Welcome back to this ongoing series of posts about using Large Language Models(LLMs) with Node.js. In the previous post we took a look at creating and using a LLM chat bot with Node.js. We showed it in the context of a fictitious claims management for an insurance company.
This next post will take a look at adding a feature that will help generate an email summarization as well as making sure the data returned from our LLM is structured as JSON.
Summarization Prompting
As a small refresher, the example application was a mock claims management system that an insurance company might use to service their claims. The previous post added a chatbot where the claims adjuster could ask it questions about the information in the claims summaries.
This post will add a new feature that will direct our LLM to write an email response that they could then send to the insured regarding their claim.
Instructions for running this part of the example is in its own feature branch here
The UI has a new side tab called Email Generate and is just a TextArea to input the original customer correspondence.
When that is submitted, a new prompt that we will take a look at in a second, will send that to our LLM to produce a JSON response that we can send back to the front-end. The front-end expects that our JSON object has both a subject and message property. The final response on the UI could look something like this:
The prompt that we use to generate the response is fairly simple, so let’s take a look at what that:
const prompt = ChatPromptTemplate.fromTemplate(`
You are a helpful, respectful and honest assistant named Parasol Assistant.
You work for Parasol Insurance
Formatting Instructions: {format_instructions}
User Input: {input}
`);The prompt is pretty simple. In this example we are using the basic fromTemplate method(Last post we used a little more advanced method). The first two lines are also very simple, instructing the LLM that it should think of itself as someone who works for our company.
The next line we will cover in the next section, but this where we will tell our LLM that we want a more structured response, like JSON.
The final line is a placeholder for the input coming from the form on the UI
Formatting Instructions
In the chatbot example, you might remember that our response was accessed through the content parameter of an object that langchain.js creates. While this might be ok to use, we can actually instruct the LLM to respond in a more structured way.
Langchain.js provides a few different output parsers for developers to use. This is nice, because maybe in your own application the AI portion might need to pass its response off to another service, like a REST endpoint and having the LLM format the response in that services format can reduce code.
The object format that our front-end expects is a JSON object with the two properties, subject and message. Something like this:
{
subject:'',
message: ''
}For this, we use the langchain.js class StructedOutputParser. It contains two methods for defining our object schema, and for this example we are using the one that allows our definition to be more advanced. This is the fromZodSchema
Just as a quick aside, Zod, while being a Superman villain, is also a popular TypeScript-first schema declaration and validation library.
Below is the code that is used to define our objects schema:
// Setup the output parser using the Zod Schema
const outputParser = StructuredOutputParser.fromZodSchema(
z.object({
subject: z.string().describe('Subject of your response, suitable to use as an email subject line.'),
message: z.string().describe('Response text that summarizes the information they gave, and asks for any other missing information needed from Parasol.')
})
);Here we have our two properties and we describe what type they should be, in this case strings, and we also describe what they are. For example, we define our message property saying that it should be some sort of response text that summarizes the information they gave, and asks for any other missing information needed from Parasol. I’d like to call out the ask for any other missing information line. In our screenshot earlier showing the result, our response was asking for a number of items, so that part worked!
Piecing it Together
It is now time to piece everything together. Similar to our chatbot, we create a chain and the only difference this time is that we add another pipe method to the end that includes the output parser that was just created:
const chain = prompt.pipe(model).pipe(outputParser);Finally we can invoke this chain to get the response. In the prompt we left placeholders for both the user's input as well as the formatting instructions, which needs to be passed to the chain. This looks something like this:
return await chain.invoke({
input: userResponse,
format_instructions: outputParser.getFormatInstructions()
});
In this example, the invoke method is used to get the response instead of the streaming method that was used in the chatbot example in the previous post. Once this is invoked, the LLM should provide a result similar to this console.log that I did:
{
subject: 'Claim Regarding Vehicle Accident - Dominic Toretto',
message: 'Dear Mr. Toretto,\n' +
'\n' +
'Thank you for reaching out to Parasol Insurance regarding your recent vehicle accident. We appreciate your detailed account of the circumstances surrounding the incident.\n' +
'\n' +
'To expedite your claim process, please provide us with the following information:\n' +
'\n' +
'1. Date and time of the accident\n' +
'2. Police report number (if any)\n' +
'3. Contact information for witnesses (if any)\n' +
'4. Photos or videos of the damaged vehicle\n' +
'5. Your policy number and any other relevant policy details\n' +
'\n' +
'Once we have this information, we can proceed with processing your claim promptly. If you have any questions or concerns, please do not hesitate to contact us.\n' +
'\n' +
'Looking forward to working with you.\n' +
'\n' +
'Best regards,\n' +
'Parasol Insurance Team'
}
This is then passed back to the UI and properly formatted on screen.
Conclusion
While this post was a bit shorter than the first post, it does show how simple returning structured output from the LLM can be.
The next post in the series will take a look at how to give your LLM more little more context when answering questions using the Retrieval Augmented Generation (RAG) concept.
As always if you want to learn more about what the Red Hat Node.js team is up to check these out:
https://developers.redhat.com/topics/nodejs
https://developers.redhat.com/topics/nodejs/ai
https://github.com/nodeshift/nodejs-reference-architecture
