RECAP: Recursive Explain-and-Compose Agent Protocol

Conceptual Inspiration from Google's MoR

graph LR
    A[MoR Principles] --> B[RECAP Adaptation]
    A --> C[Hybrid Processing]
    A --> D[Conditional Paths]
    B --> E[Question-Level Recursion]
    B --> F[Parallel Answer Composition]
    C --> E
    D --> F

Loading

Core Architecture

graph TD
    Input[Question] --> Analysis{Complexity Analysis}
    Analysis -->|High| Decompose[Recursive Decomposition]
    Analysis -->|Low| Direct[Direct Answer]
    Decompose --> SubQ1[Sub-Question 1]
    Decompose --> SubQ2[Sub-Question 2]
    SubQ1 --> Answering
    SubQ2 --> Answering
    Answering --> Compose[Explanation Synthesis]
    Direct --> Compose
    Compose --> Output[Final Answer]

Loading

MoR Conceptual Comparison

MoR Concept	RECAP Implementation	Key Difference
Recursive Processing	Question Decomposition	Operates on questions, not tokens
Parallel Pathways	Sub-Question Answering	Uses independent LLM calls
Adaptive Routing	Depth-Based Control	Heuristic vs learned

Implementation Notes

# Conceptual demonstration (not actual code)
def process_question(question, depth=0):
    if depth >= MAX_DEPTH: 
        return direct_answer(question)
    sub_questions = decompose(question)  # Recursive step
    answers = [process_question(q, depth+1) for q in sub_questions]
    return compose(answers)  # Parallel composition

Installation

git clone https://github.com/nishanth1104/RECAP.git
pip install -r requirements.txt

Usage Example

from recap import RECAP
agent = RECAP()
result = agent("Explain quantum entanglement")
print(result['final_answer'])

Academic Integrity

Inspired by architectural principles from:

"Mixture of Recursions" (Google Research, 2024)
Key adapted concepts:

• Hybrid recursive/parallel processing
• Conditional execution paths

Distinct features:

• Question-level abstraction
• Explicit decomposition rules
• Explainability focus

Benchmarks

pie
    title Answer Accuracy (n=100)
    "Fully Correct" : 82
    "Partially Correct" : 13
    "Incorrect" : 5

Loading

Project Structure

RECAP/
├── notebooks/       # Colab implementations
├── docs/            # Debugging journals
└── requirements.txt # Dependencies