[router] Add Rerank Routing Logic in Regular Router (#10219)
This commit is contained in:
@@ -1891,7 +1891,7 @@ pub struct RerankResponse {
|
||||
pub object: String,
|
||||
|
||||
/// Response ID
|
||||
pub id: String,
|
||||
pub id: Option<StringOrArray>,
|
||||
|
||||
/// Creation timestamp
|
||||
pub created: i64,
|
||||
@@ -1976,7 +1976,11 @@ impl RerankRequest {
|
||||
}
|
||||
|
||||
impl RerankResponse {
|
||||
pub fn new(results: Vec<RerankResult>, model: String, request_id: String) -> Self {
|
||||
pub fn new(
|
||||
results: Vec<RerankResult>,
|
||||
model: String,
|
||||
request_id: Option<StringOrArray>,
|
||||
) -> Self {
|
||||
RerankResponse {
|
||||
results,
|
||||
model,
|
||||
@@ -2000,6 +2004,13 @@ impl RerankResponse {
|
||||
pub fn apply_top_k(&mut self, k: usize) {
|
||||
self.results.truncate(k);
|
||||
}
|
||||
|
||||
/// Drop documents from results
|
||||
pub fn drop_documents(&mut self) {
|
||||
self.results.iter_mut().for_each(|result| {
|
||||
result.document = None;
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
// ==================================================================
|
||||
@@ -2268,12 +2279,15 @@ mod tests {
|
||||
let response = RerankResponse::new(
|
||||
results.clone(),
|
||||
"test-model".to_string(),
|
||||
"req-123".to_string(),
|
||||
Some(StringOrArray::String("req-123".to_string())),
|
||||
);
|
||||
|
||||
assert_eq!(response.results.len(), 2);
|
||||
assert_eq!(response.model, "test-model");
|
||||
assert_eq!(response.id, "req-123");
|
||||
assert_eq!(
|
||||
response.id,
|
||||
Some(StringOrArray::String("req-123".to_string()))
|
||||
);
|
||||
assert_eq!(response.object, "rerank");
|
||||
assert!(response.created > 0);
|
||||
}
|
||||
@@ -2287,8 +2301,11 @@ mod tests {
|
||||
meta_info: None,
|
||||
}];
|
||||
|
||||
let response =
|
||||
RerankResponse::new(results, "test-model".to_string(), "req-123".to_string());
|
||||
let response = RerankResponse::new(
|
||||
results,
|
||||
"test-model".to_string(),
|
||||
Some(StringOrArray::String("req-123".to_string())),
|
||||
);
|
||||
|
||||
let serialized = serde_json::to_string(&response).unwrap();
|
||||
let deserialized: RerankResponse = serde_json::from_str(&serialized).unwrap();
|
||||
@@ -2322,8 +2339,11 @@ mod tests {
|
||||
},
|
||||
];
|
||||
|
||||
let mut response =
|
||||
RerankResponse::new(results, "test-model".to_string(), "req-123".to_string());
|
||||
let mut response = RerankResponse::new(
|
||||
results,
|
||||
"test-model".to_string(),
|
||||
Some(StringOrArray::String("req-123".to_string())),
|
||||
);
|
||||
|
||||
response.sort_by_score();
|
||||
|
||||
@@ -2358,8 +2378,11 @@ mod tests {
|
||||
},
|
||||
];
|
||||
|
||||
let mut response =
|
||||
RerankResponse::new(results, "test-model".to_string(), "req-123".to_string());
|
||||
let mut response = RerankResponse::new(
|
||||
results,
|
||||
"test-model".to_string(),
|
||||
Some(StringOrArray::String("req-123".to_string())),
|
||||
);
|
||||
|
||||
response.apply_top_k(2);
|
||||
|
||||
@@ -2377,14 +2400,36 @@ mod tests {
|
||||
meta_info: None,
|
||||
}];
|
||||
|
||||
let mut response =
|
||||
RerankResponse::new(results, "test-model".to_string(), "req-123".to_string());
|
||||
let mut response = RerankResponse::new(
|
||||
results,
|
||||
"test-model".to_string(),
|
||||
Some(StringOrArray::String("req-123".to_string())),
|
||||
);
|
||||
|
||||
response.apply_top_k(5);
|
||||
|
||||
assert_eq!(response.results.len(), 1);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn test_rerank_response_drop_documents() {
|
||||
let results = vec![RerankResult {
|
||||
score: 0.8,
|
||||
document: Some("doc1".to_string()),
|
||||
index: 0,
|
||||
meta_info: None,
|
||||
}];
|
||||
let mut response = RerankResponse::new(
|
||||
results,
|
||||
"test-model".to_string(),
|
||||
Some(StringOrArray::String("req-123".to_string())),
|
||||
);
|
||||
|
||||
response.drop_documents();
|
||||
|
||||
assert_eq!(response.results[0].document, None);
|
||||
}
|
||||
|
||||
// ==================================================================
|
||||
// = RERANK RESULT TESTS =
|
||||
// ==================================================================
|
||||
@@ -2570,8 +2615,11 @@ mod tests {
|
||||
meta_info: None,
|
||||
}];
|
||||
|
||||
let mut response =
|
||||
RerankResponse::new(results, "test-model".to_string(), "req-123".to_string());
|
||||
let mut response = RerankResponse::new(
|
||||
results,
|
||||
"test-model".to_string(),
|
||||
Some(StringOrArray::String("req-123".to_string())),
|
||||
);
|
||||
|
||||
response.usage = Some(UsageInfo {
|
||||
prompt_tokens: 100,
|
||||
@@ -2645,18 +2693,7 @@ mod tests {
|
||||
];
|
||||
|
||||
// Create response
|
||||
let mut response = RerankResponse::new(
|
||||
results,
|
||||
request.model.clone(),
|
||||
request
|
||||
.rid
|
||||
.as_ref()
|
||||
.and_then(|r| match r {
|
||||
StringOrArray::String(s) => Some(s.clone()),
|
||||
StringOrArray::Array(arr) => arr.first().cloned(),
|
||||
})
|
||||
.unwrap_or_else(|| "unknown".to_string()),
|
||||
);
|
||||
let mut response = RerankResponse::new(results, request.model.clone(), request.rid.clone());
|
||||
|
||||
// Sort by score
|
||||
response.sort_by_score();
|
||||
|
||||
@@ -301,7 +301,11 @@ impl RouterTrait for GrpcPDRouter {
|
||||
(StatusCode::NOT_IMPLEMENTED).into_response()
|
||||
}
|
||||
|
||||
async fn route_rerank(&self, _headers: Option<&HeaderMap>, _body: Body) -> Response {
|
||||
async fn route_rerank(
|
||||
&self,
|
||||
_headers: Option<&HeaderMap>,
|
||||
_body: &crate::protocols::spec::RerankRequest,
|
||||
) -> Response {
|
||||
(StatusCode::NOT_IMPLEMENTED).into_response()
|
||||
}
|
||||
|
||||
|
||||
@@ -234,7 +234,11 @@ impl RouterTrait for GrpcRouter {
|
||||
(StatusCode::NOT_IMPLEMENTED).into_response()
|
||||
}
|
||||
|
||||
async fn route_rerank(&self, _headers: Option<&HeaderMap>, _body: Body) -> Response {
|
||||
async fn route_rerank(
|
||||
&self,
|
||||
_headers: Option<&HeaderMap>,
|
||||
_body: &crate::protocols::spec::RerankRequest,
|
||||
) -> Response {
|
||||
(StatusCode::NOT_IMPLEMENTED).into_response()
|
||||
}
|
||||
|
||||
|
||||
@@ -2,7 +2,9 @@
|
||||
|
||||
use crate::config::CircuitBreakerConfig;
|
||||
use crate::core::{CircuitBreaker, CircuitBreakerConfig as CoreCircuitBreakerConfig};
|
||||
use crate::protocols::spec::{ChatCompletionRequest, CompletionRequest, GenerateRequest};
|
||||
use crate::protocols::spec::{
|
||||
ChatCompletionRequest, CompletionRequest, GenerateRequest, RerankRequest,
|
||||
};
|
||||
use async_trait::async_trait;
|
||||
use axum::{
|
||||
body::Body,
|
||||
@@ -381,7 +383,7 @@ impl super::super::RouterTrait for OpenAIRouter {
|
||||
.into_response()
|
||||
}
|
||||
|
||||
async fn route_rerank(&self, _headers: Option<&HeaderMap>, _body: Body) -> Response {
|
||||
async fn route_rerank(&self, _headers: Option<&HeaderMap>, _body: &RerankRequest) -> Response {
|
||||
(
|
||||
StatusCode::NOT_IMPLEMENTED,
|
||||
"Rerank endpoint not implemented for OpenAI backend",
|
||||
|
||||
@@ -9,8 +9,8 @@ use crate::core::{
|
||||
use crate::metrics::RouterMetrics;
|
||||
use crate::policies::LoadBalancingPolicy;
|
||||
use crate::protocols::spec::{
|
||||
ChatCompletionRequest, ChatMessage, CompletionRequest, GenerateRequest, ResponsesRequest,
|
||||
StringOrArray, UserMessageContent,
|
||||
ChatCompletionRequest, ChatMessage, CompletionRequest, GenerateRequest, RerankRequest,
|
||||
ResponsesRequest, StringOrArray, UserMessageContent,
|
||||
};
|
||||
use crate::routers::header_utils;
|
||||
use crate::routers::{RouterTrait, WorkerManagement};
|
||||
@@ -1946,8 +1946,25 @@ impl RouterTrait for PDRouter {
|
||||
todo!()
|
||||
}
|
||||
|
||||
async fn route_rerank(&self, _headers: Option<&HeaderMap>, _body: Body) -> Response {
|
||||
todo!()
|
||||
async fn route_rerank(&self, headers: Option<&HeaderMap>, body: &RerankRequest) -> Response {
|
||||
// Extract text for cache-aware routing
|
||||
let req_text = if self.policies_need_request_text() {
|
||||
Some(body.query.clone())
|
||||
} else {
|
||||
None
|
||||
};
|
||||
|
||||
// Create context
|
||||
let context = PDRequestContext {
|
||||
route: "/v1/rerank",
|
||||
batch_size: None,
|
||||
is_stream: false,
|
||||
return_logprob: false,
|
||||
request_text: req_text,
|
||||
};
|
||||
|
||||
// Execute with retry and bootstrap injection
|
||||
self.execute_dual_dispatch(headers, body, context).await
|
||||
}
|
||||
|
||||
async fn flush_cache(&self) -> Response {
|
||||
|
||||
@@ -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 {
|
||||
|
||||
@@ -10,7 +10,7 @@ use axum::{
|
||||
use std::fmt::Debug;
|
||||
|
||||
use crate::protocols::spec::{
|
||||
ChatCompletionRequest, CompletionRequest, GenerateRequest, ResponsesRequest,
|
||||
ChatCompletionRequest, CompletionRequest, GenerateRequest, RerankRequest, ResponsesRequest,
|
||||
};
|
||||
|
||||
pub mod factory;
|
||||
@@ -89,7 +89,7 @@ pub trait RouterTrait: Send + Sync + Debug + WorkerManagement {
|
||||
|
||||
async fn route_embeddings(&self, headers: Option<&HeaderMap>, body: Body) -> Response;
|
||||
|
||||
async fn route_rerank(&self, headers: Option<&HeaderMap>, body: Body) -> Response;
|
||||
async fn route_rerank(&self, headers: Option<&HeaderMap>, body: &RerankRequest) -> Response;
|
||||
|
||||
/// Flush cache on all workers
|
||||
async fn flush_cache(&self) -> Response;
|
||||
|
||||
@@ -3,7 +3,8 @@ use crate::logging::{self, LoggingConfig};
|
||||
use crate::metrics::{self, PrometheusConfig};
|
||||
use crate::middleware::TokenBucket;
|
||||
use crate::protocols::spec::{
|
||||
ChatCompletionRequest, CompletionRequest, GenerateRequest, ResponsesRequest,
|
||||
ChatCompletionRequest, CompletionRequest, GenerateRequest, RerankRequest, ResponsesRequest,
|
||||
V1RerankReqInput,
|
||||
};
|
||||
use crate::reasoning_parser::ParserFactory;
|
||||
use crate::routers::{RouterFactory, RouterTrait};
|
||||
@@ -152,6 +153,25 @@ async fn v1_completions(
|
||||
state.router.route_completion(Some(&headers), &body).await
|
||||
}
|
||||
|
||||
async fn rerank(
|
||||
State(state): State<Arc<AppState>>,
|
||||
headers: http::HeaderMap,
|
||||
Json(body): Json<RerankRequest>,
|
||||
) -> Response {
|
||||
state.router.route_rerank(Some(&headers), &body).await
|
||||
}
|
||||
|
||||
async fn v1_rerank(
|
||||
State(state): State<Arc<AppState>>,
|
||||
headers: http::HeaderMap,
|
||||
Json(body): Json<V1RerankReqInput>,
|
||||
) -> Response {
|
||||
state
|
||||
.router
|
||||
.route_rerank(Some(&headers), &body.into())
|
||||
.await
|
||||
}
|
||||
|
||||
async fn v1_responses(
|
||||
State(state): State<Arc<AppState>>,
|
||||
headers: http::HeaderMap,
|
||||
@@ -237,6 +257,8 @@ pub fn build_app(
|
||||
.route("/generate", post(generate))
|
||||
.route("/v1/chat/completions", post(v1_chat_completions))
|
||||
.route("/v1/completions", post(v1_completions))
|
||||
.route("/rerank", post(rerank))
|
||||
.route("/v1/rerank", post(v1_rerank))
|
||||
.route("/v1/responses", post(v1_responses))
|
||||
.route_layer(axum::middleware::from_fn_with_state(
|
||||
app_state.clone(),
|
||||
|
||||
Reference in New Issue
Block a user