brainstorm-fundamental-building-blocks-for-ai-systems.txt

brainstorm-fundamental-building-blocks-for-ai-systems.txt

LEGEND
    <processing>
    [data]
    + binary operator for structurally separable data
    & binary operator for structurally inseparable data 
        (e.g. data and instructions in prompts are mixed up)
    █ lines are the parts that are FUNCTION IMPLEMENTATION
    | lines are the parts that are just scaffolding/definitions

BAML provides all scaffolding to create LLM FUNCTIONS
how does a good scaffolding AGENT FUNCTION look like



/======= FUNCTION =======\    # Actions are also FUNCTIONS
|
█   <processing>
|
|   [structuredOutput]
|
\========================/


/=========== LLM FUNCTION ==========\     
|
█   <llmRequestBuilder>
█
█   [  llmId 
       + modelParams
       + instruction 
         & instructionContext 
         & structuredOutputDescr   ]
|
|   <llm>                     # baml sends request
|
|   [completion]
|
|   <parser>                  # baml parses completion
|
|   [structuredOutput]
|
\====================================/


/===== LOOP STATE MANAGER FUNCTION ===\
|
|   loop {
|
█       <loop state builder>
█       
█       [   functionId 
            + task 
            + taskContext 
            + workingMemory 
            + actionsWithParamsDescr         ]
|            
|
|       <function>       # action routing AND action param filling
                         # usually this will be a single LLM FUNCTION
                         # but doesn't (!) HAVE to be
|
|
|       [actionWithParams]
|
|
|       <runner>
|
|   }
|
|   [structuredOutput]   # also external effects of the actions taken
|
\====================================/