Eval-driven development: the only way we ship agents
We write the test suite before the agent. Here is how a golden set, a grader and a CI gate keep our AI systems honest.
Retrieval-augmented generation looks simple in a tutorial: split documents, embed them, search, answer. In production, every one of those steps hides a failure mode.
Chunk by meaning, not by length
Fixed-size chunks cut tables in half and separate questions from answers. We split on document structure — headings, list items, table rows — and keep a link back to the source section.
Use hybrid search
Vector search is great at meaning and poor at exact terms like product codes. Combining it with keyword search fixes most “it couldn’t find the obvious thing” complaints.
Always cite
Every answer links to the passage it came from. Citations make wrong answers easy to spot and right answers easy to trust.