System One in production
Jev is not a reranker
Public benchmarks found that Jev reranking alone does not beat vector retrieval. That is the right answer to the wrong question.
· 7 min read
WRITING
Notes from production.
What we learn running Navlu: typed decision models, retrieval and product discovery in ecommerce.
System One in production
One call, sixteen questions
A typed decision model answers a dictionary of questions against one state in a single pass. That changes the unit of design from the question to the decision point.
· 6 min read
Patterns
Make the model choose, not invent
Asking a model what the budget is invites a hallucinated number. Handing it every number in the sentence and asking which one is the budget does not.
· 6 min read
Hardening
Schema valid is not correct
Constrained decoding guarantees a well-formed answer, not a right one, and not one you should trust.
· 6 min read
Calibration
Choosing a confidence threshold
No published reliability curve backs the calibration claim. That does not stop you shipping, as long as you measure it on your own traffic.
· 5 min read