Retrieve Context Information

If you're using the MongoDB Chatbot Server to perform RAG, you must retrieve context information to include in your answer. The primary way of doing this is with semantic search.

You can add the information that you retrieve using the MongoDB RAG Ingest CLI. Information retrieval is the single point of contact between the MongoDB RAG Ingest CLI and the MongoDB Chatbot Server.

The FindContentFunc Function

To perform semantic search, you must implement a FindContentFunc function. To see the default implementation using Atlas Vector Search, refer to the following Find Content with Atlas Vector Search section.

Pass the FindContentFunc to the MakeRagGenerateUserPromptParams.findContent property.

import { makeRagGenerateUserPrompt } from "mongodb-chatbot-server";
import { someFindContentFunc } from "./someFindContentFunc"; // example

const ragGenerateUserPrompt = makeRagGenerateUserPrompt({
  findContent: someFindContentFunc,
  // ...other config
});

Find Content with Atlas Vector Search

To use the MongoDB Chatbot Server with Atlas Vector Search for semantic search, you can use the makeDefaultFindContentFunc().

This function retrieves data from an EmbeddedContentStore. To learn more about how to add data to an EmbeddedContentStore, refer to the Ingest CLI documentation.

Pass a MakeDefaultFindContentFuncArgs object to the makeDefaultFindContentFunc() function.

import { makeRagGenerateUserPrompt } from "mongodb-chatbot-server";

// Create function that creates vector embeddings
// for user query.
const embedder = makeOpenAiEmbedder({
  openAiClient,
  deployment: OPENAI_EMBEDDING_DEPLOYMENT,
  backoffOptions: {
    numOfAttempts: 3,
    maxDelay: 5000,
  },
});

// Data store that is used to store the vector embeddings.
// Used to look up matching content.
const embeddedContentStore = makeMongoDbEmbeddedContentStore({
  connectionUri: MONGODB_CONNECTION_URI,
  databaseName: MONGODB_DATABASE_NAME,
});

const args: MakeDefaultFindContentFuncArgs = {
  embedder,
  store: embeddedContentStore,
  findNearestNeighborsOptions: {
    k: 5,
    path: "embedding",
    indexName: VECTOR_SEARCH_INDEX_NAME,
    // Note: you may want to adjust the minScore depending
    // on the embedding model you use. We've found 0.9 works well
    // for OpenAI's text-embedding-ada-02 model for most use cases,
    // but you may want to adjust this value if you're using a different model.
    minScore: 0.9,
  },
};
const findContent = makeDefaultFindContentFunc(args);

const ragGenerateUserPrompt = makeRagGenerateUserPrompt({
  findContent,
  // ...other config
});

Boost Results

You can modify the results returned by the default find content function with SearchBooster objects. SearchBoosters mutate the results returned by the default find content function.

You could use a SearchBooster to do things like:

• Always results from a specific data source
• Ensure a data source isn't over-represented in the results

To use one or more SearchBoosters, pass them to the MakeDefaultFindContentFuncArgs.searchBoosters property. The searchBoosters property is an array of SearchBooster objects, which are applied in the order of the array.

The following is an example of using a SearchBooster to ensure that results from a specific data source are always returned.

/**
 * Ensure that results from data source 'foo' are always returned
 * if query contains 'foo'.
 */
const boostFoo: SearchBooster = {
  async shouldBoostFunc({ text }: { text: string }) {
    return text.includes("foo");
  },
  async boost({
    embedding,
    store,
    existingResults,
  }: {
    embedding: number[];
    store: EmbeddedContentStore;
    existingResults: WithScore<EmbeddedContent>[];
  }) {
    const boostedResults = await store.findNearestNeighbors(
      embedding,
      {
        k: 2,
        path: "embedding",
        indexName: VECTOR_SEARCH_INDEX_NAME,
        minScore: 0, // no min score for 'foo'
      },
      {
        filter: {
          dataSource: "foo",
        },
      }
    );
    const fewerExistingResults = existingResults.slice(0, 3);
    // No duplicates
    const newResults = fewerExistingResults.filter((result) =>
      boostedResults.every((manualResult) => manualResult.text !== result.text)
    );
    //
    return [...boostedResults, ...fewerExistingResults].sort(
      (a, b) => b.score - a.score
    );
  },
};

const args: MakeDefaultFindContentFuncArgs = {
  // ...other args
  searchBoosters: [boostFoo],
};

Include Search Booster Filters In Your Atlas Vector Search Index

If you are using an Atlas Vector Search filter in a booster, you must include the filter in your index definition. For more information on Atlas Vector Search filters, refer to Atlas Vector Search filter index definition in the MongoDB Atlas documentation.

For example, you might have a booster that finds data from a specific data source by including a filter on a field named sourceName during vector search. For the search to run, you must include the dataSource field and any other filtered fields in your vector search index definition:

{
  "fields": [
    {
      "type": "vector"
      // ...
    },
    {
      "type": "filter",
      "path": "sourceName"
    }
    // ...
  ]
}

Then you can include the filter in the $search query in your booster.