dottxt

dottxt: make LLMs return valid, schema-shaped output on every single call

dottxt (styled .txt) exists to solve one of the most annoying problems in production AI: models that return almost-JSON. Instead of prompting, retrying, and defensively parsing, dottxt uses constrained decoding — the model is mathematically prevented from emitting tokens that would violate your JSON Schema, regular expression, or context-free grammar. The output is valid on the first call, every call, with no retries or validation loops.

The company is best known as the maker of Outlines, the open-source structured generation library that has become a default choice in the Python LLM stack. On top of it, dottxt sells a hosted API (api.dottxt.ai, pay-per-token on open-source models) and self-hosted, drop-in replacements for vLLM, SGLang, and TensorRT-LLM inference servers, plus composable libraries — dotjson for JSON Schema enforcement, dotgrammar for grammars, and dotlambda for function calling. The pitch is simple: make AI behave like software, with enforced contracts between the model and your application.

Key Features

• Constrained decoding that guarantees outputs match JSON Schema, regex, or context-free grammars — no retries or defensive parsing
• Outlines, the open-source structured generation library with 65M+ downloads and 14,000+ GitHub stars
• Hosted API on open-source models with pay-per-token pricing
• Self-hosted drop-in replacements for vLLM, SGLang, and TensorRT-LLM inference servers
• Composable libraries: dotjson (schemas), dotgrammar (grammars), dotlambda (function calling)
• Negligible latency overhead versus unconstrained generation

Ideal Use Case

dottxt is for engineering teams putting LLMs into real pipelines — extraction, classification, function calling, synthetic data, agent tool-use — where a malformed response breaks downstream code. If you run open-source models on your own inference infrastructure and are tired of writing retry-and-repair loops around model output, Outlines or a dottxt-powered inference server slots directly into what you already run. Teams that only call closed APIs with light formatting needs will feel less pain and less benefit.

How dottxt differentiates

Structured generation is dottxt's entire company, not a feature bolted onto a broader platform. The technical depth shows: OpenAI publicly credited Outlines as an inspiration for its own Structured Outputs feature. The company raised $11.9M across a $3.2M pre-seed led by Elaia and an $8.7M seed led by EQT Ventures in 2024, with angels including Hugging Face CTO Julien Chaumond, and its work has been featured by AWS for production structured generation. Unlike validation-based tools such as Instructor or Guardrails that check output after the fact, dottxt constrains the token stream itself.

FAQ

How is dottxt different from the Outlines library? Outlines is dottxt's free, open-source library. The company adds a hosted API, hardened self-hosted inference servers, and enterprise-grade libraries built on the same constrained-decoding tech.

Does it work with OpenAI or Anthropic models? Constrained decoding requires logit-level access, so dottxt's guarantees apply to open-source models on its API or your own inference servers, not closed third-party APIs.

Does constraining output hurt quality or speed? dottxt's research shows structured generation adds negligible latency and can improve task accuracy by eliminating malformed responses.

What does it cost? Outlines is free and open source; the hosted API is pay-per-token, and self-hosted deployments are priced by inquiry.

tl;dr

dottxt, the $11.9M-funded company behind Outlines (65M+ downloads), guarantees LLM output matches your schema on every call — free open source, paid API and self-hosted servers.

Related

Looking for more options? Browse the Developer Tools directory or read our best AI coding tools listicle. dottxt is also tracked on Crunchbase.