The Synergy Between LangChain and Azure OpenAI

Hello again from #XenAIBlog! Before we dig deeper into the intricacies of language models, we believe it's important to establish some foundational knowledge. And so, this week, our main focus will be on leveraging LangChain to create custom LLM applications, and specifically integrating Azure OpenAI with it. Are you ready? Let's begin!

If you're a ChatGPT user, you've likely marveled at its impressive text generation, question answering, and conversation capabilities. You may have wondered if it can be integrated into local systems or use it to create new applications. The answer is “Yes, you can!”. But, as with any technology, there are certain limitations to consider.

Firstly, data isolation - OpenAI models operate independently and cannot tap into our internal data. This restriction prevents them from retrieving data from company databases or delivering real-time updates on internal affairs. Secondly, using the OpenAI API can lead to costs, as it is billed based on token usage. Finally, ChatGPT's knowledge is limited to information up to September 2021, so if you're seeking the most current updates from sources like Google or Wikipedia, it may not have the latest information.

So, how do we surmount these challenges and elevate the ChatGPT experience? Allow us to introduce you to LangChain, a robust framework meticulously crafted for constructing applications driven by language models. Its core philosophy is that the most powerful and distinctive applications should not merely access language models via an API but also possess data-awareness. This means we can seamlessly merge our language model with our organization's internal databases, thus opening up fresh avenues for data-driven applications. Furthermore, LangChain empowers the language model to be agentic, enabling it to interact with its surroundings and take actions based on its responses.

We'll now demonstrate the development of a basic yet effective scenario where we harnessed LangChain and GPT-3.5 to craft a healthcare advisory tool directly within our in-house system. To achieve this, we combined the capabilities of LangChain with Azure's OpenAI API, focusing on the GPT-3.5 Turbo model.

Why Azure OpenAI, you ask? Unlike OpenAI, Azure OpenAI retains data exclusively within Microsoft Azure and enforces automatic encryption for training data and models, guaranteeing strict adherence to organizational security and compliance standards. What sets Azure OpenAI apart is its seamless integration into the Azure ecosystem, simplifying deployment, scalability, and AI model management. Moreover, it provides essential features like private networking, regional availability, and built-in Responsible AI Content Filtering, ensuring a comprehensive and secure AI solution.

Let's get our hands dirty now, shall we? We used a Python version = 3.8.1, Azure OpenAI key, Azure OpenAI endpoint, and Azure OpenAI's model deployment name. A detailed guide for creating and deploying an Azure OpenAI Service resource can be accessed here. You can also find instructions to obtain the key and endpoint in the same link.

Pro tip: Do make sure to choose a proper deployment name for your model as it will be used in your code to call the model in the deployment_id parameter.

Just like a determined adventurer equipping themselves for a journey, we first ensured that we had the necessary tools. Here's a straightforward code snippet that we used for installing the LangChain Python package:

To securely store our API key, we created an environment variable by following the provided code below. As we worked with Azure's OpenAI API key, it was essential to specify the type and version of the API, apart from the key and endpoint, and save them as environment variables, like this:

Next, we were able to seamlessly incorporate the gpt-3.5 model into our code:

In the code above, our initial action involved importing the LangChain OpenAI module. Following this, we instantiated an OpenAI object, configuring specific parameters, such as setting the temperature to 0.5. To tailor the code to our needs, we replaced the "deployment_id" with the deployment name corresponding to the Azure OpenAI model we deployed. We then utilized the llm.predict() method, providing a prompt that transformed the model into our personal health advisor. The prompt was as follows:

"""You are my personal health advisor. Please give me some tips on how to lead a healthy lifestyle."""

This elicited a response containing several essential tips. It's worth noting that fine-tuning the prompt can cater to specific requirements in your specific use-case. This is truly quite marvelous!

The fusion of a powerful LLM in Azure OpenAI, coupled with LangChain, has yielded an astonishing result—a healthcare advisor readily available on our local system. As we look ahead, LangChain promises to be a groundbreaking tool that empowers us to fully harness language models, creating applications that are both data-aware and highly responsive, marking the beginning of an exciting journey into the future of AI-driven applications.

Stay tuned for our upcoming article next week, where we'll decode the 'chain' in LangChain. Until then, take care!