Skip to main content

How to efficiently build AI chat applications for your own documents with MongoDB Atlas

· 4 min read
Duncan Blythe

Despite the huge surge in popularity in building AI applications with LLMs and vector search, we haven't seen any walkthroughs boil this down to a super-simple, few-command process. With SuperDuperDB together with MongoDB Atlas, it's easier and more flexible than ever before.

info

We have built and deployed an AI chatbot for questioning technical documentation to showcase how efficiently and flexibly you can build end-to-end Gen-AI applications on top of MongoDB with SuperDuperDB: https://www.question-the-docs.superduperdb.com/

Implementing a (RAG) chat application like a question-your-documents service can be a tedious and complex process. There are several steps involved in doing this:

  • Serve a model or forward requests to convert text-data in the database to vectors in a vector-database
  • Setting up a vector-index in a vector-database which efficiently finds similar vectors
  • Setting up an endpoint to either run a self hosted LLM or forward requests to a question-answering LLM such as OpenAI
  • Setting up an endpoint to:
    • Convert a question to a vector
    • Find relevant documents to the question using vector-search
    • Send the documents as context to the question-answering LLM

This process can be tedious and complex, involving several pieces of infrastructure, especially if developers would like to use other models than those hosted behind OpenAI's API.

What if we told you that with SuperDuperDB together with MongoDB Atlas, these challenges are a thing of the past, and can be done more simply than with any other solution available?

Let's dive straight into the solution:

Connect to MongoDB Atlas with SuperDuperDB

from superduperdb.db.base.build import build_datalayer
from superduperdb import CFG
import os

ATLAS_URI = "mongodb+srv://<user>@<atlas-server>/<database_name>"
OPENAI_API_KEY = "<your-open-ai-api-key>"

os.environ["OPENAI_API_KEY"] = OPENAI_API_KEY

CFG.data_backend = ATLAS_URI
CFG.vector_search = ATLAS_URI

db = build_datalayer()

After connecting to SuperDuperDB, setting up question-your-documents in Python boils down to 2 commands.

Set up a vector-index

from superduperdb.container.vector_index import VectorIndex
from superduperdb.container.listener import Listener
from superduperdb.ext.openai.model import OpenAIEmbedding

collection = Collection('documents')

db.add(
    VectorIndex(
        identifier='my-index',
        indexing_listener=Listener(
            model=OpenAIEmbedding(model='text-embedding-ada-002'),
            key='txt',
            select=collection.find(),
        ),
    )
)

In this code snippet, the model used for creating vectors is OpenAIEmbedding. This is completely configurable. You can also use:

  • CohereAI API
  • Hugging-Face transformers
  • sentence-transformers
  • Self built models in torch

The Listener component sets up this model to listen for new data, and compute new vectors as this data comes in.

The VectorIndex connects user queries with the computed vectors and the model.

By adding this nested component to db, the components are activated and ready to go for vector-search.

Add a question-answering component

from superduperdb.ext.openai.model import OpenAIChatCompletion

chat = OpenAIChatCompletion(
    model='gpt-3.5-turbo',
    prompt=(
        'Use the following content to answer this question\n'
        'Do not use any other information you might have learned\n'
        'Only base your answer on the content provided\n'
        '{context}\n\n'
        'Here\'s the question:\n'
    ),
)

db.add(chat)

This command creates and configures an LLM hosted on OpenAI to operate together with MongoDB. The prompt can be configured to ingest the context using the {context} format variable. The results of the vector search are pasted into this format variable.

Question your documents!

input = 'Explain to me the reasons for the change of strategy in the company this year.'

response, context = db.predict(
    'gpt-3.5-turbo',
    input=input,
    context=collection.like({'txt': input}, vector_index='my-index').find()
)

This command executes the vector-search query in the context parameter. The results of this search are added to the prompt to prime the LLM to ground its answer on the documents in MongoDB.

Contributors are welcome!​

SuperDuperDB is open-source and permissively licensed under the Apache 2.0 license. We would like to encourage developers interested in open-source development to contribute in our discussion forums, issue boards and by making their own pull requests. We'll see you on GitHub!

Become a Design Partner!​

We are looking for visionary organizations which we can help to identify and implement transformative AI applications for their business and products. We're offering this absolutely for free. If you would like to learn more about this opportunity please reach out to us via email: partnerships@superduperdb.com