jina-reranker-m0 - a reranking-only provider for a self-hosted server.

provider = ExAgent.Providers.JinaRerankerM0.new(base_url: System.fetch_env!("RERANKER_URL"))

{:ok, result} = ExAgent.rerank(provider, "who restarts processes?", chunks, top_n: 5)

best = ExAgent.Reranking.take(result, chunks)

It implements ExAgent.Provider.rerank/4 only. chat/3 returns {:error, %ExAgent.Error{type: :unsupported}}, and there is no embed/3 - a reranker scores query/document pairs and has no single-text vector to give.

Where it belongs in retrieval

Reranking is the second stage, not a replacement for the first:

{:ok, hits} = ExAgent.embed(embedder, [query], task: :retrieval, args: [prompt_name: :query])
shortlist = vector_search(hits, limit: 100)

{:ok, ranked} = ExAgent.rerank(reranker, query, shortlist, top_n: 10)

Embeddings compare vectors computed independently, which is what makes them fast enough for a corpus. A cross-encoder reads the query and one document together - more accurate, and far too slow to run over everything. Batches are capped at 512 documents, which is the shape of the intended use.

Behind Modal's proxy auth

ExAgent.Providers.JinaRerankerM0.new(
  base_url: System.fetch_env!("MODAL_RERANKER_URL"),
  api_key: System.fetch_env!("MODAL_API_KEY"),
  headers: [
    {"Modal-Key", System.fetch_env!("MODAL_KEY")},
    {"Modal-Secret", System.fetch_env!("MODAL_SECRET")}
  ]
)

Scores

Higher is more relevant, and that is the only guarantee - see ExAgent.Reranking on why an absolute cutoff does not port between models.

The server contract

Verified against a running deployment. POST {base_url}/v1/rerank:

{"query": "...", "documents": ["..."], "top_n": 5, "return_documents": false}

answering

{"model": "...",
 "results": [{"index": 0, "relevance_score": 0.28, "document": {"text": "..."}}]}

Unknown body fields are rejected by the server, so unrecognized options are reported rather than forwarded. return_documents defaults to true there and false here: :index already identifies each document, and sending a corpus back to have it re-quoted is waste.

This is not the shape of Jina's hosted api.jina.ai/v1/rerank service, whose documents accept {"text": ...} / {"image": ...} objects for multimodal ranking. Pointing this provider there will not work.