Introduce hermesllm library to handle llm message translation (#501)

crates/common/Cargo.toml

@@ -18,6 +18,7 @@ serde_json = "1.0"
 hex = "0.4.3"
 urlencoding = "2.1.3"
 url = "2.5.4"
+hermesllm = { version = "0.1.0", path = "../hermesllm" }
 
 [dev-dependencies]
 pretty_assertions = "1.4.1"

crates/common/src/configuration.rs

@@ -1,3 +1,4 @@
+use hermesllm::providers::openai::types::{ModelDetail, ModelObject, Models};
 use serde::{Deserialize, Serialize};
 use std::collections::HashMap;
 use std::fmt::Display;
@@ -206,6 +207,29 @@ pub struct LlmProvider {
     pub usage: Option<String>,
 }
 
+pub trait IntoModels {
+    fn into_models(self) -> Models;
+}
+
+impl IntoModels for Vec<LlmProvider> {
+    fn into_models(self) -> Models {
+        let data = self
+            .iter()
+            .map(|provider| ModelDetail {
+                id: provider.name.clone(),
+                object: "model".to_string(),
+                created: 0,
+                owned_by: "system".to_string(),
+            })
+            .collect();
+
+        Models {
+            object: ModelObject::List,
+            data,
+        }
+    }
+}
+
 impl Default for LlmProvider {
     fn default() -> Self {
         Self {

crates/common/src/errors.rs

@@ -1,6 +1,7 @@
 use proxy_wasm::types::Status;
 
 use crate::{api::open_ai::ChatCompletionChunkResponseError, ratelimit};
+use hermesllm::providers::openai::types::OpenAIError;
 
 #[derive(thiserror::Error, Debug)]
 pub enum ClientError {
@@ -39,4 +40,6 @@ pub enum ServerError {
     BadRequest { why: String },
     #[error("error in streaming response")]
     Streaming(#[from] ChatCompletionChunkResponseError),
+    #[error("error parsing openai message: {0}")]
+    OpenAIPError(#[from] OpenAIError),
 }

crates/common/src/tokenizer.rs

@@ -14,7 +14,7 @@ pub fn token_count(model_name: &str, text: &str) -> Result<usize, String> {
             );
             "gpt-4"
         }
-        true => model_name
+        true => model_name,
     };
 
     // Consideration: is it more expensive to instantiate the BPE object every time, or to contend the singleton?