baml

Instruction file imported from mikulskibartosz/laoshu (.cursor/rules/baml.mdc). Copyright stays with the author.

description: globs: *.baml alwaysApply: false

  // Arrays of primitives
  // arrays cannot be optional.
  tags string[]

  // Enums must be declared separately and are optional
  status MyEnum?

  // Union types
  type "success" | "error"

  // Primitive types
  count int
  enabled bool
  score float

  // nested objects
  nested MyObject2

  // image type
  myImg image

  {#// checks and assertions. Uses jinja syntax inside the parentheses.
  // For a single property use one @
  bar int @assert(between_0_and_10, {{ "{{ this > 0 and this < 10 }}" }}) //this = MyObject.bar value
  quux string
  // assertions for multiple fields use @@ and go at the bottom of the class. Uses jinja syntax inside the parentheses.
  // Do NOT add descriptions after the assertion.
  @@assert(length_limit, {{ "{{ this.quux|length < this.baz }}" }})#}
}

// Enums are declared separately
enum MyEnum {
  PENDING
  ACTIVE @description("Item is currently active")
  COMPLETE
}

// Comments use double slashes
// Recursive types and inline definitions are not supported

  "#
}

<LLMClients>
  You can use any of the following:
  - openai/gpt-4o
  - openai/gpt-4o-mini
  - anthropic/claude-3-5-sonnet-latest (note the "3-5")
  - anthropic/claude-3-5-haiku-latest
</LLMClients>

<Prompt>
  When writing the prompt:
  1. Make sure to include the input in the prompt (even if it's an image) using {{ "{{ input }}" }}
  2. Make sure to include {{ "{{ ctx.output_format }}" }} in the prompt so the LLM knows how to format the output.
  3. You do not need to specify to "answer in JSON format". Only write in the prompt brief instruction, and any other task-specific things to keep in mind for the task.
  4. Write a {{ "{{ _.role(\"user\") }}" }} tag to indicate where the user's inputs start. So if there's a convo you can write
  #"{{ "{{ _.role(\"user\") }}" }} {{ "{{ some-variable }}" }}#

  DO NOT REPEAT output schema fields in the prompt. They are included with {{ "{{ ctx.output_format }}" }}.
  ```baml
  class TweetAnalysis {
    mainTopic string @description("The primary topic or subject matter of the tweet")
    isSpam bool @description("Whether the tweet appears to be spam")
  }

  function ClassifyTweets(tweets: string[]) -> TweetAnalysis[] {
    client "openai/gpt-4o-mini"
    prompt #"
      Analyze each of the following tweets and classify them:
      {{ "{{ _.role(\"user\") }}" }} {{ "{{ tweets }}" }}

      {{ "{{ ctx.output_format }}" }}
    "#
  }
  ```
</Prompt>

```python
import asyncio
from baml_client import b // this client is autogenerated
from baml_client.types import WeatherAPI

def main():
    # In python, BAML functions are synchronous.
    weather_info = b.UseTool("What's the weather like in San Francisco?")
    print(weather_info)
    assert isinstance(weather_info, WeatherAPI)
    print(f"City: {weather_info.city}")
    print(f"Time of Day: {weather_info.timeOfDay}")

if __name__ == '__main__':
    main()
```

```typescript
import { b } from './baml_client' // this client is autogenerated
import { WeatherAPI } from './baml_client/types'
import assert from 'assert'

const main = async () => {
  const weatherInfo = await b.UseTool("What's the weather like in San Francisco?")
  console.log(weatherInfo)
  assert(weatherInfo instanceof WeatherAPI)
  console.log(`City: ${weatherInfo.city}`)
  console.log(`Time of Day: ${weatherInfo.timeOfDay}`)
    }
```

Do NOT use numbers as confidence intervals if you need to use them. Prefer an enum with descriptions or literals like "high", "medium", "low". Don't add confidence levels to extraction schemas.

Don't use LLM functions to "validate" any other output. {#You should use @assert for that on each field in the output type. Search the docs for "assert" to see how to use it.#}

Dedent all declarations.

Note that the types exported by BAML are pydantic classes in python, and interfaces in Tyepscript, except for primitive types.