Content material: lengthy textual content), after which imported knowledge from this publicly out there small dataset containing textual content with primary information about Asian international locations. Use the CSV and hyperlink choices to import the information into the desk. Extra details about creating tables and importing knowledge is in this text.
After getting your information base prepared in Airtable, let’s swap to Pipedream for the remainder of the method! Register when you haven’t carried out so but, and log in. There are a number of methods to provoke a workflow. Considered one of them is by describing what you wish to create to Pipedream’s AI agent, which is intuitive however might lead to errors needing fixes, as we’ll illustrate shortly. Pipedream has two parallel UIs: one for manually constructing and enhancing workflows, and one for conversations with its AI agent, known as String.
Pipedream’s AI conversational interface (String)
Should you go for letting String construct an approximate model of your required workflow, describe clearly what to construct, with a immediate that will look one thing like:
I would prefer to orchestrate a easy toy RAG utility that makes use of an Airtable desk containing textual content, and OpenAI’s GPT fashions (I’ve a paid API key).
Pipedream’s language mannequin might return a solution so that you can approve (at your individual threat), like this:
I will create a easy RAG (Retrieval-Augmented Era) agent that connects your Airtable desk with OpenAI’s GPT fashions. Here is what I will construct:
Arrange a set off to just accept queries (by way of webhook or handbook execution)
Search and retrieve related textual content out of your Airtable desk primarily based on the question
Use the retrieved textual content as context for OpenAI GPT to generate an knowledgeable response
Return the AI-generated reply that is grounded in your Airtable knowledge
This gives you a primary RAG system the place you may ask questions and get solutions primarily based on the content material saved in your Airtable, enhanced by GPT’s reasoning capabilities.
For the sake of experimentation and “seeing what occurs,” approve the prompt plan of action to acquire a three-block workflow just like the one within the picture proven above.
Since this isn’t the core Pipedream editor UI, swap to it by clicking “Open Pipedream” within the higher proper nook. A brand new tab will open with the precise Pipedream workflow editor.
For the set off block, a URL is mechanically generated with a syntax just like this one I obtained for mine: https://eoupscprutt37xx.m.pipedream.internet. Click on it and, within the settings pane that opens on the right-hand aspect, guarantee the primary couple of choices are set to “Full HTTP request” and “Return a static response.”
For the second block (Airtable motion) there could also be slightly work to do. First, connect with your Airtable base. Should you’re working in the identical browser, this is likely to be simple: register to Airtable from the pop-up window that seems after clicking “Join new account,” then observe the on-screen steps to specify the bottom and desk to entry:
Pipedream workflow editor: connecting to Airtable
Right here comes the tough half (and a purpose I deliberately left an imperfect immediate earlier when asking the AI agent to construct the skeleton workflow): there are a number of kinds of Airtable actions to select from, and the precise one we’d like for a RAG-style retrieval mechanism is “Record information.” Chances are high, this isn’t the motion you see within the second block of your workflow. If that’s the case, take away it, add a brand new block within the center, choose “Airtable,” and select “Record information.” Then reconnect to your desk and check the connection to make sure it really works.
That is what a efficiently examined connection appears like:
Pipedream workflow editor: testing connection to Airtable
Final, arrange and configure OpenAI entry to GPT. Maintain your API key helpful. In case your third block’s secondary label isn’t “Generate RAG response,” take away the block and exchange it with a brand new OpenAI block with this subtype.
Begin by establishing an OpenAI connection utilizing your API key:
Establishing OpenAI connection
The person query area must be set as {{ steps.set off.occasion.physique.check }}, and the information base information (your textual content “paperwork” for RAG from Airtable) have to be set as {{ steps.list_records.$return_value }}.
You possibly can maintain the remainder as default and check, however chances are you’ll encounter parsing errors widespread to those sorts of workflows, prompting you to leap again to String for assist and computerized fixes utilizing the AI agent. Alternatively, you may instantly copy and paste the next into the OpenAI element’s code area on the backside for a sturdy resolution:
import openai from "@pipedream/openai"
export default defineComponent({
title: "Generate RAG Response",
description: "Generate a response utilizing OpenAI primarily based on person query and Airtable information base content material",
kind: "motion",
props: {
openai,
mannequin: {
propDefinition: [
openai,
"chatCompletionModelId",
],
},
query: {
kind: "string",
label: "Person Query",
description: "The query from the webhook set off",
default: "{{ steps.set off.occasion.physique.check }}",
},
knowledgeBaseRecords: {
kind: "any",
label: "Information Base Information",
description: "The Airtable information containing the information base content material",
default: "{{ steps.list_records.$return_value }}",
},
},
async run({ $ }) {
// Extract person query
const userQuestion = this.query;
if (!userQuestion) {
throw new Error("No query supplied from the set off");
}
// Course of Airtable information to extract content material
const information = this.knowledgeBaseRecords;
let knowledgeBaseContent = "";
if (information && Array.isArray(information)) {
knowledgeBaseContent = information
.map(report => {
// Extract content material from fields.Content material
const content material = report.fields?.Content material;
return content material ? content material.trim() : "";
})
.filter(content material => content material.size > 0) // Take away empty content material
.be a part of("nn---nn"); // Separate totally different information base entries
}
if (!knowledgeBaseContent) {
throw new Error("No content material present in information base information");
}
// Create system immediate with information base context
const systemPrompt = `You're a useful assistant that solutions questions primarily based on the supplied information base. Use solely the data from the information base beneath to reply questions. If the data just isn't out there within the information base, please say so.
Information Base:
${knowledgeBaseContent}
Directions:
- Reply primarily based solely on the supplied information base content material
- Be correct and concise
- If the reply just isn't within the information base, clearly state that the data just isn't out there
- Cite related elements of the information base when doable`;
// Put together messages for OpenAI
const messages = [
{
role: "system",
content: systemPrompt,
},
{
role: "user",
content: userQuestion,
},
];
// Name OpenAI chat completion
const response = await this.openai.createChatCompletion({
$,
knowledge: {
mannequin: this.mannequin,
messages: messages,
temperature: 0.7,
max_tokens: 1000,
},
});
const generatedResponse = response.generated_message?.content material;
if (!generatedResponse) {
throw new Error("Did not generate response from OpenAI");
}
// Export abstract for person suggestions
$.export("$abstract", `Generated RAG response for query: "${userQuestion.substring(0, 50)}${userQuestion.size > 50 ? '...' : ''}"`);
// Return the generated response
return {
query: userQuestion,
response: generatedResponse,
model_used: this.mannequin,
knowledge_base_entries: information ? information.size : 0,
full_openai_response: response,
};
},
})
If no errors or warnings seem, try to be prepared to check and deploy. Deploy first, after which check by passing a person question like this within the newly opened deployment tab:
Testing deployed workflow with a immediate asking what’s the capital of Japan
If the request is dealt with and all the things runs appropriately, scroll right down to see the response returned by the GPT mannequin accessed within the final stage of the workflow:
GPT mannequin response
Effectively carried out! This response is grounded within the information base we in-built Airtable, so we now have a easy prototype RAG system that mixes Airtable and GPT fashions by way of Pipedream.
# Wrapping Up
This text confirmed construct, with little or no coding, an orchestration workflow to prototype a RAG system that makes use of Airtable textual content databases because the information base for retrieval and OpenAI’s GPT fashions for response era. Pipedream permits defining orchestration workflows programmatically, manually, or aided by its conversational AI agent. Via the writer’s experiences, we succinctly showcased the professionals and cons of every method.
Iván Palomares Carrascosa is a pacesetter, author, speaker, and adviser in AI, machine studying, deep studying & LLMs. He trains and guides others in harnessing AI in the actual world.
