Skip to main content

Schema-Aligned Parsing

Schema-Aligned Parsing (SAP) is BAML's approach to converting free-form LLM text into structured, validated data.

The Challenge​

LLMs output text, but applications need structured data:

LLM says: "The sentiment is definitely positive! :)"
You need: { sentiment: "positive" }

How BAML Solves This​

BAML's parser understands your schema and extracts matching data:

enum Sentiment {
  positive
  negative
  neutral
}

function Classify(text: string) -> Sentiment {
  client GPT4
  prompt #"Classify: {{ text }}"#
}

Even if the LLM responds with extra text, BAML extracts positive.

Parsing Strategies​

1. Exact Match​

For enums and simple types, BAML looks for exact matches:

Output: "I believe this is positive"
Parsed: Sentiment.positive

2. Structural Extraction​

For classes, BAML extracts fields flexibly:

Output: "Name: John, he is 25 years old"
Parsed: { name: "John", age: 25 }

3. JSON Mode​

When enabled, BAML uses the provider's JSON mode:

function Extract(text: string) -> User {
  client GPT4
  prompt #"..."#
  output_format json  // Use JSON mode
}

Error Handling​

When parsing fails, BAML provides detailed errors:

try {
  const result = await b.Classify(text);
} catch (e) {
  if (e instanceof BAMLParseError) {
    console.log(e.rawOutput);  // What the LLM actually said
    console.log(e.expected);   // What schema was expected
  }
}