[router] Add Rerank Routing Logic in Regular Router (#10219)
This commit is contained in:
@@ -6,10 +6,12 @@ use crate::core::{
|
||||
use crate::metrics::RouterMetrics;
|
||||
use crate::policies::LoadBalancingPolicy;
|
||||
use crate::protocols::spec::{
|
||||
ChatCompletionRequest, CompletionRequest, GenerateRequest, GenerationRequest, ResponsesRequest,
|
||||
ChatCompletionRequest, CompletionRequest, GenerateRequest, GenerationRequest, RerankRequest,
|
||||
RerankResponse, RerankResult, ResponsesRequest,
|
||||
};
|
||||
use crate::routers::header_utils;
|
||||
use crate::routers::{RouterTrait, WorkerManagement};
|
||||
use axum::body::to_bytes;
|
||||
use axum::{
|
||||
body::Body,
|
||||
extract::Request,
|
||||
@@ -1124,6 +1126,25 @@ impl Router {
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
async fn build_rerank_response(
|
||||
req: &RerankRequest,
|
||||
response: Response,
|
||||
) -> anyhow::Result<Response> {
|
||||
let (_, response_body) = response.into_parts();
|
||||
let body_bytes = to_bytes(response_body, usize::MAX).await?;
|
||||
let rerank_results = serde_json::from_slice::<Vec<RerankResult>>(&body_bytes)?;
|
||||
let mut rerank_response =
|
||||
RerankResponse::new(rerank_results, req.model.clone(), req.rid.clone());
|
||||
rerank_response.sort_by_score();
|
||||
if let Some(top_k) = req.top_k {
|
||||
rerank_response.apply_top_k(top_k);
|
||||
}
|
||||
if !req.return_documents {
|
||||
rerank_response.drop_documents();
|
||||
}
|
||||
Ok(Json(rerank_response).into_response())
|
||||
}
|
||||
}
|
||||
|
||||
use async_trait::async_trait;
|
||||
@@ -1223,8 +1244,26 @@ impl RouterTrait for Router {
|
||||
todo!()
|
||||
}
|
||||
|
||||
async fn route_rerank(&self, _headers: Option<&HeaderMap>, _body: Body) -> Response {
|
||||
todo!()
|
||||
async fn route_rerank(&self, headers: Option<&HeaderMap>, body: &RerankRequest) -> Response {
|
||||
if let Err(e) = body.validate() {
|
||||
return (StatusCode::BAD_REQUEST, e).into_response();
|
||||
}
|
||||
let response = self.route_typed_request(headers, body, "/v1/rerank").await;
|
||||
if response.status().is_success() {
|
||||
match Self::build_rerank_response(body, response).await {
|
||||
Ok(rerank_response) => rerank_response,
|
||||
Err(e) => {
|
||||
error!("Failed to build rerank response: {}", e);
|
||||
return (
|
||||
StatusCode::INTERNAL_SERVER_ERROR,
|
||||
"Failed to build rerank response".to_string(),
|
||||
)
|
||||
.into_response();
|
||||
}
|
||||
}
|
||||
} else {
|
||||
response
|
||||
}
|
||||
}
|
||||
|
||||
async fn flush_cache(&self) -> Response {
|
||||
|
||||
Reference in New Issue
Block a user