Merge branch 'main' into debug-streaming-v2

crates/common/src/consts.rs

@@ -7,10 +7,13 @@ pub const SYSTEM_ROLE: &str = "system";
 pub const USER_ROLE: &str = "user";
 pub const TOOL_ROLE: &str = "tool";
 pub const ASSISTANT_ROLE: &str = "assistant";
-pub const GPT_35_TURBO: &str = "gpt-3.5-turbo";
-pub const ARC_FC_CLUSTER: &str = "arch_fc";
 pub const ARCH_FC_REQUEST_TIMEOUT_MS: u64 = 120000; // 2 minutes
 pub const MODEL_SERVER_NAME: &str = "model_server";
+pub const ZEROSHOT_INTERNAL_HOST: &str = "zeroshot";
+pub const ARCH_FC_INTERNAL_HOST: &str = "arch_fc";
+pub const HALLUCINATION_INTERNAL_HOST: &str = "hallucination";
+pub const EMBEDDINGS_INTERNAL_HOST: &str = "embeddings";
+pub const GUARD_INTERNAL_HOST: &str = "guard";
 pub const ARCH_ROUTING_HEADER: &str = "x-arch-llm-provider";
 pub const ARCH_MESSAGES_KEY: &str = "arch_messages";
 pub const ARCH_PROVIDER_HINT_HEADER: &str = "x-arch-llm-provider-hint";

crates/common/src/errors.rs

@@ -23,11 +23,12 @@ pub enum ServerError {
     Serialization(serde_json::Error),
     #[error("{0}")]
     LogicError(String),
-    #[error("upstream error response authority={authority}, path={path}, status={status}")]
+    #[error("upstream application error host={host}, path={path}, status={status}, body={body}")]
     Upstream {
-        authority: String,
+        host: String,
         path: String,
         status: String,
+        body: String,
     },
     #[error("jailbreak detected: {0}")]
     Jailbreak(String),

crates/common/src/http.rs

@@ -1,6 +1,9 @@
-use crate::{errors::ClientError, stats::{Gauge, IncrementingMetric}};
+use crate::{
+    errors::ClientError,
+    stats::{Gauge, IncrementingMetric},
+};
 use derivative::Derivative;
-use log::debug;
+use log::{debug, trace};
 use proxy_wasm::{traits::Context, types::Status};
 use serde::Serialize;
 use std::{cell::RefCell, collections::HashMap, fmt::Debug, time::Duration};
@@ -45,9 +48,10 @@ pub trait Client: Context {
         call_args: CallArgs,
         call_context: Self::CallContext,
     ) -> Result<u32, ClientError> {
-        debug!(
+        trace!(
             "dispatching http call with args={:?} context={:?}",
-            call_args, call_context
+            call_args,
+            call_context
         );
 
         match self.dispatch_http_call(

crates/common/src/lib.rs

@@ -4,10 +4,10 @@ pub mod common_types;
 pub mod configuration;
 pub mod consts;
 pub mod embeddings;
+pub mod errors;
 pub mod http;
 pub mod llm_providers;
 pub mod ratelimit;
 pub mod routing;
 pub mod stats;
 pub mod tokenizer;
-pub mod errors;