To use the results from SID in your app, most adopt the following flow:

  1. Get input from the user. Usually, this is from a chat interface like ChatGPT.
  2. Forward that input to our query endpoint to get relevant context.
  3. Feed the results and the original user request to a language model like ChatGPT.

Using Query Results as Context

Once you have received the results from the query endpoint you will want to feed them as context to a language model like ChatGPT. We’ve collected a few tips on how to do this well.

The most important thing is the prompt. You need to inform the model that it can now access the personal files a user has connected, and can use that information to answer the request. We suggest adding a section with Instructions and a section with Context to your prompt.

In general, if you observe behavior that the model should not exhibit, you can add a section to your prompt that instructs the model to not do that.

import openai

const userInput = "Write a paragraph describing how SID can help developers."

const accessToken = "" // retrieve the access token from wherever you store it, usually a database
const queryResults = await"", {query: query, limit: 5}, {headers: {Authorization: `Bearer ${accessToken}`}})

const systemMessage = "You are an expert writer and editor. You hold yourself to high journalistic standards and never invent or misrepresent information. You are provided with a writing task.\n\n" +
    Instructions: You complete the writing task using the context provided below. Do not say that you do not know or need more information. Be concise and specific. Never repeat yourself. Refrain from using vacuous phrases that do not convey concrete information."

const completion = await{
    messages: [{ role: "system", content: systemMessage }, { role: "user", content: "Writing Task: " + userInput + "\n\nContext:\n" + queryResults.results }],
    model: "gpt-3.5-turbo",
    max_tokens: 2048,

Formatting Query Results

If you’re inserting the raw JSON results from the query endpoint, the model will be inclined to writing code instead of text. To circumvent this, we recommend formatting the results as plain text or markdown. Here is a helper function that may be useful.

function formatResultsToMarkdown(results: Result[]): string {
    if (results.length === 0) {
        return "- no additional context provided\n";

    return => {
        const scoreAsPercentage = (result.score * 100).toFixed(2);
        return `- **${result.title} (Confidence: ${scoreAsPercentage}%):** "${result.text}"`;

The query endpoint may return an empty results array if there are no relevant results. Be sure to tell the model that no context was found.