fix: streaming

src/api/routes/v1/models.rs

@@ -13,7 +13,7 @@ use axum::{
         (
             status = 200,
             description = "List of locally available Ollama models",
-            body = api::types::ModelsResponse,
+            body = api::types::ApiModelsResponse,
             content_type = "application/json",
         ),
         (
@@ -27,7 +27,7 @@ use axum::{
 // #[axum::debug_handler]
 pub async fn list_models(
     State(state): State<SharedState>,
-) -> Result<Json<api::types::ModelsResponse>, api::errors::ApiError> {
+) -> Result<Json<api::types::ApiModelsResponse>, api::errors::ApiError> {
     let models = state.chat_service.list_models().await?;
 
     Ok(Json(models.into()))
@@ -41,7 +41,7 @@ pub async fn list_models(
         ("model" = String, Path, description = "Name of the model to load into memory (e.g. 'llama3')")
     ),
     request_body(
-        content = api::types::LoadModelRequest,
+        content = api::types::ApiLoadModelRequest,
         description = "Load model request",
         content_type = "application/json",
         example = json!({ "keep_alive": "10m" })
@@ -50,7 +50,7 @@ pub async fn list_models(
         (
             status = 200,
             description = "Model successfully loaded into memory",
-            body = api::types::LoadModelResponse,
+            body = api::types::ApiLoadModelResponse,
             content_type = "application/json",
         ),
         (
@@ -79,8 +79,8 @@ pub async fn list_models(
 pub async fn load_model(
     State(state): State<SharedState>,
     Path(model): Path<String>,
-    Json(body): Json<api::types::LoadModelRequest>,
-) -> Result<Json<api::types::LoadModelResponse>, api::errors::ApiError> {
+    Json(body): Json<api::types::ApiLoadModelRequest>,
+) -> Result<Json<api::types::ApiLoadModelResponse>, api::errors::ApiError> {
     let response = state
         .chat_service
         .load_model(crate::core::llm::models::LoadModelRequest {
@@ -89,7 +89,7 @@ pub async fn load_model(
         })
         .await?;
 
-    Ok(Json(api::types::LoadModelResponse {
+    Ok(Json(api::types::ApiLoadModelResponse {
         model: response.model,
         keep_alive: body.keep_alive,
         status: "loaded".to_string(),