Quest arg generation should always use gpt-4

src/generators/server_settings_field_generators/adventure_fixed_quest_arc_generator.py

@@ -8,6 +8,7 @@
 from generators.server_settings_field_generator import ServerSettingsFieldGenerator
 from generators.utils import safe_format
 from schema.quest import QuestDescription
+from utils.context_utils import get_quest_arc_generator
 
 
 class AdventureFixedQuestArcGenerator(ServerSettingsFieldGenerator):
@@ -54,6 +55,11 @@ def inner_generate(
         context: AgentContext,
         generation_config: Optional[dict] = None,
     ) -> Block:
+
+        # NOTE: We use a specific generator for this because it's important it's gpt-4
+        # otherwise the prompts have been a bit too shaky.
+        generator = get_quest_arc_generator(context)
+
         prompt = self.prompt
 
         task = generator.generate(

src/schema/server_settings.py

@@ -229,7 +229,7 @@ class ServerSettings(BaseModel):
     """
 
     # Language Generation Settings - Function calling
-    default_function_capable_llm_model: str = Field("gpt-3.5-turbo", description="")
+    default_function_capable_llm_model: str = Field("gpt-4", description="")
     default_function_capable_llm_temperature: float = Field(0.0, description="")
     default_function_capable_llm_max_tokens: int = Field(512, description="")
 

src/utils/context_utils.py

@@ -21,7 +21,6 @@
 from typing import List, Optional, Union
 
 from steamship import Block, PluginInstance
-from steamship.agents.llms.openai import ChatOpenAI
 from steamship.agents.logging import AgentLogging
 from steamship.agents.schema import ChatHistory, ChatLLM, FinishAction
 from steamship.agents.schema.agent import AgentContext
@@ -38,6 +37,7 @@
 )
 _BACKGROUND_MUSIC_GENERATOR_KEY = "background-music-generator"
 _NARRATION_GENERATOR_KEY = "narration-generator"
+_QUEST_GENERATOR_KEY = "quest-arc-generator"
 _SERVER_SETTINGS_KEY = "server-settings"
 _GAME_STATE_KEY = "user-settings"
 
@@ -101,6 +101,33 @@ def get_story_text_generator(
     return generator
 
 
+def get_quest_arc_generator(
+    context: AgentContext, default: Optional[PluginInstance] = None
+) -> Optional[PluginInstance]:
+    generator = context.metadata.get(_QUEST_GENERATOR_KEY, default)
+
+    if not generator:
+        # Lazily create
+        server_settings: ServerSettings = get_server_settings(context)
+
+        # Hard coded.
+        model_name = "gpt-4"
+        plugin_handle = "gpt-4"
+
+        generator = context.client.use_plugin(
+            plugin_handle,
+            config={
+                "model": model_name,
+                "max_tokens": server_settings.default_story_max_tokens,
+                "temperature": server_settings.default_story_temperature,
+            },
+        )
+
+        context.metadata[_QUEST_GENERATOR_KEY] = generator
+
+    return generator
+
+
 def get_background_music_generator(
     context: AgentContext, default: Optional[PluginInstance] = None
 ) -> Optional[PluginInstance]:
@@ -343,14 +370,38 @@ def switch_history_to_current_quest(
 
 
 def get_function_capable_llm(
-    context: AgentContext, default: Optional[ChatLLM] = None  # noqa: F821
-) -> Optional[ChatLLM]:  # noqa: F821
-    llm = context.metadata.get(_FUNCTION_CAPABLE_LLM, default)
-    if not llm:
-        # Lazy create
-        llm = ChatOpenAI(context.client)
-        context.metadata[_FUNCTION_CAPABLE_LLM] = llm
-    return llm
+    context: AgentContext, default: Optional[PluginInstance] = None
+) -> Optional[PluginInstance]:
+    generator = context.metadata.get(_FUNCTION_CAPABLE_LLM, default)
+
+    if not generator:
+        # Lazily create
+        server_settings: ServerSettings = get_server_settings(context)
+
+        open_ai_models = ["gpt-3.5-turbo", "gpt-4-1106-preview", "gpt-4"]
+
+        model_name = server_settings._select_model(
+            open_ai_models,
+            default=server_settings.default_function_capable_llm_model,
+            preferred="gpt-4",
+        )
+
+        plugin_handle = None
+        if model_name in open_ai_models:
+            plugin_handle = "gpt-4"
+
+        generator = context.client.use_plugin(
+            plugin_handle,
+            config={
+                "model": model_name,
+                "max_tokens": server_settings.default_function_capable_llm_max_tokens,
+                "temperature": server_settings.default_function_capable_llm_temperature,
+            },
+        )
+
+        context.metadata[_FUNCTION_CAPABLE_LLM] = generator
+
+    return generator
 
 
 def _key_for_question(blocks: List[Block], key: Optional[str] = None) -> str: