Rerank search results

Short version: you cannot rerank with Strata today. The command exists and no available provider implements it. This page says so plainly so you do not spend an afternoon finding out.

What happens if you try

strata ./docs inference rank openai:gpt-4o-mini "query" "passage one" "passage two"

inference.unsupported_operation: not supported: `openai:gpt-4o-mini` is served by
openai, which cannot rank.
  hint: Use an inference operation supported by the selected provider and model.

Check before calling rather than catching the error:

strata ./docs inference capability openai:gpt-4o-mini

can_generate  true
can_embed     true
can_rank      false

can_rank is false for both built-in cloud models, and no model in the local catalogue is a rank model. openai, anthropic and google are the three providers, and none of them offers a reranking endpoint through Strata.

What to do instead

Retrieve more and cut. Ask for more neighbours than you need and apply your own relevance rule. Scores are comparable within one query, so a cutoff works even though it needs calibrating to your corpus.

strata ./docs vector query docs --text "..." -k 20

Filter rather than rerank. Metadata filtering runs before scoring, so narrowing by document type or recency often does what reranking was going to do, at no extra cost. Build semantic search has the filter shape.

Let the generator judge. Pass the top handful into inference generate and ask it to pick, which costs one call and is usually good enough for small candidate sets.

Rerank outside Strata. A cross-encoder in your own process, with Strata as the retrieval layer.

When this changes

The surface is already there, so this becomes a one-line change when a provider supports it: inference rank takes a query and passages and returns them scored. capability is the thing to check, and it reports per model, so a future model that can rank will say so without this page needing to be right.

Next

Build semantic search for the retrieval half, or Build RAG for the generation half.