Flowrite Boosts Copy Generation Capabilities with Cohere

Overview

Writing takes time, effort, and focus — all of which can be hard to muster on any given day. Busy people often don’t have the bandwidth to craft thoughtful communications that help them stay connected and productive. Enter Flowrite, a generative AI platform that helps people stay on top of their communications backlog by drafting emails and messages for them across their web browser.

Flowrite is purposely designed for frequent and responsive email and messaging workflows, offering a broad collection of smart email templates for sales, marketing, recruiting, support, and other use cases. But Flowrite is not only for business use; customers also use the service to communicate with friends and family.

Improving Product Reliability and Output Quality

Under the hood, Flowrite uses large language models (LLMs) to power its text generation capabilities, and from the beginning, the startup took the managed LLM approach. “We didn’t have the budget for building a model in-house, and we didn’t want to manage the infrastructure,” recalls Bernardo García del Río, AI Lead at Flowrite. “Using an API instead of building and deploying our own model was the most sensible path at that time, and it still makes sense for us.”

Flowrite was originally built using OpenAI models, and the AI team was looking to expand their stack to include other LLM solutions in order to boost reliability and unlock new opportunities. They also wanted greater data security for enterprise customers, as well as access to expert technical support to help them achieve their goals faster.

García says, “We've been monitoring the LLM landscape from the very beginning, and we thought that Cohere was a clear competitor to OpenAI. And that's why we wanted to start to form a partnership early in the journey.” Not only did Cohere’s models perform well against OpenAI, and in addition, Cohere’s Accelerator program offered the startup access to technical expertise for tailored guidance and support.

Copy Generation, Data Diversity, and Technical Support

Flowrite’s primary use case for Cohere was copy generation, and the AI team proceeded to run Cohere’s Command model through their testing process, focusing on prompt engineering backed by real data to test models for specific tasks. The team looked at a mix of data points from user activity, along with visual output comparison by linguistics experts and company-wide testing. After conducting extensive testing with Command, the AI team was happy with the results and began actively working towards adding it to their production stack.

Flowrite has also used Cohere to help them prepare a diverse dataset in order to be able to generate content for a diverse range of topics. The AI team used Cohere's embeddings model to cluster data and evaluate the level of diversity of their dataset, so it could help them effectively fine-tune their models and improve latency and quality. “And the results were really good to be honest,” says García. “We found that using Cohere to calculate the distance between the vectors resulted in better clusters compared to OpenAI.”

To support these use cases, and more to come, Flowrite relies on Cohere’s technical expertise available through the Cohere Accelerator program. Flowrite was one of the first startups to join the program and, in fact, their early feedback helped to shape it.

As members, the team can reach out for support anytime on Slack, and gain early access to Beta releases, which helps them plan their roadmap. They also appreciate being closer to Cohere in a collaborative environment where they can give real-world feedback on Cohere products and features. And when Flowrite needs to scale in future, having a good partnership will help the company grow more easily.

Next Steps for the AI Team

For García and his team, getting hands-on with Cohere has been a great experience overall. “Cohere is really easy and straightforward to use,” he says, “In four or five lines of code, you’re generating an email. Cohere is also integrated into popular frameworks like LangChain, as well as the ChromaDB vector database that can automatically create embeddings. It makes things really efficient.”

Once testing is complete and the AI team has developed prompts that work well for Cohere, they plan to ship Cohere’s generative models to production. At that point, it will just be a matter of changing a little bit of code. As Flowrite grows its user base of consumers and businesses, they’ll be able to continuously gather more data to further improve model performance across their LLM portfolio.