ChatGPT is an excellent general-purpose example of how we can use AI to answer casual questions, but it could do better when the questions require domain-specific knowledge. Thanks to this ChatGPT starter kit, you can train the model on websites you define.

header image was generated using midjourney

What is gannonh/gpt3.5-turbo-pgvector?

This starter app was put together by @gannonh and makes great use of the Supabase pgvectors and OpenAI Embedding feature. The app leverages Next.js to stand up a simple prompt interface.

Live demo: https://astro-labs.app/docs

How does it work?

This starter app uses embeddings to generate a vector representation of a document and then uses vector search to find the most similar documents to the query. The results of the vector search are then used to construct a prompt for GPT-3, which is then used to generate a response. The response is then streamed to the user.

Web pages are scraped, stripped to plain text, and split into 1000-character documents.

// Stripe text from HTML
// pages/api/generate-embeddings.ts

async function getDocuments(urls: string[]) {
  const documents = [];
  for (const url of urls) {
    const response = await fetch(URL);
    const html = await response.text();
    const $ = cheerio.load(html);
    // tag based e.g. <main>
    const articleText = $("body").text();
    // class based e.g. <div class="docs-content">
    // const articleText = $(".docs-content").text();

    let start = 0;
    while (start < articleText.length) {
      const end = start + docSize;
      const chunk = articleText.slice(start, end);
      documents.push({ url, body: chunk });
      start = end;
    }
  }
  return documents;
}

Once the URLs are stripped down to the text, they are sent to the Supabase after some embedding creation using the text-embedding-ada-002 model.

The OpenAI docs recommend using text-embedding-ada-002 for nearly all use cases. Fun fact, this is the same embedding Notion's AI tool uses under the hood. It's better, cheaper, and simpler to use.

text-embedding-ada-002 announcement

// Create embeddings from URLs 
// pages/api/generate-embeddings.ts

const documents = await getDocuments(urls);

for (const {
    url,
    body
  }
  of documents) {
  const input = body.replace(/\n/g, " "); 

  console.log("\nDocument length: \n", body.length);
  console.log("\nURL: \n", url);

  const embeddingResponse = await openAi.createEmbedding({
    model: "text-embedding-ada-002",
    input
  });

  console.log("\nembeddingResponse: \n", embeddingResponse);

  const [{
    embedding
  }] = embeddingResponse.data.data;

  // In production we should handle possible errors
  await supabaseClient.from("documents").insert({
    content: input,
    embedding,
    URL
  });
}

gpt3.5-turbo-pgvector is an excellent starter for folks looking to try out OpenAI on their own data or sites. I see this being extremely useful in the documentation and now understand why OpenAI doesn't need to search in their docs (this is a joke, they should add search). Search in docs could be replaced by projects setting up their own embeddings.

Share in the comments if you have a use case for this.

Also, if you have a project leveraging OpenAI or similar, leave a link in the comments. I'd love to take a look and include it in my 9 days of OpenAI series.

Find more AI projects using OpenSauced

Stay saucy.