refactor: several optimizations (#749)

.gitignore

@@ -1,3 +1,4 @@
 /target
 /tmp
+/.env
 *.log

src/client/common.rs

@@ -391,7 +391,6 @@ pub trait Client: Sync + Send {
 
     async fn embeddings(&self, data: EmbeddingsData) -> Result<Vec<Vec<f32>>> {
         let client = self.build_client()?;
-        self.model().guard_max_batch_size(&data)?;
         self.embeddings_inner(&client, data)
             .await
             .context("Failed to call embeddings api")

src/client/model.rs

@@ -1,7 +1,6 @@
 use super::{
     list_chat_models, list_embedding_models, list_reranker_models,
     message::{Message, MessageContent},
-    EmbeddingsData,
 };
 
 use crate::config::Config;
@@ -254,13 +253,6 @@ impl Model {
         }
         Ok(())
     }
-
-    pub fn guard_max_batch_size(&self, data: &EmbeddingsData) -> Result<()> {
-        if data.texts.len() > self.max_batch_size() {
-            bail!("Exceed max_batch_size limit");
-        }
-        Ok(())
-    }
 }
 
 #[derive(Debug, Clone, Default, Serialize, Deserialize)]

src/serve.rs

@@ -163,7 +163,7 @@ impl Server {
             self.arena_page()
         } else {
             status = StatusCode::NOT_FOUND;
-            Err(anyhow!("The requested endpoint was not found."))
+            Err(anyhow!("Not Found"))
         };
         let mut res = match res {
             Ok(res) => {
@@ -171,7 +171,9 @@ impl Server {
                 res
             }
             Err(err) => {
-                status = StatusCode::BAD_REQUEST;
+                if status == StatusCode::OK {
+                    status = StatusCode::BAD_REQUEST;
+                }
                 error!("{method} {uri} {} {err}", status.as_u16());
                 ret_err(err)
             }