SAGE.spec.yaml

protocol:
  name: SAGE
  full_name: Sequential Agent Goal Execution Protocol
  version: 0.1.0
  description: |
    SAGE is a modular protocol for dynamically managing multi-LLM workflows. It decomposes user prompts into validated, goal-driven sub-tasks, routes them to the best local language models (via Ollama), and ensures each step meets a success threshold before proceeding. The protocol is designed for extensibility, reliability, and transparency in complex AI task orchestration.

components:
  - name: DecomposerAgent
    description: Breaks down the main user prompt into logical, meaningful sub-prompts (sub-tasks).
    input: User prompt (string), optional context (dict)
    output: List of SubPrompt objects
  - name: RouterAgent
    description: Selects the best local LLM/model (Ollama) for each sub-prompt based on task type and configuration, using a meta-router LLM.
    input: SubPrompt object
    output: ModelAssignment object
  - name: ExecutionManager
    description: Executes sub-prompts sequentially using the assigned local Ollama model, maintaining context.
    input: SubPrompt object, ModelAssignment object
    output: ExecutionResult object
  - name: Evaluator
    description: Evaluates the output of each sub-prompt against its expected goal using an LLM-based judgment protocol. The evaluator prompts a selected LLM (default: Deepseek, configurable) to act as an expert judge, returning YES/NO and a confidence score.
    input: ExecutionResult object
    output: EvaluationResult object
  - name: Retry/Reassign Handler
    description: Handles retries or reassignments if a sub-task does not meet the success threshold.
    input: SubPrompt object, EvaluationResult object
    output: ModelAssignment object (new or adjusted)
  - name: Aggregator
    description: Aggregates all sub-task results into a final, coherent response for the user.
    input: List of ExecutionResult objects
    output: AggregatedResponse object

workflow:
  steps:
    - Receive user prompt
    - Decompose prompt into sub-prompts
    - Assign each sub-prompt to the best local model (Ollama)
    - Execute sub-prompts sequentially, chaining context
    - After each execution:
        - Evaluate output against expected goal using LLM-based judgment (YES/NO, confidence)
        - If confidence >= threshold and answer is YES, proceed
        - Else, retry or reassign
    - Aggregate all results into a final response
    - Return final response to user

configuration:
  file: config/settings.yaml
  parameters:
    similarity_threshold: float (default: 0.9)
    max_retries: int (default: 3)
    default_model: string (e.g., 'gemma3:4b')
    available_models: list of strings (e.g., ['gemma3:4b', 'deepseek-r1:1.5b'])
    model_assignments: mapping of task type to model
    model_parameters: per-model parameter settings
    retry_strategy: max_attempts, backoff_factor, initial_delay
    logging: level, format
    evaluator_model: string (e.g., 'deepseek-r1:1.5b') # Model used for LLM-based evaluation

models:
  - SubPrompt
  - ModelAssignment
  - ExecutionResult
  - EvaluationResult
  - AggregatedResponse
  - SAGEConfig

extensibility:
  - Add new agent types (e.g., for planning, validation, or post-processing)
  - Integrate additional LLM providers (cloud or local)
  - Customize decomposition, routing, or evaluation logic
  - Plug in custom similarity metrics or feedback mechanisms

usage:
  - Initialize SAGE with configuration
  - Call process_prompt(prompt) to process a user prompt
  - Receive AggregatedResponse with final result and details

supported_models:
  - gemma3:4b (Ollama local)
  - deepseek-r1:1.5b (Ollama local)
  - qwen3:1.7b (Ollama local)

license: MIT