Before you dive in
What you’ll find in the original
- Phrase the query as a predicate over each line, including actor and negation constraints when lexical matches would be ambiguous.
- Batch 30 lines per request and use bounded concurrency rather than sending every line separately.
- Return original matching lines with probabilities; unlike vector retrieval, each line is judged jointly with the question.
Worth knowing
Requires a TypeSafe API key and sends searched text to the service. The stated under-one-second timing for a 210-line file is project-reported and depends on network and service conditions.