BAML Launch Week Day 4

Developer Spotlights

Launch Week isn't just about us - it's also about the incredible engineering work of amazing companies building with BAML!

Daniel Edrisian

Founder•Alex

What does your company do?

Alex is an AI coding assistant for Xcode, enhancing iOS and Mac development with features like sidebar chat, inline completions, and image-to-code generation.

How has BAML helped you?

BAML is the foundation of our chat functionality at Alex. It is a dead simple interface which lets us create Agentic experiences within days. The ability to stream structured outputs can't be found anywhere else - even in OpenAI/Anthropic's own libraries.

Faizan Sattar

CEO & Cofounder•Sherlock AI

What does your company do?

Sherlock is the ChatGPT for retail traders and investors. They're able to get AI-powered insights, track social trends, and execute trades effortlessly.

How has BAML helped you?

BAML has been crucial in building our Analyst AI agent. Structured responses with parallel function calling does not work natively with most LLMs. BAML types has unlocked this capability for all LLMs, making it super easy to map user input to the right output/behaviors.

Daniel Edrisian

Founder•Alex

What does your company do?

Alex is an AI coding assistant for Xcode, enhancing iOS and Mac development with features like sidebar chat, inline completions, and image-to-code generation.

How has BAML helped you?

BAML is the foundation of our chat functionality at Alex. It is a dead simple interface which lets us create Agentic experiences within days. The ability to stream structured outputs can't be found anywhere else - even in OpenAI/Anthropic's own libraries.

Faizan Sattar

CEO & Cofounder•Sherlock AI

What does your company do?

Sherlock is the ChatGPT for retail traders and investors. They're able to get AI-powered insights, track social trends, and execute trades effortlessly.

How has BAML helped you?

BAML has been crucial in building our Analyst AI agent. Structured responses with parallel function calling does not work natively with most LLMs. BAML types has unlocked this capability for all LLMs, making it super easy to map user input to the right output/behaviors.

Previous slideNext slide

Semantic Streaming

Structured Outputs vs. Streaming Data

LLMs provide almost magical reasoning abilities. The current challenge for application developers building on top of LLMs has been retaining the reliability and latency that users expect from normal web apps. We try to meet these goals by encoding our domain objects in structured output frameworks like BAML and JSON-mode, and using streaming APIs to deliver results incrementally as tokens come back from the LLM.

Sadly, these two techniques are at odds - a partially streamed message will not match the semantics you crafted in your structured output. Even if lenient parsers can patch up streaming json objects that are still waiting to receive their closing quotes and braces, the values inside those objects undergo violent janking as objects and numbers receive their new fields and digits, one token at a time.

We are excited to present "Semantic Streaming", a method for preserving domain semantics while streaming. Semantic Streaming allows you to specify invariants on structured outputs, and receive streamed outputs, not as a series of new tokens, but as a stream of semantically valid messages.

Semantic Streaming

Consider this BAML type, which has a new attribute @stream.done.

class PersonAssignment {
  person Person
  assignment string
}
 
class Person {
  name string @stream.done
  age int
}

The Person.name string @stream.done field indicates that a person is not semantically valid until we can be sure that the field is fully available in the stream. PersonAssignment.assignment string does not have this annotation, so PersonAssignment is valid before assignment is fully known.

With this semantic streaming, your application will only see valid states. You will never be given a Person with a name that is incomplete.

The following example shows what message would be delivered to your application at each step in the streaming process. Note that name doesn't appear one character at a time in the output messages, because it is marked @stream.done, whereas assignment does appear one character at a time.

Semantic Streaming Annotations

Another example shows the full set of annotations available for semantic streaming:

class Agenda {
  start_time string @stream.done // (1)
  items (Talk | Social)[]
  description string @stream.with_state // (2)
}
 
class Talk {
  type "talk" @stream.not_null // (3)
  speaker string
  title string
  duration_minutes int
}
 
class Social {
  type "social" @stream.not_null
  duration_minutes int // (4)
  @@stream.done
}

1. This @stream.done means that start_time will only be streamed when enough tokens have been processed to be sure that start_time is done. If you render start_time in your application, it will not change from 6 to 6:0 to 6:15PM as new tokens arrive.
2. @stream.with_state on description means that when your application uses the description field, that field will have extra metadata attached to it called stream_state. You can check the stream_state metadata field to make rendering decisions, for example the description text color could pulse while the description is streaming, but remain black when streaming is complete. In the video above, every field was marked with @stream.with_state so that we could render loading spinners or check marks for all the data.
3. @stream.not_null on the type field indicates that the parent class, Talk will not be streamed to your application until the type field is present. For fields like type, whose job is to indicate whether a JSON object is meant to represent an instance of Talk or an instance of Social, it is important that your application has this information before it begins to render an item.
4. There is no annotation on duration_minutes. BAML automatically considers numbers and literal types to be @stream.done. Your application will never need to render a 4 only for it to become 400000000 several tokens later.

Using semantic streaming to build progressive data views

The annotations you put onto your BAML types are reflected in the Python, TypeScript, and Ruby types you generate.

To understand these changes, remember that BAML generates a "Partial" version of your BAML types returned by streaming functions. A type Partial<T> is created from any type T by the following rules:

• If T is a class, all its fields will be replaced by partial versions of those fields, and the field may be null.
• If T is some other type that builds on top of other types, such as a tuple or a union, then it becomes a tuple or a union over the "Partial" versions of those other types.
• If T is a primitive type, it becomes T | null.

"Partialization" allows BAML to flexibly parse anything the LLM streams back. But this level of flexibility means that streamed values are unlikely to be semantically valid. Let's see how we can make the parser more strict with Semantic Streaming, and what impact that has on our generated types.

Returning to our example above with Agendas, Talks and Socials, this code will be generated:

# partial_types.py
 
class Agenda(BaseModel):
    start_time: Optional[str] = None # (1)
    items: List[Optional[Union["Talk", "types.Social"]]] # (2)
    description: StreamState[Optional[str]] # (3)
 
class Talk(BaseModel):
    type: Literal["talk"]
    speaker: Optional[str] = None
    title: Optional[str] = None
    duration_minutes: Optional[int] = None

# partial_types.py
 
export interface Agenda {
    start_time: string // (1)
    items?: (Talk | null | types.Social | null)[] // (2)
    description?: StreamState<(string | null)> // (3)
}
 
export interface Talk {
    type: "talk"
    speaker?: (string | null)
    title?: (string | null)
    duration_minutes?: (number | null)
}

1. start_time is still Optional, despite being marked as @stream.done. This may be confusing. @stream.done does not imply that the value is not null, in only implies that if a value is present, then it is complete.
2. items is a union of Talk and types.Social. types.Social is the version of Social defined in the non-streaming types.py file, and it is used here becase the entire Social BAML type was marked @stream.done.
3. description is wrapped in the StreamState struct because that BAML field was marked @stream.with_state. You can access the value at description.value and the streaming state at description.streaming_state.

To summarize the effect of Semantic Streaming on generated code: the invariants you specify through semantic streaming attributes will result in stricter types.

Better living through Domain Specific Languages

The central thesis of BAML is that LLMs present a such a novel interface to computation that they require a different set of tools to collaborate with them effectively. That's why we wrote a new language. The best developer experience for integrating LLMs into traditional applications necessarily has new language semantics to bridge the gap. In the past, BAML introduced LLM functions, LLM annotations, prompt syntax, and partial streaming, and prompting playgrounds.

As our customers encountered the tension between structured data invariants and streaming, the fact that we use a domain specific language to interface with LLMs made it simple to add Semantic Streaming attributes to the syntax, and to uphold semantic invariants in the runtime you use in your generated client code.

Here's one last example of Semantic Streaming. We stream the Ingredient quantities as whole numbers, and render each one at a time as it comes in. Note the spinner in the UI.

Learning a new language can be hard, but we think this is the best way of wrangling the complexity of LLMs for application developers. Please give Semantic Streaming a try! You can play with BAML at promptfiddle.com, learn more at docs.boundaryml.com, and hang out in our thriving Discord community.

Happy BAML-ing!