Pre-Process User Queries

One of the most powerful and flexible features that you can add to your chatbot is a pre-processing step for user queries. With a query preprocessor, you can mutate a query before it's used to retrieve relevant content. This makes the query more conversationally relevant, and therefore more likely to get better results.

You can also use a preprocessor to reject queries that are inappropriate or otherwise should not be sent to the chatbot.

Query preprocessors run on the POST /conversations/:conversationId/messages endpoint.

Adding a preprocessor is optional. If you don't add one, the user query is sent directly to find relevant content and the query will never be rejected.

Generally, preprocessors use LLMs to perform the transformation.

For example, say a user sends the following message to a chatbot that answers questions about the MongoDB documentation:

User: add data node

To make this query more relevant, you can add a preprocessor that transforms the query to something like:

User: How do I add a data to MongoDB using the Node.js driver?

This transformed query is more likely to return relevant results, which are then used to generate a better answer.

The Preprocessor Function

Preprocessors have the QueryPreprocessorFunc signature.

To run a preprocessor when adding user messages, include a QueryPreprocessorFunc in the MakeRagGenerateUserPromptParams.queryPreprocessor property.

Mutate Queries

You can use a preprocessor to mutate the user query to make it more relevant.

If you don't want to mutate the query, do not return a query property from the result or keep it the same as the original query.

Here's an example of preprocessor that transforms the user query to a more semantically relevant question:

import {
  QueryPreprocessorFunc,
  makeRagGenerateUserPrompt,
} from "mongodb-chatbot-server";
import { OpenAIClient, AzureKeyCredential } from "@azure/openai";

const openAiClient = new OpenAIClient(
  OPENAI_ENDPOINT,
  new AzureKeyCredential(OPENAI_API_KEY)
);
const querySystemPrompt: SystemPrompt = {
  role: "system",
  content: `Transform the query into a relevant question about MongoDB products or company.
Example 1:
User: add data node
Assistant: How do I add a data to MongoDB using the Node.js driver?
Example 2:
User: get started atlas
Assistant: How do I get started with MongoDB Atlas?`,
};

const queryToQuestionPreprocessor: QueryPreprocessorFunc = async function ({
  query,
  messages,
}) {
  if (query === undefined) {
    return { query, rejectQuery: false };
  }
  const {
    choices: [choice],
  } = await openAiClient.getChatCompletions(OPENAI_DEPLOYMENT, [
    querySystemPrompt,
    { role: "user", content: query },
  ]);
  const { message } = choice;
  return {
    query: message,
    rejectQuery: false,
  };
};

const ragGenerateUserPrompt = makeRagGenerateUserPrompt({
  queryPreprocessor: queryToQuestionPreprocessor,
  // ...other config
});

Reject Queries

You can also add preprocessor logic to reject user queries that you do not want to send to the chatbot. To reject a query, return { rejectQuery: true } from the QueryPreprocessorFunc.

For example, say you want to reject queries that contain profanity. You can use a preprocessor to check if the query contains profanity, and reject the query if it does.

import { makeRagGenerateUserPrompt } from "mongodb-chatbot-server";
import { profanityCheck } from "./profanityCheck"; // checks if query contains profanity

async function rejectProfanePreprocessor({ query, messages }) {
  if (query === undefined) {
    return { query, rejectQuery: false };
  }
  const isProfane = await profanityCheck(query);
  return {
    rejectQuery: isProfane,
  };
}

const ragGenerateUserPrompt = makeRagGenerateUserPrompt({
  queryPreprocessor: rejectProfanePreprocessor,
  // ...other config
});