new context handling

backend_kobold_cpp.yaml

@@ -1,8 +1,7 @@
 URL: "http://localhost:5001"
 ENDPOINT: "/api/v1/generate"
-STREAM: False
+STREAM: True
 STREAM_ENDPOINT: "/api/extra/generate/stream"
 DATA_ENDPOINT: "/api/extra/generate/check"
 DEFAULT_BODY: '{"stop_sequence": "\n\n\n\n", "max_length":750, "max_context_length":4096, "temperature":0.5, "top_k":120, "top_a":0.0, "top_p":0.85, "typical_p":1.0, "tfs":1.0, "rep_pen":1.2, "rep_pen_range":256, "sampler_order":[6,0,1,3,4,2,5], "seed":-1}'
-ANALYSIS_BODY: '{}'
 GENERATION_BODY: '{"stop_sequence": "\n\n\n\n", "max_length":750, "max_context_length":4096, "temperature":1.0, "top_k":120, "top_a":0.0, "top_p":0.85, "typical_p":1.0, "tfs":1.0, "rep_pen":1.2, "rep_pen_range":256, "sampler_order":[6,0,1,3,4,2,5], "seed":-1}'

llm_config.yaml

@@ -1,12 +1,12 @@
-WORD_LIMIT: 200
+WORD_LIMIT: 200 # max number of words the model is encoraged to generate. not a hard limit
+SHORT_WORD_LIMIT: 25 # max number of words when asked to write something short. not a hard limit
 BACKEND: "kobold_cpp" # valid options: "openai", "llama_cpp", "kobold_cpp"
-ANALYSIS_BODY: '{}'
 MEMORY_SIZE: 512
-PRE_PROMPT: 'You are a creative game keeper. You craft detailed worlds and interesting characters with unique and deep personalities for the player to interact with.'
-BASE_PROMPT: "[Story context: {story_context}]; [History: {history}]; [USER_START] Rewrite [{input_text}] in your own words using the supplied Context and History to create a background for your text. Use about {max_words} words."
-ACTION_PROMPT: "[Story context: {story_context}]; [History: {history}]; [USER_START] Rewrite the Action, and nothing else, in your own words using the supplied Context and Location. History is what happened before. Use less than {max_words} words. [Action: {input_text}]."
-DIALOGUE_PROMPT: 'Story context: {story_context}; Location: {location}; The following is a conversation between {character1} and {character2}; {character1}:{character1_description}; {character2}:{character2_description}; {character2}s sentiment towards {character1}: {sentiment}. Write a single response as {character2} in third person pov, using {character2} description. If {character2} has a quest active, they will discuss it based on its status. Respond in JSON using this template: {{"response":"may be both dialogue and action.", "sentiment":"sentiment based on response", "give":"if any physical item of {character2}s is given as part of the dialogue. Or nothing."}}. [USER_START]Continue the following conversation as {character2}: {previous_conversation}'
-ITEM_PROMPT: 'Items:[{items}];Characters:[{character1},{character2}]  \n\n [USER_START] Decide if an item was explicitly given, taken, dropped or put somewhere in the following text:[Text:{text}]. Insert your thoughts about whether an item was explicitly given, taken, put somewhere or dropped in "my thoughts", and the results in "item", "from" and "to", or make them empty if . Insert {character1}s sentiment towards {character2} in a single word in "sentiment assessment". Fill in the following JSON template: {{ "thoughts":"", "result": {{ "item":"", "from":"", "to":""}}, {{"sentiment":"sentiment assessment"}} }} End of example. \n\n Write your response in valid JSON.'
+DIALOGUE_TEMPLATE: '{"response":"may be both dialogue and action.", "sentiment":"sentiment based on response", "give":"if any physical item of {character2}s is given as part of the dialogue. Or nothing."}'
+ACTION_TEMPLATE: '{"goal": reason for action, "thoughts":thoughts about performing action, "action":action chosen, "target":character, item or exit or description, "text": if anything is said during the action}'
+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.'
+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: 'The following is a combat scene between user {attacker} and {victim} in {location}, {location_description} into a vivid description. [USER_START] Rewrite the following combat result in about 150 words, using the characters weapons and their health status: 1.0 is highest, 0.0 is lowest. Combat Result: {attacker_msg}'
 PRE_JSON_PROMPT: 'Below is an instruction that describes a task, paired with an input that provides further context. Write a response in valid JSON format that appropriately completes the request.'
 CREATE_CHARACTER_PROMPT: 'Story context: {story_context}; World info: {world_info};[USER_START] For a {story_type}, create a diverse character with rich personality that can be interacted with using the story context and keywords. {{quest_prompt}} Do not mention height. Story context: {story_context}; keywords: {keywords}. Fill in this JSON template and write nothing else: {{"name":"", "description": "50 words", "appearance": "25 words", "personality": "50 words", "money":(int), "level":"", "gender":"m/f/n", "age":(int), "race":""}}'
@@ -28,6 +28,6 @@ PLAYER_ENTER_PROMPT: 'Story context: {story_context}; World info: {world_info};
 QUEST_PROMPT: '[Story context: {story_context}]; World info: {world_info}; Zone info: {zone_info}; Character: {character_card}; [USER_START] In an RPG described as {story_type}, {character_name} needs someone to perform a task. Based on the following input, come up with a suitable reason for it, using {character_name}s personality and history. Task info: {base_quest}. Fill in this JSON template and do not write anything else: {{"reason":""}} \n\n '
 NOTE_QUEST_PROMPT: '[Story context: {story_context}]; World info: {world_info}; Zone info: {zone_info}; [USER_START]For an RPG described as {story_type}, generate a quest that starts from reading a note. The reader must find and talk to a person. Fill in the following JSON template and write nothing else.: {{"reason": "only what the note says. 50 words.", "type":"talk", "target":"who to talk to", "location":"", "name":"name of quest"}}'
 NOTE_LORE_PROMPT: '[Story context: {story_context}]; World info: {world_info}; Zone info: {zone_info}; [USER_START]For an RPG described as {story_type}, Decide what is written on a note that has been found. Use the provided story and world information to generate a piece of lore. Use about 50 words.'
-ACTION_PROMPT: 'Story context: {story_context};<location>{location}</location> \nAct as as {character_name} in a {story_type} RPG and perform an action.<description>{character}</description> <actions>{actions}</actions>. <items>{location_items}</items>. <characters>{characters}</characters> <exits>{exits}</exits>.\nPick an action according to {character_name}s description and mood. If suitable, select something to perform the action on (target). Make sure the action is from the list and is related to {character_name}s thoughts.\n Respond using JSON in the following format: """{{"goal": reason for action, "thoughts":thoughts about performing action, "action":action chosen, "target":character, item or exit or description, "text": if anything is said during the action}}"""<history>{history}</history>'
+ACTION_PROMPT: '<context>{context}</context>\nAct as as {character_name}.\nUsing the information supplied inside the <context> tag, pick an action according to {character_name}s description and mood. If suitable, select something to perform the action on (target). The action should be in the supplied list and should be related to {character_name}s thoughts.\n Respond using JSON in the following format: """{action_template}"""'
 USER_START: '### Instruction:'
-USER_END: '### Response:'
+USER_END: '### Response:\n'

tale/llm/contexts/ActionContext.py

@@ -0,0 +1,30 @@
+
+
+import json
+from tale.base import Location
+from tale.llm.contexts.BaseContext import BaseContext
+
+
+class ActionContext(BaseContext):
+
+    def __init__(self, story_context: str, story_type: str, character_name: str, character_card: str, event_history: str, location: Location):
+        super().__init__(story_context)
+        self.story_type = story_type
+        self.character_name = character_name
+        self.character_card = character_card
+        self.event_history = event_history
+        self.location = location
+
+
+    def to_prompt_string(self) -> str:
+        actions = ', '.join(['move, say, attack, wear, remove, wield, take, eat, drink, emote'])
+        characters = {}
+        for living in self.location.livings:
+            if living.visible and living.name != self.character_name.lower():
+                if living.alive:
+                    characters[living.name] = living.short_description
+                else:
+                    characters[living.name] = f"{living.short_description} (dead)"
+        exits = self.location.exits.keys()
+        items = [item.name for item in self.location.items if item.visible]
+        return f"Story context:{self.story_context}; Story type:{self.story_type}; Available actions: {actions}; Location:{self.location.name}, {self.location.description}; Available exits: {exits}; Self: {self.character_card}; Present items: {items}; Present characters: {json.dumps(characters)}; History:{self.event_history};"

tale/llm/contexts/BaseContext.py

@@ -0,0 +1,9 @@
+
+
+class BaseContext():
+
+    def __init__(self, story_context: str) -> None:
+        self.story_context = story_context
+
+    def to_prompt_string(self) -> str:
+        pass

tale/llm/contexts/DialogueContext.py

@@ -0,0 +1,24 @@
+
+
+from tale.llm.contexts.BaseContext import BaseContext
+
+
+class DialogueContext(BaseContext):
+
+    def __init__(self, 
+            story_context: str,
+            location_description: str,
+            speaker_card: str,
+            speaker_name: str,
+            target_name: str,
+            target_description: str):
+        super().__init__(story_context)
+        self.location_description = location_description
+        self.speaker_card = speaker_card
+        self.speaker_name = speaker_name
+        self.target_name = target_name
+        self.target_description = target_description
+
+
+    def to_prompt_string(self) -> str:
+        return f"Story context:{self.story_context}; Location:{self.location_description}; Self:{self.speaker_name}:{self.speaker_card}; Listener:{self.target_name}:{self.target_description};"

tale/llm/contexts/EvokeContext.py

@@ -0,0 +1,10 @@
+from tale.llm.contexts.BaseContext import BaseContext
+
+class EvokeContext(BaseContext):
+
+    def __init__(self, story_context: str, history: str) -> None:
+        super().__init__(story_context)
+        self.history = history
+
+    def to_prompt_string(self) -> str:
+        return f"Story context:{self.story_context}; History:{self.history};"

tale/llm/llm_utils.py

@@ -6,8 +6,11 @@
 from tale.base import Location
 from tale.image_gen.base_gen import ImageGeneratorBase
 from tale.llm.character import CharacterBuilding
+from tale.llm.contexts.ActionContext import ActionContext
+from tale.llm.contexts.EvokeContext import EvokeContext
 from tale.llm.llm_ext import DynamicStory
 from tale.llm.llm_io import IoUtil
+from tale.llm.contexts.DialogueContext import DialogueContext
 from tale.llm.quest_building import QuestBuilding
 from tale.llm.story_building import StoryBuilding
 from tale.llm.world_building import WorldBuilding
@@ -37,9 +40,10 @@ def __init__(self, io_util: IoUtil = None):
         self.default_body = json.loads(backend_config['DEFAULT_BODY'])
         self.memory_size = config_file['MEMORY_SIZE']
         self.pre_prompt = config_file['PRE_PROMPT'] # type: str
-        self.base_prompt = config_file['BASE_PROMPT'] # type: str
+        self.evoke_prompt = config_file['BASE_PROMPT'] # type: str
         self.combat_prompt = config_file['COMBAT_PROMPT'] # type: str
         self.word_limit = config_file['WORD_LIMIT']
+        self.short_word_limit = config_file['SHORT_WORD_LIMIT']
         self.story_background_prompt = config_file['STORY_BACKGROUND_PROMPT'] # type: str
         self.json_grammar = config_file['JSON_GRAMMAR'] # type: str
         self.__story = None # type: DynamicStory
@@ -83,13 +87,11 @@ def evoke(self, player_io: TextBuffer, message: str, short_len : bool=False, rol
             return output_template.format(message=message, text=cached_look), rolling_prompt
 
         trimmed_message = parse_utils.remove_special_chars(str(message))
-
-        amount = 25
+        context = EvokeContext(story_context=self.__story_context, history=rolling_prompt if not skip_history or alt_prompt else '')
         prompt = self.pre_prompt
-        prompt += alt_prompt or (self.base_prompt.format(
-            story_context=self.__story_context,
-            history=rolling_prompt if not skip_history or alt_prompt else '',
-            max_words=self.word_limit if not short_len else amount,
+        prompt += alt_prompt or (self.evoke_prompt.format(
+            context=context.to_prompt_string(),
+            max_words=self.word_limit if not short_len else self.short_word_limit,
             input_text=str(trimmed_message)))
 
         request_body = deepcopy(self.default_body)
@@ -98,9 +100,9 @@ def evoke(self, player_io: TextBuffer, message: str, short_len : bool=False, rol
             text = self.io_util.synchronous_request(request_body, prompt=prompt)
             llm_cache.cache_look(text, text_hash_value)
             return output_template.format(message=message, text=text), rolling_prompt
-
+        text = self.io_util.stream_request(request_body=request_body, player_io=player_io, prompt=prompt, io=self.connection)     
         player_io.print(output_template.format(message=message, text=text), end=False, format=True, line_breaks=False)
-        text = self.io_util.stream_request(request_body, player_io, self.connection, prompt=prompt)
+
         llm_cache.cache_look(text, text_hash_value)
 
         return '\n', rolling_prompt
@@ -114,16 +116,17 @@ def generate_dialogue(self, conversation: str,
                           location_description = '',
                           event_history='',
                           short_len : bool=False):
-        return self._character.generate_dialogue(conversation, 
-                                                 character_card=character_card, 
-                                                 character_name=character_name, 
-                                                 target=target, 
-                                                 target_description=target_description, 
-                                                 sentiment=sentiment, 
-                                                 location_description=location_description, 
-                                                 story_context=self.__story_context, 
-                                                 event_history=event_history,
-                                                 short_len=short_len)
+        dialogue_context = DialogueContext(story_context=self.__story_context,
+                                           location_description=location_description,
+                                           speaker_card=character_card,
+                                           speaker_name=character_name,
+                                           target_name=target,
+                                           target_description=target_description)
+        return self._character.generate_dialogue(context=dialogue_context,
+                                                conversation=conversation,
+                                                sentiment=sentiment,
+                                                event_history=event_history,
+                                                short_len=short_len)
 
     def update_memory(self, rolling_prompt: str, response_text: str):
         """ Keeps a history of the last couple of events"""
@@ -234,7 +237,13 @@ def generate_image(self, character_name: str, character_appearance: dict = '', s
         return result
 
     def free_form_action(self, location: Location, character_name: str,  character_card: str = '', event_history: str = ''):
-        return self._character.free_form_action(self.__story_context, self.__story_type, location, character_name, character_card, event_history)
+        action_context = ActionContext(story_context=self.__story_context,
+                                       story_type=self.__story_type,
+                                       character_name=character_name,
+                                       character_card=character_card,
+                                       event_history=event_history,
+                                       location=location)
+        return self._character.free_form_action(action_context)
 
 
     def set_story(self, story: DynamicStory):

tests/test_llm_utils.py

@@ -65,7 +65,7 @@ def test_perform_travel_action(self):
     def test_generate_dialogue(self):
         # mostly testing that prompt works
         self.llm_util._character.io_util.response = ['{"response":"Hello there", "sentiment":"cheerful", "give":"ale"}']
-        result, item, sentiment = self.llm_util._character.generate_dialogue(conversation='test conversation', 
+        result, item, sentiment = self.llm_util.generate_dialogue(conversation='test conversation', 
                                                             character_card='{}', 
                                                             character_name='Norhardt', 
                                                             target='Arto', 
@@ -80,7 +80,7 @@ def test_generate_dialogue(self):
     def test_generate_dialogue_json(self):
         # mostly testing that prompt works
         self.llm_util._character.io_util.response = ["{\n  \"response\": \"Autumn greets Test character with a warm smile, her golden hair shining in the sunlight. She returns the greeting, her voice filled with kindness, \'Hello there, how can I assist you today?\'\"\n}"]
-        result, item, sentiment = self.llm_util._character.generate_dialogue(conversation='test conversation', 
+        result, item, sentiment = self.llm_util.generate_dialogue(conversation='test conversation', 
                                                             character_card='{}', 
                                                             character_name='Norhardt', 
                                                             target='Arto',