sales insight snapshot In this tutorial, you will learn how to use Nile’s tenant virtualization and vector embedding to build a SaaS application that helps browse and query new code-bases. The web application we’ll build is designed to be end-to-end serverless RAG architecture. We’ll use Modal to deploy the web application and also deploy a Llama 3.1 7B model. We’ll use Nile’s serverless Postgres to store the embeddings and the application data and retrieve the relevant parts of the sales transcripts. Embeddings for the code in each repository are generated with OpenAI’s text-embedding-3-small model and stored in Nile with the relevant chunks of the transcripts. The response to each question is generated by querying the code snippets table using pg_vector extension, and then sending the relevant documents to the Llama 3.1 7B model, deployed on A100 GPU from Modal. The response is streamed back to the user in real-time. Because of Nile’s virtual tenant databases, the retrieved transcripts will only be from the tenant the user selected. Nile validates that the user has permissions to view this tenant’s data. No risk of accidentally retrieving code that belongs to the wrong tenant.

​
1. Create a database

  1. Sign up for a Nile account if you don’t have one already
  2. You should see a welcome message. Click on “Lets get started”
  3. Give your workspace and database names, or you can accept the default auto-generated names.

​
2. Create tables

After you created a database, you will land in Nile’s query editor. Create the following table for storing our embeddings: 
create table call_chunks (
   tenant_id uuid,
   conversation_id varchar(50),
   chunk_id int,
   speaker_role varchar(20),
   content text,
   embedding vector(1536) -- must match the embedding model dimensions
);
Note the embedding column is of type vector(1536). Vector type is provided by the pg_vector extension for storing embeddings. The size of the vector has to exactly match the number of dimensions in the model you use. The table has tenant_id column, which makes it tenant aware. By storing embeddings in a tenant-aware table, we can use Nile’s built-in tenant isolation to ensure that sales transcripts won’t leak between tenants. If all went well, you’ll see the new table in the panel on the left hand side of the query editor. You can also see Nile’s built-in tenant table next to it. You can also explore the schema in the “Schema Visualizer”

​
3. Getting Nile credentials

In the left-hand menu, click on “Settings” and then select “Credentials”. Generate credentials and keep them somewhere safe. These give you access to the database. In addition, you’ll need the API URL. You’ll find it under “Settings” in the “General” page.

​
4. Third party dependencies

This project uses Modal for hosting the web application and the LLM. You will need to have a Modal account, set up the Modal CLI, nd link your Nile credentials to your Modal account. Our Modal Integration doc will help you through this. We are also using Hugging Face for the Llama 3.1 7B model. You will need to have a Hugging Face account for this, and an HuggingFace API key. Set up the API key as a secret in Modal using Hugging Face integration. We are using OpenAI to generate embeddings for sales transcripts and questions. You will need to have an OpenAI account and an OpenAI API key. Store the OpenAI key as a secret in Modal. The way this example is set up, you’ll want to create a secret called “embedding-config” and store two keys in it: OPENAI_API_KEY and EMBEDDING_MODEL (e.g. text-embedding-3-small).

​
5. Setting the environment

  • If you haven’t cloned this project yet, now will be an excellent time to do so. Since it uses NextJS, we can use create-next-app for this: 
    git clone https://github.com/niledatabase/niledatabase.git
cd niledatabase/examples/sales_insight
  • Create a .env file and add your Nile connection string: 
      DATABASE_URL=postgresql://user:password@us-west-2.db.thenile.dev:5432/mydb
  • Install dependencies with pip install -r requirements.txt

​
6. Loading data

We prepared example data, that is already chunked and embedded. This will let you start faster and save on OpenAI credits. To get the prepared data, you need to:
  • Install git-lfs. You can download from their website or “brew install git-lfs”
  • git lfs install
  • Upload your SSH public key to your Hugging Face account.
  • Configure ssh to use the key you just uploaded to Hugging Face when accessing the repo. You’ll need to put this snippet in your .ssh/config file:
Host hf.co
  AddKeysToAgent yes
  UseKeychain yes
  IdentityFile ~/.ssh/myhfkey
Now, you can load the data by running: 
 python -m ingest.load_embeddings

​
7. Download the Llama 3.1 7B model to Modal

This step is important, as it will take a while to download the model and its weights, and you don’t want to do it on every deployment. 
modal run download_llama.py
Note that Llama models require approval from Meta before downloading. You can request access to the Llama 3.1 7B model by clicking on the “Request Access” button on the model page in Hugging Face. At the end of the process, you’ll have a volume with the model and its weights. We’ll use this volume in the llm container.

​
8. Building the UI

The example has a React frontend. We pre-build it with Vite and then serve the static files in the Modal app. 
cd ui
npm install
npm run build
cd ..

​
9. Running the app

You run the app in Modal. We will start by running it in dev mode, which will start a modal app and a web server. This will let you see the logs in the console, and it will automatically re-deploy if you make changes. 
modal serve web_app.py
If all went well, modal will print a URL. Go this URL in your browser to see the app. Your browser should show you the first page in the app, asking you to login or sign up. After you sign up as a user of this example app, you’ll be able to see this user by going back to Nile Console and running select * from users in the query editor. Once you choose a tenant, you select a sales transcript and can ask questions about the transcript.

​
Learn More

To learn more about how this example works and how to use Nile: