Merge pull request #87 from neph1/update-v0.31.1

llama3 "template"

Author: neph1

llm_config.yaml

@@ -16,7 +16,7 @@ DUNGEON_LOCATION_TEMPLATE: '{"index": (int), "name": "", "description": 25 words
 CHARACTER_TEMPLATE: '{"name":"", "description": "50 words", "appearance": "25 words", "personality": "50 words", "money":(int), "level":"", "gender":"m/f/n", "age":(int), "race":""}'
 FOLLOW_TEMPLATE: '{{"response":"yes or no", "reason":"50 words"}}'
 ITEM_TYPES: ["Weapon", "Wearable", "Health", "Money", "Trash", "Food", "Drink", "Key"]
-PRE_PROMPT: 'You are a creative game keeper for a role playing game (RPG). You craft detailed worlds and interesting characters with unique and deep personalities for the player to interact with. Do not acknowledge the task, just perform it.'
+PRE_PROMPT: 'You are a creative game keeper for a role playing game (RPG). You craft detailed worlds and interesting characters with unique and deep personalities for the player to interact with. Do not acknowledge the task or speak directly to the user, just perform it.'
 BASE_PROMPT: '<context>{context}</context>\n[USER_START]Rewrite [{input_text}] in your own words using the information found inside the <context> tags to create a background for your text. Use about {max_words} words.'
 DIALOGUE_PROMPT: '<context>{context}</context>\nThe following is a conversation between {character1} and {character2}; {character2}s sentiment towards {character1}: {sentiment}. Write a single response as {character2} in third person pov, using {character2} description and other information found inside the <context> tags. If {character2} has a quest active, they will discuss it based on its status. Respond in JSON using this template: """{dialogue_template}""". [USER_START]Continue the following conversation as {character2}: {previous_conversation}'
 COMBAT_PROMPT: '<context>{context}</context>\nThe following is a combat scene between {attackers} and {defenders} in {location}. [USER_START] Describe the following combat result in about 150 words in vivid language, using the characters weapons and their health status: 1.0 is highest, 0.0 is lowest. Combat Result: {input_text}'

tale/llm/LivingNpc.py

@@ -303,7 +303,8 @@ def _defer_result(self, action: str, verb: str="idle-action"):
         self.tell_action_deferred(verb)
 
     def tell_action_deferred(self, verb: str):
-        actions = '\n'.join(self.deferred_actions) + '\n'
+        actions = '\n'.join(self.deferred_actions) + '\n\n'
+        actions = actions.replace('\n\n\n', '\n\n')
         deferred_action = ParseResult(verb=verb, unparsed=actions, who_info=None)
         self.tell_others(actions)
         self.location._notify_action_all(deferred_action, actor=self)

tale/llm/io_adapters.py

@@ -13,11 +13,13 @@
 
 class AbstractIoAdapter(ABC):
 
-    def __init__(self, url: str, stream_endpoint: str, user_start_prompt: str, user_end_prompt: str):
+    def __init__(self, url: str, stream_endpoint: str, user_start_prompt: str, user_end_prompt: str, system_start_prompt: str = '', prompt_end: str = ''):
         self.url = url
         self.stream_endpoint = stream_endpoint
+        self.system_start_prompt = system_start_prompt
         self.user_start_prompt = user_start_prompt
         self.user_end_prompt = user_end_prompt
+        self.prompt_end = prompt_end
 
     @abstractmethod
     def stream_request(self, request_body: dict, io = None, wait: bool = False) -> str:
@@ -37,9 +39,7 @@ def set_prompt(self, request_body: dict, prompt: str, context: str = '') -> dict
 
 class KoboldCppAdapter(AbstractIoAdapter):
 
-
-
-    def __init__(self, url: str, stream_endpoint: str, data_endpoint: str, user_start_prompt: str, user_end_prompt: str):
+    def __init__(self, url: str, stream_endpoint: str, data_endpoint: str, user_start_prompt: str, user_end_prompt: str, system_start_prompt: str = '', prompt_end: str = ''):
         super().__init__(url, stream_endpoint, user_start_prompt, user_end_prompt)
         self.data_endpoint = data_endpoint
         self.place_context_in_memory = False
@@ -87,6 +87,8 @@ def parse_result(self, result: str) -> str:
         return json.loads(result)['results'][0]['text']
 
     def set_prompt(self, request_body: dict, prompt: str, context: str = '') -> dict:
+        if self.system_start_prompt:
+            prompt = self.system_start_prompt + prompt
         if self.user_start_prompt:
             prompt = prompt.replace('[USER_START]', self.user_start_prompt)
         if self.user_end_prompt:
@@ -96,6 +98,8 @@ def set_prompt(self, request_body: dict, prompt: str, context: str = '') -> dict
             request_body['memory'] = f'<context>{context}</context>'
         else:
             prompt = prompt.replace('<context>{context}</context>', f'<context>{context}</context>')
+        if self.prompt_end:
+            prompt = prompt + self.prompt_end
         request_body['prompt'] = prompt
         return request_body
 
@@ -143,12 +147,16 @@ def parse_result(self, result: str) -> str:
             raise LlmResponseException("Error parsing result from backend")
 
     def set_prompt(self, request_body: dict, prompt: str, context: str = '') -> dict:
+        if self.system_start_prompt:
+            prompt = self.system_start_prompt + prompt
         if self.user_start_prompt:
             prompt = prompt.replace('[USER_START]', self.user_start_prompt)
         if self.user_end_prompt:
             prompt = prompt + self.user_end_prompt
         if context:
             prompt = prompt.replace('<context>{context}</context>', f'<context>{context}</context>')
             #request_body['messages'][0]['content'] = f'<context>{context}</context>'
+        if self.prompt_end:
+            prompt = prompt + self.prompt_end
         request_body['messages'][1]['content'] = prompt
         return request_body

tale/llm/llm_io.py

@@ -18,9 +18,9 @@ def __init__(self, config: dict = None, backend_config: dict = None):
             headers['Authorization'] = f"Bearer {backend_config['OPENAI_API_KEY']}"
             self.openai_json_format = json.loads(backend_config['OPENAI_JSON_FORMAT'])
             self.headers = headers
-            self.io_adapter = LlamaCppAdapter(self.url, backend_config['STREAM_ENDPOINT'], config['USER_START'], config['USER_END'])
+            self.io_adapter = LlamaCppAdapter(self.url, backend_config['STREAM_ENDPOINT'], config['USER_START'], config['USER_END'], config.get('SYSTEM_START', ''), config.get('PROMPT_END', ''))
         else:
-            self.io_adapter = KoboldCppAdapter(self.url, backend_config['STREAM_ENDPOINT'], backend_config['DATA_ENDPOINT'], config['USER_START'], config['USER_END'])
+            self.io_adapter = KoboldCppAdapter(self.url, backend_config['STREAM_ENDPOINT'], backend_config['DATA_ENDPOINT'], config['USER_START'], config['USER_END'], config.get('SYSTEM_START', ''), config.get('PROMPT_END', ''))
             self.headers = {}
 
         self.stream = backend_config['STREAM']

tale/player.py

@@ -94,8 +94,7 @@ def tell(self, message: str, *, end: bool=False, format: bool=True, evoke: bool=
             msgs = msg.split('\n\n')
             if len(msgs) > 1:
                 for msg in msgs:
-                    self._output.print(msg, end=end, format=format)
-                    self._output.p()
+                    self._output.print(msg, end=True, format=format)
             else:
                 self._output.print(msg, end=end, format=format)
 

tests/test_llm_ext.py

@@ -175,7 +175,7 @@ def test_do_say(self):
         self.npc.do_say(what_happened='something', actor=self.npc2)
         assert(self.npc.sentiments['actor'] == 'kind')
         assert(len(self.npc._observed_events) == 2)
-        assert ["test : Hello there, how can I assist you today?\n"] == self.msg_trace_npc.messages
+        assert ["test : Hello there, how can I assist you today?\n\n"] == self.msg_trace_npc.messages
 
     @responses.activate
     def test_idle_action(self):
@@ -186,8 +186,8 @@ def test_idle_action(self):
         self.llm_util._character.io_util.response = []
         action = self.npc.idle_action()
         assert(action == 'sits down on a chair')
-        assert(llm_cache.get_events(self.npc2._observed_events) == 'test : sits down on a chair\n')
-        assert ["test : sits down on a chair\n"] == self.msg_trace_npc.messages
+        assert(llm_cache.get_events(self.npc2._observed_events) == 'test : sits down on a chair\n\n')
+        assert ["test : sits down on a chair\n\n"] == self.msg_trace_npc.messages
 
     @responses.activate
     def test_do_react(self):
@@ -256,7 +256,7 @@ def test_give_action(self):
         self.npc.autonomous_action()
         assert self.npc.search_item('test item', include_location=False) == None
         assert(self.npc2.search_item('test item', include_location=False))
-        assert ["test : Test gives test item to test\n"] == self.msg_trace_npc.messages
+        assert ["test : Test gives test item to test\n\n"] == self.msg_trace_npc.messages
 
     @responses.activate
     def test_move_action(self):