As we've reported before, enterprise CIOs are taking generative AI in their stride. One reason is that AI doesn't fit into existing software engineering workflows, because it literally doesn't speak the same language. For example, LLMs (also known as large language models) require a lot of persuasion to deliver valid JSON.
That's where a US-based startup called puntotxt comes in, promising to “make AI talk to the computer.” The company is run by the team behind the open source project. Contornosque helps developers get what they need from ChatGPT and other generative AI models without having to resort to clunky tactics like injecting emotional blackmail into prompts (“write the code or the kitty understands it”).
Software libraries such as Outlines, a Python library, or Microsoft Steering, or LMQL (also known as Language Mannequin Question Language), make it possible to guide LLMs in a more sophisticated way than quick hacking, using an approach that It is known as structured generation (or sometimes constrained generation).
As the name suggests, the focus of the technique is on the output of the LLMs, rather than the input. Or, in other words, it's about telling AI models as respond, says Dottxt CEO Rémi Louf.
The approach “makes it possible to return to the traditional engineering workflow,” he told TechCrunch. “You refine the grammar until you get it right.”
Dottxt aims to build a powerful structured generation solution by being model-agnostic and offering more features (and, he says, better performance) than the open source project (Outlines) from which it was born.
Louf, a Frenchman who has a PhD and several degrees, has a background in Bayesian statistics, as do several other members of the Dottxt team. This grounding in probability theory probably opened their eyes to the potential of structured generation. Familiarity with IT beyond AI also influenced his decision to create a company focused on helping others usefully leverage generative AI.
Louf's reference to grammar is not informal. Dottxt's is based on the premise that most of the text we interact with is highly structured. There is code, of course, but also many other templates that LLMs should be able to follow to make them truly useful in work environments.
OpenAI, maker of GPT, recently introduced a form of structured generation it calls Structured Outputs, and nodded to Outlines as part of its “inspiration.”
Meanwhile, Louf sees the popularity of Outlines as a sign that there is demand for another versatile approach with more details. And investors seem to agree: Dottxt has raised $11.9 million in a matter of months.
The startup raised a $3.2 million seed round led by deep tech venture capital firm Elaia in 2023, followed by an $8.7 million seed round led by EQT Ventures in August. In the meantime, Louf and his co-founders have focused on working to prove that their approach doesn't impact performance. During this time, demand for open source schemes has skyrocketed; They say it has been downloaded more than 2.5 million times, which has encouraged them to think bigger.
Raising more funding made sense for another reason: Dottxt's co-founders now knew they wanted to use the money to hire more people so they could respond to the growing demand for structured generation tools. The startup's fully remote team will reach a staff of 17 by the end of the month, up from eight people in June, according to Louf.
The new staff includes two DevRel (developer relations) professionals, reflecting Dottxt's ecosystem building priority. “Our goal in the next 18 months is to accelerate adoption, rather than the commercial aspect,” Louf said. Although he also said that marketing will still begin within the next six months, with a focus on enterprise customers.
This could be a potentially risky approach if the AI hype ends when Dottxt seeks more funding. But the startup is convinced there is substance behind the bubble; Their hope is precisely to help companies unlock the current value of AI.
Hugging Face CTO Julien Chaumond, an investor in Dottxt, recently called structured generation “the future of LLM.” Therefore, the hype is also reaching this segment of the GenAI technology stack.
With other tailwinds, such as AI agents and the emergence of smaller AI models, Dottxt's adoption gamble could pay off. “Within a few years everyone will be using structured generation, there is no doubt about that,” Louf predicted.