Inverted Responses
I am trying to use jinaai/jina-reranker-v3 as a reranker in llama-server. The model loads successfully, but when sending a request to the /v1/rerank endpoint, the model outputs completely incorrect, microscopic scores, and actually scores irrelevant documents higher than perfect matches.
curl http://127.0.0.1:8282/v1/rerank ^
-H "Content-Type: application/json" ^
-d "{"model": "jina-reranker-v3-Q8_0", "query": "What is the capital of France?", "documents": ["The sky is blue.", "Paris is the capital of France."]}"
{"model":"jina-reranker-v3-Q8_0","object":"list","usage":{"prompt_tokens":26,"total_tokens":26},"results":[{"index":0,"relevance_score":0.00010479514457983896},{"index":1,"relevance_score":2.028539682896735e-07}]}
is this correct ?