Creating GPT-Driven Applications Using LangChain

Generated using Dall-E 2 [by author]

Introduction

Large Language Models (LLMs) like OpenAI ChatGPT are called foundational models because even though they are trained for a relatively small set of tasks, they work exceptionally well for multiple unseen downstream tasks. While there is still some debate on how they are so good, at a high level it is quite easy to under what they do — they just predict the next word (read tokens). And all the cool tools you see built using these models, are nothing but the smart application of this feature.

All said and done, they are also quite infamous for a lot of things — sometimes the generations don’t make sense, the models hallucinate (generating the same things over and over again), generations are not factually correct, they can’t do maths well, and more. While one research paradigm believes training bigger models with more data can fix some of these problems, another approach is to improve the existing model by connecting them with external apps. The idea is quite simple, use LLMs for what they are good at, and for the rest use external experts (read Apps or Scripts). This blog will shed some light on the second paradigm and how you can connect any LLM with your apps!

Power of Prompt Engineering

The hero of the second paradigm discussed above is “prompt”, so let’s talk about prompt engineering. Essentially, prompt engineering involves crafting effective prompts or input sequences that help guide the model’s generation toward desired outputs. This can involve using various techniques such as conditioning the model on specific input formats or injecting additional information into the prompt to improve the model’s performance on specific tasks. Effective prompt engineering can be key to achieving high accuracy and performance on a wide range of natural language processing (NLP) tasks. Let’s look into some examples,

• To begin with, if you want OpenAI’s text-davinci-003 model to write a poem, your prompt might look something like Write a poem on {topic}, where you can replace {topic} variable with the desired topic. Btw this is called zero-shot prompting as only the intention was mentioned and no examples.
• A little complex example where you want the model to write a SQL query for your DB…