now ChatGPT writes to you instead of recruiters

My name is Serezha, I’m a product in the Selection team. Today I will tell you how we made a ChatGPT-based email generator for recruiters.

Why can’t a recruiter write a normal letter himself?

Maybe, but:

  1. Recruiters spend a lot of time searching for candidates and making the first contact. On average, it takes a recruiter 15 minutes to create a personal letter. Imagine that you need to write at least 20 such letters a day. That’s 5 working hours for 20 good emails. Of course, the quality of the writing is sacrificed in the end. Most often, recruiters simply use templates – and this is ok, but candidates, of course, expect a personal approach.

  2. Developers are unhappy with the messages and emails they receive. You don’t want to feel like you’re 10th in line for a “cool opportunity.”

And what did we decide to do?

We have taken on an email generator that will allow you to create personalized messages based on the vacancy and the profile of the candidate. According to our estimates, this will save the recruiter about 18 hours a week. This means focusing on other important tasks – interviews, finding only suitable candidates, screening and collecting analytics throughout this process 🙂

It is important that the letters are different, readable, coherent and personalized. When we took on this task, ChatGPT 4.0 had already been released, but we decided to work with ChatGPT 3.5-turbo, and not with the latest version, because:

  • First, ChatGPT 4.0 is 10-15 times slower than ChatGPT 3.5 due to increased model size. If ChatGPT 3.5 requires an average of 5 to 10 seconds per response to generate emails, then ChatGPT 4.0 had the best result of 45 seconds. It’s very slow.

  • Secondly, new technologies are often unstable, and ChatGPT 4.0 is no exception. During testing, frequent chat disconnects were noticed, which can last up to 5 minutes. Although this may improve over time, at the moment the stability of ChatGPT 3.5 in our tests was higher, which reduces the risks for us.

  • Finally, in order for ChatGPT 3.5 to be what it is now, the OpenAI team did some additional training and model tuning. Thanks to their painstaking work, the chat responses seem very human-like. ChatGPT 4.0 has no issues with humanity and, in some cases, even surpasses its smaller brother. However, the tests carried out on our target sample showed that a person cannot notice a big difference in quality between them.

So ChatGPT 3.5 is the best option for us right now. While there are many other large language models on the market, no one has yet achieved the quality of responses that the OpenAI model has.

We used ChatGPT 3.5-turbo for the generator.

To make the emails personal and interesting, we took into account many parameters, for example, what technologies the candidate uses, what experience he has, what projects he worked on, and much more.

How will this feature affect the lives of recruiters and candidates?

Recruiters will spend ~1 minute writing a letter. It will be personal, specific, engaging. The recruiter will be able to reach more candidates in a short time, which means filling vacancies faster and spending time more efficiently (on other cool tasks).

Candidates will receive clear and simple emails with proposals. If you use the generator together with the Selection, then the offers will be significantly more relevant.

What should be inside the letter?

Here is the real message I received a week ago:

Sergey, good afternoon! My name is [имя скрою, чтобы никого не обидеть]career service [компанию тем более не буду светить, вдруг они наши клиенты 😀]. Colleagues shared your contact with me. We have a vacancy at work right now that I would like to discuss with you. Tell me, please, would you be interested in talking with me?”

I have questions “What vacancy?”, “What company?”, “Why are they interested in me?”. You need to understand that this is a cold letter, I’m not currently searching, respectively, in order to find out whether it makes sense to communicate, I need to answer and start a dialogue. There is no information in this letter that could interest me, besides, I am too lazy to start a long conversation and ask the recruiter, like most of those who are not actively searching.

But if they wrote to me:

Sergey, good afternoon!
My name is Varya. Colleagues shared your contact with me. We are at work now there is a vacancy for Product Manager in a large telecommunications company. This company starts HR-Tech project, and in your experience I saw that you are working in this area. Tell me, please, would you be interested in talking to me?

Agree, I would be hooked that a person studied my experience, voiced my value for their project. I think I would answer.

I gave this example to show how much information a recruiter needs to know about a job and a candidate in order to write a quality letter. It is important to hook the candidate with something, to show why he is suitable for this vacancy. So, how to hook a candidate?

Personal letter. Yes, it’s that simple.

Ok, so how do you personalize an email?

The main task of the generator is to connect the candidate with the vacancy. Find arguments that will convince a person that the vacancy is interesting and suits him.

To get a quality letter, we need to know enough information about both the vacancy and the candidate. There are no problems with information about the candidate, we specialize in this (aggregate data about candidates from 15+ open sources ala Linkedin and Github), but we don’t know anything about the vacancy yet.

So we ask the user:

  • About the company: the name of the recruiter, the name of the company, the area (for example, is it fintech or ed-tech)

  • About the vacancy: stack (we pull up our dictionaries), location (remote work is also possible), duties (free field), salary fork

  • About the candidate: name and link to his Linkedin (or profile in the Selection)

What we ask from ChatGPT:

Write a letter that will attract the attention of the candidate. We set restrictions on the format, tone, and give directions on how we can use the sources we have (for example, where we found the candidate and what interesting things we learned about him).

How was it integrated into the product?

We are currently testing two scenarios:

  1. Letter Generator – in fact, this is just a page where you can fill in all the above fields and immediately receive a letter.

  2. Letter generator based on vacancy – we also have such a scenario in which the recruiter needs even fewer actions to quickly create a message for the candidate.

What you need to do to generate letters in 10 seconds

What you need to do to generate letters in 10 seconds

Testing and improvements (the most interesting)

We made the back part of the generator and began to test it. Of course, we were still spinning the promt, but here’s what we still encountered.

Paid account

We tested on the free Open AI API key, the average response time is about 15 seconds. Sometimes I had to wait for a response from Open AI for 30 seconds. We’re trying to save users time, and instead make them wait 30 seconds per page while looking at a spinning loading spinner.

Next, we issued a paid API key, the response time was reduced to 4-5 seconds. In rare cases it is 10 seconds. And the limits on the paid key are much higher.

Free key

Paid key

20 requests per minute

3500 requests per minute

Let’s see if so many requests per minute will be enough for us, if not, then we will try to increase our limits through the support of Open AI, they are loyal guys.

Open AI errors

When we tested on a free key for every 50 requests, we caught a 429 error. In their documentation this error is responsible for both exceeding the limits and overloading the server. We definitely did not exceed the limit of 20 requests per minute, as we did them manually.
The day before the Open AI test, the cost of chatgpt was reduced by 10 times, and we assume that this is why there was an overload.

Further, with a paid key, we never caught such an error. Whether this is due to the fact that the hype has died down or we don’t know if we have a paid key, but we have added error handling.

We selected the coefficient of creativity

Also during the tests, we selected the values ​​of two parameters:

  • temperature – parameter responsible for model creativity

  • frequency_penalty – responsible for the frequency of repetition of already existing words, the higher the parameter, the model tries to use new words that have not previously been found in the text

If you do not change the temperature values, but change the frequency_penalty, then creativity also increases markedly. It is logical, because the models need to select new formulations so that they do not repeat with the past.

For example, when we raised the frequency_penalty, the generator began to set deadlines for responding to a letter, come up with new addresses to which the candidate must answer or say that Samokat is an online store for electric transport))

To find the perfect combination, we changed the parameter values ​​in increments of 0.05 and generated several dozen letters to evaluate the effect of the changes.

Example

We got a job SberMarket and generated a letter for them. Here’s what happened:

I left a screen on purpose :) Really generated letter: not perfect yet, but already better than many templates.

I left a screen on purpose 🙂 Really generated letter: not perfect yet, but already better than many templates.

Where can you poke

We have done open generator, it can be used once a day. The created letter will fall to your mail.

What functionality we already have, but we haven’t rolled it out yet

So far, we have made a beta version to debug the prompt and see how recruiters will use it. What else will we do:

  1. Telegram message vs Letter The recruiter will be able to choose whether to write a short informal message for the messenger or a letter with more details.

  2. Sentiment You can choose: formal / informal / creative message.

  3. Adding benefits, a story about the company, important details in free form. This will allow us to make the letters even more personal and even more in line with the tone of voice of a particular company.

In general, if you now see an email from a recruiter, ChatGPT may have written it.

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *