[Improvement]Vector search similarity threshold is too high in the search API

[earayu]

The current similarity threshold for vector search is set to 0.7, which is too high and often results in recalling zero documents.

I suggest either removing the similarity filter or significantly lowering the threshold.

A better approach would be to let the reranker module handle the final selection by picking the topK results from a larger pool of candidates. This would improve the document recall rate.

The search interface should call a rerank model by default to improve result relevance.

If no specific rerank model is configured, a default rerank strategy should be applied.