The Controller class is responsible for traversing the Graph of Operations (GoO), which is a static structure that is constructed once, before the execution starts. GoO prescribes the execution plan of thought operations and the Controller invokes their execution, generating the Graph Reasoning State (GRS).

In order for a GoO to be executed, an instance of Large Language Model (LLM) must be supplied to the controller. Currently, the framework supports the following LLMs:

GPT-4 / GPT-3.5 (Remote - OpenAI API)
Llama-2 (Local - HuggingFace Transformers)

The following section describes how to instantiate individual LLMs and the Controller to run a defined GoO. Furthermore, the process of adding new LLMs into the framework is outlined at the end.

LLM Instantiation

Create a copy of config_template.json named config.json.
Fill configuration details based on the used model (below).

GPT-4 / GPT-3.5

Adjust predefined chatgpt, chatgpt4 or create new configuration with an unique key.

Key	Value
model_id	Model name based on OpenAI model overview.
prompt_token_cost	Price per 1000 prompt tokens based on OpenAI pricing, used for calculating cumulative price per LLM instance.
response_token_cost	Price per 1000 response tokens based on OpenAI pricing, used for calculating cumulative price per LLM instance.
temperature	Parameter of OpenAI models that controls randomness and the creativity of the responses (higher temperature = more diverse and unexpected responses). Value between 0.0 and 2.0, default is 1.0. More information can be found in the OpenAI API reference.
max_tokens	The maximum number of tokens to generate in the chat completion. Value depends on the maximum context size of the model specified in the OpenAI model overview. More information can be found in the OpenAI API reference.
stop	String or array of strings specifying sequence of characters which if detected, stops further generation of tokens. More information can be found in the OpenAI API reference.
organization	Organization to use for the API requests (may be empty).
api_key	Personal API key that will be used to access OpenAI API.

Instantiate the language model based on the selected configuration key (predefined / custom).

lm = controller.ChatGPT(
    "path/to/config.json", 
    model_name=<configuration key>
)

Llama-2

Requires local hardware to run inference and a HuggingFace account.
Adjust predefined llama7b-hf, llama13b-hf, llama70b-hf or create a new configuration with an unique key.

Key	Value
model_id	Specifies HuggingFace Llama 2 model identifier (`meta-llama/<model_id>`).
cache_dir	Local directory where model will be downloaded and accessed.
prompt_token_cost	Price per 1000 prompt tokens (currently not used - local model = no cost).
response_token_cost	Price per 1000 response tokens (currently not used - local model = no cost).
temperature	Parameter that controls randomness and the creativity of the responses (higher temperature = more diverse and unexpected responses). Value between 0.0 and 1.0, default is 0.6.
top_k	Top-K sampling method described in Transformers tutorial. Default value is set to 10.
max_tokens	The maximum number of tokens to generate in the chat completion. More tokens require more memory.

Instantiate the language model based on the selected configuration key (predefined / custom).

lm = controller.Llama2HF(
    "path/to/config.json", 
    model_name=<configuration key>
)

Request access to Llama-2 via the Meta form using the same email address as for the HuggingFace account.
After the access is granted, go to HuggingFace Llama-2 model card, log in and accept the license ("You have been granted access to this model" message should appear).
Generate HuggingFace access token.
Log in from CLI with: huggingface-cli login --token <your token>.

Note: 4-bit quantization is used to reduce the model size for inference. During instantiation, the model is downloaded from HuggingFace into the cache directory specified in the config.json. Running queries using larger models will require multiple GPUs (splitting across many GPUs is done automatically by the Transformers library).

Controller Instantiation

Requires custom Prompter, Parser and instantiated GraphOfOperations - creation of these is described separately.
Use instantiated lm from above.
Prepare initial state (thought) as dictionary - this can be used in the initial prompts by the operations.

graph_of_operations = ...create

executor = controller.Controller(
    lm,
    graph_of_operations,
    <CustomPrompter()>,
    <CustomParser()>,
    <initial state>,
)
executor.run()
executor.output_graph("path/to/output.json")

After the run the graph is written to an output file, which contains individual operations, their thoughts, information about scores and validity and total amount of used tokens / cost.

Adding LLMs

More LLMs can be added by following these steps:

Create new class as a subclass of AbstractLanguageModel.
Use the constructor for loading configuration and instantiating the language model (if needed).

class CustomLanguageModel(AbstractLanguageModel):
    def __init__(
        self,
        config_path: str = "",
        model_name: str = "llama7b-hf",
        cache: bool = False
    ) -> None:
        super().__init__(config_path, model_name, cache)
        self.config: Dict = self.config[model_name]
        
        # Load data from configuration into variables if needed

        # Instantiate LLM if needed

Implement query abstract method that is used to get a list of responses from the LLM (call to remote API or local model inference).

def query(self, query: str, num_responses: int = 1) -> Any:
    # Support caching 
    # Call LLM and retrieve list of responses - based on num_responses    
    # Return LLM response structure (not only raw strings)

Implement get_response_texts abstract method that is used to get a list of raw texts from the LLM response structure produced by query.

def get_response_texts(self, query_response: Union[List[Dict], Dict]) -> List[str]:
    # Retrieve list of raw strings from the LLM response structure