Fixed a lot of bugs and imporved grammar generator.

examples/example_agent.py

@@ -11,7 +11,7 @@
 from llama_cpp_agent.function_calling import LlamaCppFunctionTool
 
 function_tools = [LlamaCppFunctionTool(GetFileList), LlamaCppFunctionTool(ReadTextFile),
-                  LlamaCppFunctionTool(WriteTextFile, has_triple_quoted_string=True)]
+                  LlamaCppFunctionTool(WriteTextFile, has_triple_quoted_string=True, triple_quoted_string_field_name="file_content")]
 
 function_tool_registry = LlamaCppAgent.get_function_tool_registry(function_tools)
 

examples/example_agent_auto_coder.py

@@ -11,7 +11,7 @@
 from llama_cpp_agent.function_calling import LlamaCppFunctionTool
 
 function_tools = [LlamaCppFunctionTool(SendMessageToUser), LlamaCppFunctionTool(GetFileList), LlamaCppFunctionTool(ReadTextFile),
-                  LlamaCppFunctionTool(WriteTextFile, has_triple_quoted_string=True)]
+                  LlamaCppFunctionTool(WriteTextFile, has_triple_quoted_string=True, triple_quoted_string_field_name="file_content")]
 
 function_tool_registry = LlamaCppAgent.get_function_tool_registry(function_tools)
 

examples/example_agent_models_auto_coder.py

@@ -8,7 +8,6 @@
 base_folder = "dev"
 
 
-
 def agent_dev_folder_setup(custom_base_folder=None):
     global base_folder
     base_folder = custom_base_folder
@@ -46,8 +45,9 @@ class WriteTextFile(BaseModel):
 
     # Allow free output for the File Content to Enhance LLM Output
 
-    triple_quoted_string: str = Field(...,
-                             description="Triple quoted string for unconstrained output.")
+    file_content: str = Field(...,
+                              description="Triple quoted string for unconstrained output.", triple_quoted_string=True)
+
     def run(self):
 
         if self.directory == "":
@@ -84,7 +84,7 @@ def run(self):
 
         # Write back to file
         with open(file_path, write_mode, encoding="utf-8") as file:
-            file.writelines(self.triple_quoted_string)
+            file.writelines(self.file_content)
 
         return f"Content written to '{self.filename_without_extension}{self.filename_extension}'."
 

examples/example_book_dataset_creation.py

@@ -4,14 +4,14 @@
 from pydantic import BaseModel, Field
 
 from llama_cpp_agent.llm_agent import LlamaCppAgent
-from llama_cpp_agent.gbnf_grammar_generator.gbnf_grammar_from_pydantic_models import generate_gbnf_grammar_and_documentation
+from llama_cpp_agent.gbnf_grammar_generator.gbnf_grammar_from_pydantic_models import \
+    generate_gbnf_grammar_and_documentation
 from llama_cpp_agent.providers.llama_cpp_server_provider import LlamaCppServerLLMSettings
 
 main_model = LlamaCppServerLLMSettings(
     completions_endpoint_url="http://127.0.0.1:8080/completion"
 )
 
-
 text = """The Feynman Lectures on Physics is a physics textbook based on some lectures by Richard Feynman, a Nobel laureate who has sometimes been called "The Great Explainer". The lectures were presented before undergraduate students at the California Institute of Technology (Caltech), during 1961–1963. The book's co-authors are Feynman, Robert B. Leighton, and Matthew Sands."""
 
 
@@ -38,8 +38,6 @@ class Book(BaseModel):
 gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation([Book])
 
 llama_cpp_agent = LlamaCppAgent(main_model, debug_output=True,
-                              system_prompt="You are an advanced AI, tasked to create JSON database entries for books.\n\n\n" + documentation)
-
+                                system_prompt="You are an advanced AI, tasked to create JSON database entries for books.\n\n\n" + documentation)
 
 print(llama_cpp_agent.get_chat_response(text, temperature=0.15, grammar=gbnf_grammar))
-

examples/example_dataframe_creation.py

@@ -62,7 +62,7 @@ class Database(BaseModel):
     )
 
 
-gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation([Database], False,
+gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation([Database],
                                                                       model_prefix="Response Model",
                                                                       fields_prefix="Response Model Field")
 
@@ -73,7 +73,7 @@ class Database(BaseModel):
 
 def dataframe(data: str) -> Database:
     prompt = data
-    response = llama_cpp_agent.get_chat_response(message=prompt, temperature=0.25, grammar=gbnf_grammar)
+    response = llama_cpp_agent.get_chat_response(message=prompt, temperature=0.65, grammar=gbnf_grammar)
 
     database = extract_object_from_response(response, Database)
     return database

examples/example_embodied_function_calling.py

@@ -10,7 +10,8 @@
 
 function_tools = [LlamaCppFunctionTool(SendMessageToUser), LlamaCppFunctionTool(GetFileList),
                   LlamaCppFunctionTool(ReadTextFile),
-                  LlamaCppFunctionTool(WriteTextFile, has_triple_quoted_string=True)]
+                  LlamaCppFunctionTool(WriteTextFile, has_triple_quoted_string=True,
+                                       triple_quoted_string_field_name="file_content")]
 
 function_tool_registry = LlamaCppAgent.get_function_tool_registry(function_tools)
 
@@ -42,6 +43,7 @@
                                 system_prompt=system_prompt,
                                 predefined_messages_formatter_type=MessagesFormatterType.CHATML)
 
+print(function_tool_registry.gbnf_grammar)
 response = llama_cpp_agent.get_chat_response(
     'Write a engaging rap song about the drug problem in the USA in the "USARap.txt" file under "./".',
     temperature=0.75, function_tool_registry=function_tool_registry)

examples/example_function_calling_agent.py

@@ -132,7 +132,7 @@ def send_message_to_user_callback(message: str):
     pydantic_functions=[Calculator],
     # Callback for receiving messages for the user.
     send_message_to_user_callback=send_message_to_user_callback, debug_output=True)
-
+print(function_call_agent.tool_registry.gbnf_grammar)
 while True:
     user_input = input(">")
     function_call_agent.generate_response(user_input)

examples/example_knowledge_graph.py

@@ -46,16 +46,15 @@ class KnowledgeGraph(BaseModel):
     edges: List[Edge] = Field(..., default_factory=list)
 
 
-gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation([KnowledgeGraph], False,
+gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation([KnowledgeGraph],
                                                                       model_prefix="Response Model",
                                                                       fields_prefix="Response Model Field")
 
 print(gbnf_grammar)
-grammar = LlamaGrammar.from_string(gbnf_grammar, verbose=True)
 
 llama_cpp_agent = LlamaCppAgent(main_model, debug_output=True,
-                              system_prompt="You are an advanced AI assistant responding in JSON format.\n\nAvailable JSON response models:\n\n" + documentation,
-                              predefined_messages_formatter_type=MessagesFormatterType.CHATML)
+                                system_prompt="You are an advanced AI assistant responding in JSON format.\n\nAvailable JSON response models:\n\n" + documentation,
+                                predefined_messages_formatter_type=MessagesFormatterType.CHATML)
 
 from graphviz import Digraph
 
@@ -78,7 +77,7 @@ def visualize_knowledge_graph(kg: KnowledgeGraph):
 def generate_graph(user_input: str) -> KnowledgeGraph:
     prompt = f'''Help me understand the following by describing it as a extremely detailed knowledge graph with at least 20 nodes: {user_input}'''.strip()
     response = llama_cpp_agent.get_chat_response(message=prompt, temperature=0.65, mirostat_mode=0, mirostat_tau=5.0,
-                                               mirostat_eta=0.1, grammar=grammar)
+                                                 mirostat_eta=0.1, grammar=gbnf_grammar)
     knowledge_graph = json.loads(response)
     cls = KnowledgeGraph
     knowledge_graph = cls(**knowledge_graph)

examples/example_manual_function_calling.py

@@ -16,8 +16,8 @@ def calculate_a_to_the_power_b(a: Union[int | float], b: Union[int | float]):
 
 DynamicSampleModel = create_dynamic_model_from_function(calculate_a_to_the_power_b)
 
-grammar, documentation = generate_gbnf_grammar_and_documentation([DynamicSampleModel], root_rule_class="function",
-                                                                 root_rule_content="params")
+grammar, documentation = generate_gbnf_grammar_and_documentation([DynamicSampleModel], outer_object_name="function",
+                                                                 outer_object_content="params")
 
 main_model = LlamaCppServerLLMSettings(
     completions_endpoint_url="http://127.0.0.1:8080/completion"

examples/example_multi_file_generation.py

@@ -3,7 +3,6 @@
 
 import json
 
-
 from llama_cpp_agent.messages_formatter import MessagesFormatterType
 from llama_cpp import Llama, LlamaGrammar
 from pydantic import BaseModel
@@ -29,8 +28,6 @@
 )
 
 
-
-
 class File(BaseModel):
     """
     Correctly named file with contents.
@@ -56,20 +53,20 @@ class Program(BaseModel):
     files: List[File] = Field(..., description="List of files")
 
 
-gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation(pydantic_model_list=[Program], look_markdown_code_block=False)
+gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation(pydantic_model_list=[Program])
 
 print(gbnf_grammar)
-grammar = LlamaGrammar.from_string(gbnf_grammar, verbose=True)
+
 
 llama_cpp_agent = LlamaCppAgent(main_model, debug_output=True,
-                              system_prompt="You are a world class programming AI capable of writing correct python scripts and modules. You will name files correct, include __init__.py files and write correct python code with correct imports.\n\nYou are responding in JSON format.\n\nAvailable JSON response models:\n\n" + documentation.strip() + "\n\nAlways provide full implementation to the user!!!!",
-                              predefined_messages_formatter_type=MessagesFormatterType.MIXTRAL)
+                                system_prompt="You are a world class programming AI capable of writing correct python scripts and modules. You will name files correct, include __init__.py files and write correct python code with correct imports.\n\nYou are responding in JSON format.\n\nAvailable JSON response models:\n\n" + documentation.strip() + "\n\nAlways provide full implementation to the user!!!!",
+                                predefined_messages_formatter_type=MessagesFormatterType.MIXTRAL)
 
 
 def develop(data: str) -> Program:
     prompt = data
     response = llama_cpp_agent.get_chat_response(message=prompt, temperature=0.35, mirostat_mode=2, mirostat_tau=4.0,
-                                               mirostat_eta=0.1, grammar=grammar)
+                                                 mirostat_eta=0.1, grammar=gbnf_grammar)
     json_obj = json.loads(response)
     cls = Program
     ai_program = cls(**json_obj)

examples/example_simple_function_calling.py

@@ -7,7 +7,8 @@
 
 from llama_cpp_agent.messages_formatter import MessagesFormatterType
 from llama_cpp_agent.function_calling import LlamaCppFunctionTool
-from llama_cpp_agent.providers.openai_endpoint_provider import OpenAIEndpointSettings
+from llama_cpp_agent.providers.llama_cpp_server_provider import LlamaCppServerLLMSettings
+
 
 
 # Simple calculator tool for the agent that can add, subtract, multiply, and divide.
@@ -43,8 +44,8 @@ def run(self):
 
 function_tool_registry = LlamaCppAgent.get_function_tool_registry(function_tools)
 
-main_model = OpenAIEndpointSettings(
-    "http://localhost:8080/v1/completions"
+main_model = LlamaCppServerLLMSettings(
+    "http://localhost:8080/completion"
 )
 llama_cpp_agent = LlamaCppAgent(main_model, debug_output=False,
                                 system_prompt="You are an advanced AI, tasked to assist the user by calling functions in JSON format.\n\n\n" + function_tool_registry.get_documentation(),

examples/example_two_agent_discussion.py

@@ -1,6 +1,6 @@
 from llama_cpp import Llama
 
-from llama_cpp_agent.llm_prompt_template import PromptTemplateFields, Prompter
+from llama_cpp_agent.llm_prompt_template import PromptTemplateFields, PromptTemplate
 from llama_cpp_agent.llm_agent import LlamaCppAgent
 from llama_cpp_agent.messages_formatter import MessagesFormatterType
 
@@ -38,11 +38,11 @@
 template_fields.add_field("user_role_name", user_role_name)
 
 assistant_system_prompt_template = """You are John Smith, a {assistant_role_name}, you are collaborating with Richard Steen, an expert {user_role_name}."""
-assistant_system_prompter = Prompter.from_string(assistant_system_prompt_template)
+assistant_system_prompter = PromptTemplate.from_string(assistant_system_prompt_template)
 assistant_system_prompt = assistant_system_prompter.generate_prompt(template_fields.get_fields_dict())
 
 user_system_prompt_template = """You are Richard Steen, a {user_role_name}, you are collaborating with Richard Steen, an expert {assistant_role_name}."""
-user_system_prompter = Prompter.from_string(user_system_prompt_template)
+user_system_prompter = PromptTemplate.from_string(user_system_prompt_template)
 user_system_prompt = user_system_prompter.generate_prompt(template_fields.get_fields_dict())
 
 agent_assistant = LlamaCppAgent(model=main_model, name="John Smith", system_prompt=assistant_system_prompt,

examples/function_calling_agent.json

@@ -0,0 +1,56 @@
+{
+    "llama_generation_settings": {
+        "max_tokens": 0,
+        "temperature": 0.65,
+        "top_k": 0,
+        "top_p": 0.5,
+        "min_p": 0.05,
+        "typical_p": 1.0,
+        "repeat_penalty": 1.0,
+        "mirostat_mode": 0,
+        "mirostat_tau": 5.0,
+        "mirostat_eta": 0.1,
+        "tfs_z": 0.975,
+        "stop_sequences": null,
+        "stream": true,
+        "print_output": true
+    },
+    "system_prompt": "You are an advanced AI assistant. You are interacting with your environment and the user by calling functions. You call functions by writing JSON objects, which represent specific function calls. Below is a list of your available JSON functions:\n\nFunction: send-message-to-user\n  Description: \n        Send a message to the user.\n        \n        Args:\n            message (str): The message to be sent.\n  Function Parameters:\n    message (str)\n\nFunction: calculator\n  Description: \n        Perform a math operation on two numbers.\n  Function Parameters:\n    number_one (any):\n      Description: First number.\n    operation (math-operation):\n      Description: Math operation to perform.\n    number_two (any):\n      Description: Second number.\n\nFunction: write-to-file\n  Description: \n        Write file to the user filesystem.\n        :param chain_of_thought: Your chain of thought while writing the file.\n        :param file_path: The file path includes the filename and file ending.\n        :param file_content: The actual content to write.\n  Function Parameters:\n    chain_of_thought (str)\n    file_path (str)\n    file_content (str)\n\nFunction: read-file\n  Description: \n        Read file from the user filesystem.\n        :param file_path: The file path includes the filename and file ending.\n        :return: File content.\n  Function Parameters:\n    file_path (str)\n\nFunction: get-current-weather\n  Function Parameters:\n    location (str)\n    unit (unit or none-type)\n\n",
+    "k_last_messages_from_chat_history": 0,
+    "debug_output": true,
+    "messages": [
+        {
+            "role": "user",
+            "content": "What is the weather in New York?"
+        },
+        {
+            "role": "assistant",
+            "content": "{\n  \"function\": \"get-current-weather\",\n  \"function-parameters\": {\n    \"location\": \"New York\",\n    \"unit\": \"celsius\"\n  }\n}"
+        },
+        {
+            "role": "function",
+            "content": "Function Call Result: {\"location\": \"New York\", \"temperature\": \"24\", \"unit\": \"celsius\"}"
+        },
+        {
+            "role": "assistant",
+            "content": "{\n  \"function\": \"send-message-to-user\",\n  \"function-parameters\": {\n    \"message\": \"The current temperature in New York is 24 degrees Celsius.\"\n  }\n}"
+        }
+    ],
+    "custom_messages_formatter": {
+        "PRE_PROMPT": "",
+        "SYS_PROMPT_START": "<|im_start|>system\n",
+        "SYS_PROMPT_END": "<|im_end|>\n",
+        "USER_PROMPT_START": "<|im_start|>user\n",
+        "USER_PROMPT_END": "<|im_end|>\n",
+        "ASSISTANT_PROMPT_START": "<|im_start|>assistant\n",
+        "ASSISTANT_PROMPT_END": "<|im_end|>\n",
+        "INCLUDE_SYS_PROMPT_IN_FIRST_USER_MESSAGE": false,
+        "DEFAULT_STOP_SEQUENCES": [
+            "<|im_end|>"
+        ],
+        "FUNCTION_PROMPT_START": "",
+        "FUNCTION_PROMPT_END": "",
+        "USE_USER_ROLE_FUNCTION_CALL_RESULT": true,
+        "STRIP_PROMPT": true
+    }
+}

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "llama-cpp-agent"
-version = "0.0.12"
+version = "0.0.13"
 description = "A framework for building LLM based AI agents with llama-cpp-python."
 
 readme = "ReadMe.md"

src/llama_cpp_agent/function_calling.py

@@ -30,11 +30,14 @@ class LlamaCppFunctionTool:
         __call__(*args, **kwargs): Calls the Pydantic model with the provided keyword arguments.
     """
     def __init__(self, pydantic_model: Type[BaseModel], has_markdown_code_block=False, has_triple_quoted_string=False,
+                 markdown_code_block_field_name=None, triple_quoted_string_field_name=None,
                  **additional_parameters):
         self.model = pydantic_model
         self.look_for_field_string = has_markdown_code_block or has_triple_quoted_string
         self.has_markdown_code_block = has_markdown_code_block
         self.has_triple_quoted_string = has_triple_quoted_string
+        self.markdown_code_block_field_name = markdown_code_block_field_name
+        self.triple_quoted_string_field_name = triple_quoted_string_field_name
         self.additional_parameters = additional_parameters if additional_parameters else {}
 
     def __call__(self, *args, **kwargs):
@@ -116,17 +119,15 @@ def finalize(self):
         Finalize the registry, generating the GBNF grammar and documentation.
         """
         pydantic_function_models = []
-        look_markdown_code_block = False
+
         for function_tool in self.function_tools.values():
             pydantic_function_models.append(function_tool.model)
-            if function_tool.look_for_field_string:
-                look_markdown_code_block = True
+
         for function_tool in self.function_tools_containing_field_string.values():
             pydantic_function_models.append(function_tool.model)
-            if function_tool.look_for_field_string:
-                look_markdown_code_block = True
+
         gbnf_grammar, documentation = generate_gbnf_grammar_and_documentation(
-            pydantic_function_models, look_markdown_code_block, look_markdown_code_block, self.tool_root,
+            pydantic_function_models, self.tool_root,
             self.tool_rule_content, self.model_prefix,
             self.fields_prefix)
 
@@ -169,9 +170,9 @@ def handle_function_call(self, function_call_response: str):
                     marker = "'''" if self.function_tools_containing_field_string[name].has_triple_quoted_string else "```"
                     function_call, content = parse_json_response_with_markdown_code_block_or_triple_quoted_string(function_call_response, marker)
                     if self.function_tools_containing_field_string[function_call[self.tool_root]].has_markdown_code_block:
-                        function_call[self.tool_rule_content]["markdown_code_block"] = content
+                        function_call[self.tool_rule_content][tool.markdown_code_block_field_name] = content
                     elif self.function_tools_containing_field_string[function_call[self.tool_root]].has_triple_quoted_string:
-                        function_call[self.tool_rule_content]["triple_quoted_string"] = content
+                        function_call[self.tool_rule_content][tool.triple_quoted_string_field_name] = content
 
                     output = self.intern_function_call(function_call, with_markdown_code_block=True)
                     return output