Cursor rule

1---
2description: 
3globs: *.ts,*.tsx,*.baml
4alwaysApply: false
5---
6<Overview>
7  BAML (Basically, A Made-Up Language) is a domain-specific language for building LLM prompts as functions.
8  You can build an agentic workflow with BAML.
9</Overview>
10 
11  <Schema>
12    // Define output schemas using classes
13    class MyObject {
14      // Optional string fields use ?
15      // @description is optional, but if you include it, it goes after the field.
16      name string? @description("The name of the object")
17      
18      // Arrays of primitives
19      // arrays cannot be optional.
20      tags string[]
21      
22      // Enums must be declared separately and are optional
23      status MyEnum?
24      
25      // Union types
26      type "success" | "error"
27      
28      // Primitive types
29      count int
30      enabled bool
31      score float
32 
33      // nested objects
34      nested MyObject2
35 
36      // image type
37      myImg image
38 
39      {#// checks and assertions. Uses jinja syntax inside the parentheses.
40      // For a single property use one @
41      bar int @assert(between_0_and_10, {{ "{{ this > 0 and this < 10 }}" }}) //this = MyObject.bar value
42      quux string
43      // assertions for multiple fields use @@ and go at the bottom of the class. Uses jinja syntax inside the parentheses.
44      // Do NOT add descriptions after the assertion.
45      @@assert(length_limit, {{ "{{ this.quux|length < this.baz }}" }})#}
46    }
47 
48    // Enums are declared separately
49    enum MyEnum {
50      PENDING
51      ACTIVE @description("Item is currently active")
52      COMPLETE
53    }
54 
55    // Comments use double slashes
56    // Recursive types and inline definitions are not supported
57 
58  </Schema>
59 
60  <Functions>
61    // Functions define inputs, outputs and prompts
62    // function name is always PascalCase
63    function MyFunction(input: MyObject) -> string {
64      client "openai/gpt-4o"
65      // prompt with jinja syntax inside here. with double curly braces for variables.
66      // make sure to include: \{\{ ctx.output_format \}\} in the prompt, which prints the output schema instructions so the LLM returns the output in the correct format (json or string, etc.). DO NOT write the output schema manually.
67      prompt #"
68        
69      "#
70    }
71 
72    <LLMClients>
73      You can use any of the following:
74      - openai/gpt-4o
75      - openai/gpt-4o-mini
76      - anthropic/claude-3-5-sonnet-latest (note the "3-5")
77      - anthropic/claude-3-5-haiku-latest
78    </LLMClients>
79 
80    <Prompt>
81      When writing the prompt:
82      1. Make sure to include the input in the prompt (even if it's an image) using {{ "{{ input }}" }}
83      2. Make sure to include {{ "{{ ctx.output_format }}" }} in the prompt so the LLM knows how to format the output.
84      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.
85      4. Write a {{ "{{ _.role(\"user\") }}" }} tag to indicate where the user's inputs start. So if there's a convo you can write
86      #"{{ "{{ _.role(\"user\") }}" }} {{ "{{ some-variable }}" }}#
87 
88      DO NOT REPEAT output schema fields in the prompt. They are included with {{ "{{ ctx.output_format }}" }}.
89```baml
90      class TweetAnalysis {
91        mainTopic string @description("The primary topic or subject matter of the tweet")
92        isSpam bool @description("Whether the tweet appears to be spam")
93      }
94 
95      function ClassifyTweets(tweets: string[]) -> TweetAnalysis[] {
96        client "openai/gpt-4o-mini"
97        prompt #"
98          Analyze each of the following tweets and classify them:
99          {{ "{{ _.role(\"user\") }}" }} {{ "{{ tweets }}" }}
100 
101          {{ "{{ ctx.output_format }}" }}
102        "#
103      }
104```
105    </Prompt>
106 
107  </Functions>
108 
109  <Usage in other languages>
110    You can use BAML in python, typescript, and other languages.
111 
112```python
113    import asyncio
114    from baml_client import b // this client is autogenerated
115    from baml_client.types import WeatherAPI
116 
117    def main():
118        # In python, BAML functions are synchronous.
119        weather_info = b.UseTool("What's the weather like in San Francisco?")
120        print(weather_info)
121        assert isinstance(weather_info, WeatherAPI)
122        print(f"City: {weather_info.city}")
123        print(f"Time of Day: {weather_info.timeOfDay}")
124 
125    if __name__ == '__main__':
126        main()
127```
128 
129```typescript
130    import { b } from './baml_client' // this client is autogenerated
131    import { WeatherAPI } from './baml_client/types'
132    import assert from 'assert'
133 
134    const main = async () => {
135      const weatherInfo = await b.UseTool("What's the weather like in San Francisco?")
136      console.log(weatherInfo)
137      assert(weatherInfo instanceof WeatherAPI)
138      console.log(`City: ${weatherInfo.city}`)
139      console.log(`Time of Day: ${weatherInfo.timeOfDay}`)
140        }
141```
142 
143  </Usage>
144 
145Do NOT use numbers as confidence intervals if you need to use them. Prefer an enum with descriptions or literals like "high", "medium", "low".
146Don't add confidence levels to extraction schemas.
147 
148Don'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.#}
149 
150Dedent all declarations.
151 
152Note that the types exported by BAML are pydantic classes in python, and interfaces in Tyepscript, except for primitive types.