feat: replace o1-preview by o1 (#9596)

Co-authored-by: Henry Fontanier <[email protected]>

core/src/providers/openai.rs

@@ -208,6 +208,7 @@ pub enum OpenAIChatMessageRole {
     Assistant,
     Function,
     System,
+    Developer,
     Tool,
     User,
 }
@@ -242,6 +243,7 @@ impl From<OpenAIChatMessageRole> for ChatMessageRole {
             OpenAIChatMessageRole::Assistant => ChatMessageRole::Assistant,
             OpenAIChatMessageRole::Function => ChatMessageRole::Function,
             OpenAIChatMessageRole::System => ChatMessageRole::System,
+            OpenAIChatMessageRole::Developer => ChatMessageRole::System,
             OpenAIChatMessageRole::Tool => ChatMessageRole::Function,
             OpenAIChatMessageRole::User => ChatMessageRole::User,
         }
@@ -1719,14 +1721,23 @@ pub fn to_openai_messages(
     messages: &Vec<ChatMessage>,
     model_id: &str,
 ) -> Result<Vec<OpenAIChatMessage>, anyhow::Error> {
-    messages
+    let mut oai_messages = messages
         .iter()
-        .filter_map(|m| match m {
-            // [o1-preview] Hack for OpenAI `o1-*` models to exclude system messages.
-            ChatMessage::System(_) if model_id.starts_with("o1-") => None,
-            _ => Some(OpenAIChatMessage::try_from(m)),
-        })
-        .collect::<Result<Vec<_>>>()
+        .map(|m| OpenAIChatMessage::try_from(m))
+        .collect::<Result<Vec<_>>>()?
+        .into_iter()
+        // [o1-mini] O1 mini does not support system messages, so we filter them out.
+        .filter(|m| m.role != OpenAIChatMessageRole::System || !model_id.starts_with("o1-mini"))
+        .collect::<Vec<_>>();
+
+    // [o1] O1 uses `developer` messages instead of `system` messages.
+    for m in oai_messages.iter_mut() {
+        if m.role == OpenAIChatMessageRole::System && model_id.starts_with("o1") {
+            m.role = OpenAIChatMessageRole::Developer;
+        }
+    }
+
+    Ok(oai_messages)
 }
 
 #[async_trait]
@@ -1795,8 +1806,8 @@ impl LLM for OpenAILLM {
             }
         }
 
-        // [o1-preview] Hack for OpenAI `o1-*` models to not use streaming.
-        let model_is_o1 = self.id.as_str().starts_with("o1-");
+        // [o1] Hack for OpenAI `o1*` models to not use streaming.
+        let model_is_o1 = self.id.as_str().starts_with("o1");
         let (c, request_id) = if !model_is_o1 && event_sender.is_some() {
             if n > 1 {
                 return Err(anyhow!(
@@ -2024,9 +2035,9 @@ impl LLM for OpenAILLM {
 
         let openai_messages = to_openai_messages(messages, &self.id)?;
 
-        // [o1-preview] Hack for OpenAI `o1-*` models to simulate streaming.
+        // [o1] Hack for OpenAI `o1*` models to simulate streaming.
         let is_streaming = event_sender.is_some();
-        let model_is_o1 = self.id.as_str().starts_with("o1-");
+        let model_is_o1 = self.id.as_str().starts_with("o1");
 
         let (c, request_id) = if !model_is_o1 && is_streaming {
             streamed_chat_completion(
@@ -2089,7 +2100,7 @@ impl LLM for OpenAILLM {
             .await?
         };
 
-        // [o1-preview] Hack for OpenAI `o1-*` models to simulate streaming.
+        // [o1] Hack for OpenAI `o1*` models to simulate streaming.
         if model_is_o1 && is_streaming {
             let sender = event_sender.as_ref().unwrap();
             for choice in &c.choices {

front/lib/api/assistant/global_agents.ts

@@ -30,7 +30,7 @@ import {
   MISTRAL_MEDIUM_MODEL_CONFIG,
   MISTRAL_SMALL_MODEL_CONFIG,
   O1_MINI_MODEL_CONFIG,
-  O1_PREVIEW_MODEL_CONFIG,
+  O1_MODEL_CONFIG,
 } from "@dust-tt/types";
 import assert from "assert";
 
@@ -326,16 +326,16 @@ function _getO1PreviewGlobalAgent({
     version: 0,
     versionCreatedAt: null,
     versionAuthorId: null,
-    name: "o1-preview",
-    description: O1_PREVIEW_MODEL_CONFIG.description,
+    name: "o1",
+    description: O1_MODEL_CONFIG.description,
     instructions: null,
     pictureUrl: "https://dust.tt/static/systemavatar/o1_avatar_full.png",
     status,
     scope: "global",
     userFavorite: false,
     model: {
-      providerId: O1_PREVIEW_MODEL_CONFIG.providerId,
-      modelId: O1_PREVIEW_MODEL_CONFIG.modelId,
+      providerId: O1_MODEL_CONFIG.providerId,
+      modelId: O1_MODEL_CONFIG.modelId,
       temperature: 1, // 1 is forced for O1
     },
     actions: [],

sdks/js/src/types.ts

@@ -23,7 +23,7 @@ const ModelLLMIdSchema = FlexibleEnumSchema<
   | "gpt-4o-2024-08-06"
   | "gpt-4o"
   | "gpt-4o-mini"
-  | "o1-preview"
+  | "o1"
   | "o1-mini"
   | "claude-3-opus-20240229"
   | "claude-3-5-sonnet-20240620"

types/src/front/lib/assistant.ts

@@ -89,7 +89,7 @@ export const GPT_4_TURBO_MODEL_ID = "gpt-4-turbo" as const;
 export const GPT_4O_MODEL_ID = "gpt-4o" as const;
 export const GPT_4O_20240806_MODEL_ID = "gpt-4o-2024-08-06" as const;
 export const GPT_4O_MINI_MODEL_ID = "gpt-4o-mini" as const;
-export const O1_PREVIEW_MODEL_ID = "o1-preview" as const;
+export const O1_MODEL_ID = "o1" as const;
 export const O1_MINI_MODEL_ID = "o1-mini" as const;
 export const CLAUDE_3_OPUS_2024029_MODEL_ID = "claude-3-opus-20240229" as const;
 export const CLAUDE_3_5_SONNET_20240620_MODEL_ID =
@@ -124,7 +124,7 @@ export const MODEL_IDS = [
   GPT_4O_MODEL_ID,
   GPT_4O_20240806_MODEL_ID,
   GPT_4O_MINI_MODEL_ID,
-  O1_PREVIEW_MODEL_ID,
+  O1_MODEL_ID,
   O1_MINI_MODEL_ID,
   CLAUDE_3_OPUS_2024029_MODEL_ID,
   CLAUDE_3_5_SONNET_20240620_MODEL_ID,
@@ -267,11 +267,11 @@ export const GPT_4O_MINI_MODEL_CONFIG: ModelConfigurationType = {
   toolUseMetaPrompt: LEGACY_OPEN_AI_TOOL_USE_META_PROMPT,
   supportsVision: true,
 };
-export const O1_PREVIEW_MODEL_CONFIG: ModelConfigurationType = {
+export const O1_MODEL_CONFIG: ModelConfigurationType = {
   providerId: "openai",
-  modelId: O1_PREVIEW_MODEL_ID,
-  displayName: "O1 Preview",
-  contextSize: 128_000,
+  modelId: O1_MODEL_ID,
+  displayName: "O1",
+  contextSize: 200_000,
   recommendedTopK: 32,
   recommendedExhaustiveTopK: 128, // 65_536
   largeModel: true,
@@ -601,7 +601,7 @@ export const SUPPORTED_MODEL_CONFIGS: ModelConfigurationType[] = [
   GPT_4O_MODEL_CONFIG,
   GPT_4O_20240806_MODEL_CONFIG,
   GPT_4O_MINI_MODEL_CONFIG,
-  O1_PREVIEW_MODEL_CONFIG,
+  O1_MODEL_CONFIG,
   O1_MINI_MODEL_CONFIG,
   CLAUDE_3_OPUS_DEFAULT_MODEL_CONFIG,
   CLAUDE_3_5_SONNET_20240620_DEPRECATED_MODEL_CONFIG,