The Story of BAML Error Handling

1. The Reality of AI Engineering

• LLMs = Probabilistic Failure: Unlike traditional code, every LLM call is a potential failure point (refusals, timeouts, hallucinations).
• The Need: You must handle errors, but you shouldn't have to rewrite your code to do it.
• Our Proposal: "Universal Catch". A simple, additive way to handle errors that works on functions, blocks, and expressions.

2. Design Evolution (Alternatives We Considered)

We'll evaluate each attempt against two common scenarios: 1. Declarative: ExtractResume (defining an LLM call). 2. Imperative: ProcessBatch (looping over items).

• The Base Cases:

  // 1. Declarative
  function ExtractResume(text: string) -> Resume {
    client "gpt-4o"
    prompt #"..."#
  }
  
  // 2. Imperative
  function ProcessBatch(urls: string[]) -> Resume[] {
    // If this fails, we still want to process resumes!
    let aggregator = MetricsAggregator.new() 
    let results = []
  
    for (url in urls) {
      let resume = ExtractResume(url)
      aggregator.record(resume)
      results.append(resume)
    }
    return results
  }
  

• Attempt 1: The Classic try/catch Statement

  • Imperative: The "Hoisting Tax". To handle aggregator failing, you must declare let aggregator; outside, then try { aggregator = ... }, then check if (aggregator) { ... } inside the loop.
  • Declarative: Forces wrapping the client definition (awkward).
  • Why we moved on: Structural refactoring hurts both.
• Attempt 2: Result Types (Result<T, E>)
  • Imperative: MetricsAggregator.new() returns Result. You must match or unwrap it.
  • Declarative: ExtractResume now returns Result.
  • Why we moved on: Viral complexity.
• Attempt 3: Expression-Oriented Try (let x = try { ... })
  • Imperative: let aggregator = try { ... } (Great for this case!).
  • Declarative: let resume = try { client ... } (Confusing).
  • Why we moved on: Great for imperative, but confusing for declarative.
• Attempt 4: Function Modifiers (function ... try)
  • Both: Syntactically awkward.
• Attempt 5: Wrapper Functions
  • Both: Boilerplate explosion.

3. The Solution: Universal Catch

• The Concept: catch is an operator that can be attached to ANY block.
• The Unification:
  • Attach to a Function -> Implicit Try (Scope = Function Body).
  • Attach to a Block -> Explicit Try (Scope = Block).
  • Attach to an Expression -> Inline Catch.

4. Learn by Example

• Scenario 1: The "Prototype to Production" Flow
  • Show: A clean LLM function.
  • Action: Append catch to handle a timeout. Zero indentation change.
• Scenario 2: The "Resilient Loop"
  • Show: A loop processing a batch of URLs.
  • Action: Attach catch to the loop body to prevent one failure from crashing the batch.
• Scenario 3: The "Pipeline" (Inline Catch)
  • Show: A chain of operations where one step is optional.
  • Action: Use catch { _ => null } on a single expression.
• Scenario 4: The "Complex Logic" (Imperative Try)
  • Show: A mix of safe and risky code in one function.
  • Action: Use an explicit try { ... } block for the risky part to signal intent.