Skip to content
Osarai
Get started

DocsAPI reference

Rerank and score

POST /v1/rerank and /v1/score — order documents by how well they answer a query with Qwen3-Reranker-8B on your own GPU.

HTTP
POST https://osarai.xyz/v1/rerank
POST https://osarai.xyz/v1/score

A reranker reads a query together with each candidate document and scores how well the document answers it. It is slower than embeddings but far more precise, so the usual pattern is: fetch 50–100 candidates with embeddings, then rerank them. Deploy Qwen3-Reranker-8B (qwen3-reranker-8b) to use these endpoints.

Rerank

Terminal
curl "https://osarai.xyz/v1/rerank" \
  -H "Authorization: Bearer $OSARAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-reranker-8b",
    "query": "How do I reset my password?",
    "documents": [
      "Open Settings, choose Security and press Reset password.",
      "Our office is closed on public holidays.",
      "Passwords must be at least 12 characters long."
    ],
    "top_n": 2
  }'
FieldNotes
modelqwen3-reranker-8b
queryThe search query
documentsThe candidate texts
top_nReturn only the best N (default: all)
JSON
{
  "model": "qwen3-reranker-8b",
  "results": [
    { "index": 0, "document": { "text": "Open Settings, choose Security and press Reset password." }, "relevance_score": 0.97 },
    { "index": 2, "document": { "text": "Passwords must be at least 12 characters long." }, "relevance_score": 0.31 }
  ],
  "usage": { "total_tokens": 64 }
}

results are sorted best first; index points into your documents. The format matches Jina's and Cohere's rerank APIs, so their clients work with the base URL changed.

Score

/v1/score scores pairs without sorting: one query against many texts, or pairs side by side.

Terminal
curl "https://osarai.xyz/v1/score" \
  -H "Authorization: Bearer $OSARAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-reranker-8b",
    "text_1": "How do I reset my password?",
    "text_2": ["Open Settings, choose Security and press Reset password.", "Our office is closed on public holidays."]
  }'
JSON
{
  "object": "list",
  "model": "qwen3-reranker-8b",
  "data": [
    { "index": 0, "object": "score", "score": 0.97 },
    { "index": 1, "object": "score", "score": 0.02 }
  ]
}

Scores range from 0 to 1. Request bodies can be up to 8 MiB, and each query–document pair must fit the model's 16K-token window.