boundaryml

boundaryml, also known as BAML (Boundary AI Markup Language), is a powerful and expressive language designed specifically for developers working with Large Language Models (LLMs). Its primary goal is to solve a critical challenge in AI development: reliably obtaining structured data, such as JSON, from the often unpredictable outputs of LLMs. BAML replaces fragile prompt engineering and manual parsing with a robust, type-safe framework that treats LLM interaction as a core part of the software development lifecycle.

The platform provides a comprehensive toolkit that simplifies the entire process of defining, testing, and deploying AI-powered data extraction pipelines. By defining data schemas and prompts within `.baml` files, developers can leverage static analysis, real-time feedback, and a dedicated VSCode playground to iterate quickly. BAML's intelligent parser is a standout feature, automatically correcting common LLM errors like trailing commas, unquoted keys, and other JSON formatting issues, thus preventing runtime failures and ensuring data integrity.

How to use boundaryml

Getting started with boundaryml is straightforward for developers. The process begins with installing the necessary package via pip:

$ pip install baml-py

1. Define Your Schema: Create a `.baml` file in your project. Inside this file, you define the desired output structure using BAML's intuitive syntax, which includes classes and enums, similar to modern programming languages. You also write the prompt that will be sent to the LLM, using Jinja templating for dynamic content.

2. Develop and Test in the Playground: Use the BAML VSCode extension, which provides an integrated playground. This allows you to test your prompts against different LLMs (like GPT-4o, Claude 3.5, etc.), view the real-time output, and debug any issues before writing any application code. The playground supports multimodal inputs, allowing you to test with images and audio as well.

3. Generate and Use the Client: BAML's compiler generates a type-safe client in your chosen language (e.g., Python, TypeScript). You can then import and call your BAML functions directly in your application code, just like any other library function. BAML handles the underlying LLM API calls, parsing, and error correction.

4. Deploy with Confidence: Once integrated, your application can reliably call LLMs to get structured data. For production environments, boundaryml offers Boundary Studio, an MLOps suite for observability, monitoring, and fine-tuning.

Core Features of boundaryml

• Expressive BAML Language: A dedicated syntax for defining prompts and data schemas, turning prompt engineering into a more structured coding practice.
• Advanced Error-Correcting Parser: Automatically fixes broken JSON and other formatting errors from LLM outputs, ensuring high reliability.
• Model Agnostic Function-Calling: Works seamlessly with a wide range of models, including those from OpenAI, Anthropic, Google, and open-source alternatives, often outperforming native function-calling capabilities.
• Type-Safe Client Generation: Generates clients for multiple languages (Python, TypeScript, Ruby, Go, etc.), providing full type safety and editor autocompletion.
• Integrated VSCode Playground: An interactive environment for rapid prototyping, testing, and debugging of prompts and data extraction logic.
• Semantic Streaming: A sophisticated technique for streaming structured data objects, not just raw text tokens, enabling more responsive user experiences.
• Multimodal Capabilities: Supports non-text inputs like audio and images within prompts.
• Open Source Core: The core BAML language and tools are free and open-source under the Apache 2.0 license.

Use Cases for boundaryml

boundaryml is ideal for any application that relies on structured information from LLMs:

• Data Extraction: Parsing unstructured documents like resumes, invoices, contracts, and customer emails into structured formats.
• AI Agent Development: Building reliable AI agents that use tools and functions by ensuring the LLM's output correctly matches the required function signature.
• Content Classification and Tagging: Automatically categorizing user feedback, support tickets, or articles based on their content.
• RAG (Retrieval-Augmented Generation) Systems: Structuring the output of RAG pipelines, for example, to generate answers with citations in a consistent format.
• Natural Language to API: Translating user requests in natural language into structured API calls or database queries.

Advantages of boundaryml

Developers choose boundaryml for several key advantages:

• Enhanced Reliability: Drastically reduces parsing failures and eliminates the need for complex `JSON.parse()` try-catch blocks.
• Improved Developer Experience: The code-like syntax, static analysis, and integrated testing playground significantly accelerate iteration speed and improve code quality.
• Cost and Performance Optimization: BAML's efficient prompting techniques can reduce token usage and decrease time-to-first-token without sacrificing accuracy.
• Cross-Platform and Polyglot: Define logic once in BAML and use it across different services written in various programming languages.
• State-of-the-Art Results: Benchmarks show that BAML achieves superior performance in function-calling tasks compared to native model implementations.

Pricing and Plans

boundaryml operates on a freemium model, making it accessible to everyone from individual developers to large enterprises.

• Starter Plan (Free Forever): This plan is completely free and includes the core BAML language (Apache 2.0 Licensed), the ability to get structured data from LLMs, the VSCode playground with multimodal capabilities, and community support via Discord and GitHub.
• Enterprise Plan (Custom Pricing): Tailored for businesses requiring the highest level of reliability and support. It includes everything in the Starter plan plus access to Boundary Studio (an MLOps suite with observability, data labeling, and fine-tuning support), SLA guarantees, dedicated Slack support, architectural reviews, and prioritized feature requests. Interested parties should contact sales for a quote.