AIVAX Reflex
The useful passage is there. Put it in front.
Your application already has candidate documents. Rank them around the question at hand, without creating a collection first.
Same candidates. A more useful order.
Follow the evidence as the question changes.
Compare the original order with two illustrative rankings. The documents stay the same; the question changes what belongs first.
Three candidates. Which one helps?
Your application supplies the text. Choose a question above to see an illustrative reordering.
What should I do if my delivery arrives damaged?
The passage about damaged deliveries moves to the front. The next step is available in the supplied text.
Where can I download last month's invoice?
The billing passage now belongs first. A different question changes which evidence is useful.
No collection required
- 01Update your contact address
Open account settings to change the address used for future orders.
- 02Keep a photo and your order number
Send a photo of the damaged item and your order number to the support team.
- 03Download an invoice from your account
Open Billing, select the month and download the corresponding invoice.
Worked example · editorial sample IDs and illustrative order, not a benchmark · no live call. Reflex cannot recover evidence that is absent from the supplied candidates.
Recurring documents, new questions
Don't make every return a first visit.
Exact repeated text can reuse cached processing. A changing question can still benefit from material processed earlier.
The benefit depends on the workload. A cold, one-off candidate set does not get the same reuse.
Conceptual reuse, not a measured cache-hit rate.
Have the candidates, or need the search system?
Reflex ranks supplied text. Managed RAG stores, indexes and searches a persistent corpus. Choose based on where the candidate set comes from.
Reflex
Your candidates, ordered for the current question.
Bring a question and a candidate list.
Start with examples where you know what relevant evidence looks like. Inspect the order before using it to assemble context.
Processed and cached input have different rates.
Inspect usage on a representative workload to understand how much text is reused, rather than assuming a fixed saving.
Reflex billing and limits