Reflex V1.5: a reranker that reads "from largest to smallest"
Reflex V1.5 is a new AIVAX reranker. It ranks queries that ask for an order by attribute better than Reflex V1, accepts 15,000 documents per request, and keeps the same request body. What we measured, and when V1 is still the right default.
Ask a product search for "drills, from most powerful to least" and a reranker that only scores topical relevance has no reason to put the 135 Nm hammer drill above the 30 Nm drill driver. Both are drills and both match the query. What the user asked for was the order.
Reflex V1.5 (@aivax/reflex-v1.5) went live on 16 November 2026. It is the second version of the reranker we built for AIVAX, and the clearest gain we measured over Reflex V1 is on exactly that kind of query: a request to order documents by an attribute. Reflex V1 is still what you get when you do not name a model, and there are workloads where we would keep it.
The biggest gain is ordering by attribute
We measured V1.5 against V1 on a benchmark we built for intent and attributes: 174 multilingual tasks in 8 languages, each a query with its candidate pool, with graded relevance assigned by independent LLM judges. The metric is nDCG@10. The confidence intervals come from a paired bootstrap, so they describe the difference between the two rankers on the same tasks.
| Task type | Reflex V1 | Reflex V1.5 | Difference |
|---|---|---|---|
| All tasks | 0.776 | 0.797 | +0.021 (+0.004 to +0.038) |
| Ordering by attribute | 0.769 | 0.812 | +0.043 (+0.009 to +0.080) |
| Several combined constraints (6 tasks) | 0.631 | 0.753 | +0.122 |
Ordering by attribute covers queries such as "from largest to smallest" or "from most aggressive to least". Its interval stays clear of zero, and it is the reason we would try V1.5 first on any catalog where users sort in plain language.
Queries with several constraints at once ("a waterproof hiking jacket with a hood that packs into its own pocket") were V1's weakest category, and they moved the most, from 0.631 to 0.753. There are only six of those tasks in the set, so we read that as a positive direction and would not quote the size of the gain. Implicit needs, where the query describes a situation instead of a product ("a phone for my grandmother who only calls and texts"), went up by 0.027, a positive direction that is not statistically significant on this benchmark.
On human labels, the gain is in English
The second measurement uses MIRACL, a public multilingual retrieval benchmark with human relevance labels. We ran it in German, English and Spanish: 135 queries, each with a pool of 100 candidates, scored by nDCG@10. Reflex V1 scored 0.694 and V1.5 scored 0.727, a gain of 0.033.
Most of that gain comes from English. In German the two versions are close to even, so a German or mostly non-English corpus should not expect the English gain to carry over.
Where V1 holds its ground
Two task types in the intent benchmark did not move in V1.5's favor. On filtering by category ("only the laptops"), V1 scored 0.917 and V1.5 scored 0.882. On generic topical queries the two tied. Negation ("not for gaming") and numeric filters ("under 50 dollars") are still unsolved in both versions, which get part of those cases wrong. If a condition must hold, enforce it before ranking, with metadata filters on the collection or in your own query code.
Time is the other trade-off. The first time V1.5 sees a document, it takes longer than V1 to score it. Documents the same account sent in the last 30 minutes are not processed again, and on repeated documents reranking 50 of them takes about 0.1 seconds on the server and 500 of them about 0.7 seconds. That rewards the pattern the first Reflex post argued for: a corpus whose documents keep coming back across many questions. For a pool of fresh documents on every request, where latency matters more than order, V1 is the safer pick.
Limits and price next to V1
| Limit or price | Reflex V1 | Reflex V1.5 |
|---|---|---|
| Documents per request | 10,000 | 15,000 |
| Context per document | 1,948 tokens | 1,948 tokens |
Maximum top_n | 200 | 50 |
| New input tokens | $0.015 per million | $0.015 per million |
| Cached input tokens | $0.003 per million | $0.005 per million |
With 15,000 documents, a first-pass search that returns ten or twelve thousand candidates can go to the ranker without a cut. The top_n ceiling goes the other way: V1.5 answers a request with top_n above 50 with a 400 error, while V1 returns up to 200, so a long results page reranked in one call still needs V1. New tokens cost the same in both versions. Cached tokens cost $0.005 per million on V1.5 against $0.003 on V1, so a workload made almost entirely of repeats pays somewhat more on V1.5. V1.5 requests count toward the plan's daily Reflex allowance the same way V1 requests do.
Switching is the model field
The reranking endpoint takes the same body for both versions. Send it to POST https://inference.aivax.net/api/v1/generations/rerank with an Authorization: Bearer <API_KEY> header:
{
"model": "@aivax/reflex-v1.5",
"query": "How do I rotate an API key without downtime?",
"documents": [
"Billing invoices are available in account settings.",
"Create a second key, deploy it to every service, confirm traffic on the new key, then revoke the old one.",
"API keys can be scoped to read-only permissions.",
"Revoking a key immediately rejects every request that uses it."
],
"top_n": 3
}
The response lists results[].index and relevance_score for each returned document, and usage separates input_tokens from cached_input_tokens, so you can see how much of a request was served from the cache. Sending the same body a second time came back with every token counted as cached. Omitting model still selects Reflex V1.
In collection search, set "reranker": "@aivax/reflex-v1.5" on the query. Collection search drops semantic candidates below minScore before any reranker sees them, so when you switch rankers on a collection, recheck that threshold as described in RAG similarity threshold. If the right document never reaches the pool, no version of Reflex will find it; Do I need a reranker for RAG? separates that failure from a bad order.
Our recommendation is to run both versions over the same candidate pools from your own traffic, with a few queries that sort by an attribute and a few that combine constraints, and keep whichever puts the right documents in the slice you actually send to the model. If your users sort in plain language, that is the case V1.5 was built for, and the one where we would try it first.
The benchmark figures come from the AIVAX team's measurements for the 16 November 2026 release. The example request, the top_n limit and the cached usage were checked against the live AIVAX API on 18 November 2026.
Update, 20 November 2026: Reflex V1 will be deprecated on 25 November 2026. If your requests omit model or name @aivax/reflex-v1, test them on Reflex V1.5 before that date, including any that ask for more than 50 results.