dottxt is focused on making LLM outputs more predictable
dottxt, founded by the creators of the popular Outlines library, has raised $11.9 million to date to develop advanced structured generation solutions for large language models, aiming to ensure more reliable and predictable AI outputs.
This week, EQT Ventures confirmed its financial backing of dottxt as a co-leading investor in the startup's $8.7 seed funding round this August, together with Elaia, and joined by Seedcamp, Common Magic, Kima Ventures, Script Capital, and STATION F. Previously, in December 2023, Elaia and Kima Ventures backed dottxt with $3.2 million in a pre-seed round, for a total of $11.9 million raised to date.
dottxt's co-founders Brandon T. Willard (CSO), Rémi Louf, (CEO), and Daniel Gerlanc, (CTO) happen to be the team behind Outlines: a popular Python library that provides developers with tools to interact with LLMs to ensure model outputs are reliable. This may seem like a minor qualm, but examples of model output unpredictability abound. It may be as simple as instructing an LLM to perform sentiment analysis on reviews to categorize them as positive and negative, only to find the model has also included a "neutral" category, which it was not asked to do.
However, the problem of output unpredictability may be as complex as having a model refuse to format its outputs as valid JSON, where outputting valid JSON is often a key step to guaranteeing the usability of the model's outputs within a more complex system without the need for human oversight. Indeed, the issue with what is now called 'structured generation' is so pressing, that even OpenAI acknowledged it drew inspiration from Outlines when it launched Structured Outputs on its API to coerce its models to adhere to JSON schemas supplied by developers.
The funding has helped dottxt grow its fully remote team from 8 to 17 staff members, a necessary move to cope with rising demand for its solution (Outlines has been downloaded over 2.5 million times). The startup's agenda includes building a powerful, model-agnostic, structured generation solution that offers more features and better performance than Outlines. As part of this plan, dottxt is beta-testing a model for reliable valid JSON output generation. The company says it will first focus on increasing adoption, and then start commercializing its products among enterprise customers.
