Page Comparison

- [Introduction](#introduction)
  - [Configuration Philosophy](#configuration-philosophy)
- [Prerequisites](#prerequisites)
- [Determine Configuration Values](#determine-configuration-values)
- [Configure the Azure OpenAI Models](#configure-the-azure-openai-models)

# Introduction
This guide describes how to configure Azure to produce the values required to configure the Ayfie Personal Assistant feature.

## Configuration Philosophy
As we will learn in the next section, one is to set Personal Assistant up with three models: a Main model, a High Quality model and an Embeddings model. The Embeddings model is always the same and the High Quality model is normally easy to determine; it is whichever model that, at the time, is considered to produce the most correct and accurate responses to user prompts. Deciding which model to use as the Main model can be a bit trickier. One may want to select a model with a lower cost or, if the model to be used as the High Quality model is a bit slow, then one may opt for a model with a faster response time as the Main model.

This is further complicated by the fact that various factors, such as pricing and model performance, are subject to change over time. New models or new versions of existing models are also constantly being released. This means that the conditions that led to one configuration at installation time might be very different a few months later. For an overview of released models and their regional availability, as well as links to pricing information, check out [Azure OpenAI Service Documentation](https://learn.microsoft.com/en-us/azure/ai-services/openai/overview).

The configuration recommendation provided in the following section reflects the current advantages of GPT-3.5, which include its lower cost and faster response time over the more correct and accurate GPT-4 model. We advise periodically reviewing this documentation to determine if evolving conditions warrant a reconfiguration.

# Prerequisites
These are the prerequisites that must be fulfilled before setting up Personal Assistant:
- **Obtain an Azure Subscription**
  - One must have an active Azure subscription. If one don’t have one, one can sign up for an Azure subscription on the [Azure website](https://azure.microsoft.com/en-us/). At times, Microsoft may impose subscription type-specific limitations on their OpenAI services, particularly concerning the amount of data (referred to as the token quota) that can be exchanged during chat interactions. Please see this Microsoft documentation on [token quotas](https://learn.microsoft.com/en-us/azure/ai-services/openai/quotas-limits) for details on which subscription types may have restrictions.
- **Get Azure OpenAI Approval**
  - The Azure subscription needs to be approved for [Azure Open AI](https://learn.microsoft.com/en-us/azure/ai-services/openai/overview). How to do that is described in [Ayfie Personal Assistant - How to Request Access to the Azure OpenAI Service](https://ayfie-dev.atlassian.net/wiki/spaces/SAGA/pages/3443523634/Ayfie+Personal+Assistant+-+How+to+Request+Access+to+the+Azure+OpenAI+Service).

 Once the prerequisites above have been completed, one can then start configuring Ayfie Personal Assistant.

# Determine Configuration Values
One needs to determine the following configuration values:
- **Deployment Name**
- **API Address**
- **API Key**

for each of the following 3 Azure OpenAI models:
- **Main Model**
- **High Quality Model**
- **Embeddings Model**

The intended difference between the Main Model and the High Quality Model is the quality of the chat responses when users toggle between the two in the Personal Assistant UI. The normal approach is to set the Main Model up with GPT-3.5 and the High Quality Model with GPT-4. The reason for having two options and not just a single High Quality Model option is that GPT-3.5 can, depending on the current offerings, be faster and/or lower cost than GPT-4.

The last listed model is for creating embeddings. Embeddings are numerical representations of words that are learned from large amounts of text data. Currently, there is only one supported model.

Each of the 3 models requires an API address and an API key. However, unless one chose to spread the models across geographical regions, all 3 models will be reached via the same API address and an API key.

Given the limited options for each setting, the configuration of the prerequisites in Azure described in the next section, has a very predictable outcome:
- Main Model Deployment Name: ***gpt-35-turbo***
- High Quality Model Deployment Name: ***gpt-4***
- Embeddings Model Deployment Name: ***text-embedding-ada-002***
- API Address: *the same one for all three*
- API Key: *the same one for all three*

# Configure the Azure OpenAI Models
For more information, consult [Azure OpenAI Service Documentation](https://learn.microsoft.com/en-us/azure/ai-services/openai/overview).

Follow these steps to set up the deployments for the 3 Azure OpenAI models:
- Log in to the Azure portal at [portal.azure.com](https://portal.azure.com/)
- Make sure the account that is logged in has at least one subscription
- Go to *Azure OpenAI*
  - Click *Create*, to create a Resource
    - In *Project Details*, select *Subscription*
    - In *Project Details*, select *Resource Group*
    - In *Instance Details*, select *Region*.  Not all models are available in all regions, consult with [Azure OpenAI Service models](https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/models) for availability. EU country customers should for legal reason select a region that is within EU. To avoid any data quota conflict with a current or a future deployment of the standalone version of the Personal Assistant, it is recommended to **not** select the regions *Sweden Central*, *UK South* and *Canada East*.
    - In *Instance Details*, set the *Name*
    - In *Instance Details*, select *Pricing Tier*
    - Click *Next* to go to the *Network* tab
    - In *Type*, select *All networks, including internet can access this resource.*
    - Click *Next* to go to the *Tags* tab
    - Click *Next* to go to the *Review + submit* tab
    - Click *Create*
  - When Resource is created, select the resource in *Azure OpenAI*
  - Click *Keys and Endpoint* in the left menu
    - **Copy the value of *Endpoint*, it will be required later as the API Address**
    - **Copy the value of *KEY 1*, it will be required later as the API Key** (optionally *KEY 2*, both keys are valid)
  - Click *Model deployments* in the left menu
  - Click *Manage Deployments* (this will open a new portal)
  - Click *Create new deployment* to create Main model
    - Select the model (recommended *gpt-35-turbo*)
    - Select the Model Version (Set to *Auto-update to default*)
    - Set the *Deployment Name* (must be same as model name)
      - **Copy the *Deployment Name*, it will be required later**
    - In *Advanced Options*, set *Tokens per Minute Rate Limit (thousands)* to maximum value.
  - Click *Create new deployment* to create High Quality model
    - Select the model (recommended *gpt-4*)
    - Select the Model Version (recommended *1106-Preview*)
    - Set the *Deployment Name* (must be same as model name).
      - **Copy the *Deployment Name*, it will be required later**
    - In *Advanced Options*, set *Tokens per Minute Rate Limit (thousands)* to maximum value.
  - Click *Create new deployment* to create Embeddings model
    - Select the model (must be *text-embedding-ada-002*)
    - Set the *Deployment Name* (must be same as model name)
      - **Copy the *Deployment Name*, it will be required later**
    - In *Advanced Options*, set *Tokens per Minute Rate Limit (thousands)* to maximum value