Our Crossover with LangChain

Using LangSmith to tackle LLMs in production

We’ve been working a lot with GenerativeAI lately, with a framework for building production ready apps, and researching ways to generate the best possible images. But now it’s time to talk about generating text using LLMs (Large Language Models): the technology behind ChatGPT, LLaMa, Mistral and more.

As always, our focus is to be able to build production-ready systems out of these cutting-edge technologies. If it’s not running in production and helping users, then we don’t consider it done! We have been focusing on deploying LLMs, and we’re happy to share the challenges (and solutions) we’ve found along the way.

You’ll be able to read about our experiences in this post from LangChain’s blog, where our very own Tomas Pastore has shared his experience using LangSmith to measure LLMs.

Why we chose LangSmith

LangSmith is a developer platform by LangChain, capable of debugging, monitoring, testing, and evaluating LLM applications. Considering our focus on building AI products that are run in production environments, LangSmith was the right choice. It not only sped up development time but also helped us deliver a better end product.

In our use case, we needed to improve the quality of the responses of our LLM. We already had a test suite, but manually performing and rating our tests was time-consuming and introduced the bias of people in charge of correcting the outputs. This was simply not scalable for production.

In our crossover post with LangChain, you’ll be able to read about how we used LangSmith to build a Custom Evaluator with just a few lines of code, significantly improving the quality of our test suite and its metrics.

Find out more!

That’s just one of the many topics we discussed in LangChain’s blog post, where you’ll be able to read about improving prompt quality, security risks of LLMs, and more.

Find out more about the main challenges of generating text using LLMs and our recommended solutions for each of them in this detailed exploration.