MeTTa-MCP

chevron-icon
RFP Proposals
Top
chevron-icon
project-presentation-img
user-profile-img
KeyvanMSadeghi
Project Owner

MeTTa-MCP

Expert Rating

n/a

Overview

This project introduces a Model Context Protocol (MCP) agent to integrate MeTTa and PyNeuraLogic with LLMs/VLMs, enhancing higher-order reasoning and experiential learning. MCP will enable semantic delegation of tasks to a MeTTa reasoner and a PyNeuraLogic component for rule-based processing. Experiential rules will be derived from user "vibe coding" interactions in a mobile app, inspired by AIRIS. The PoC aims to demonstrate improved LLM reasoning, interpretability, and symbolic knowledge integration, aligning with neural-symbolic DNN exploration for PRIMUS.

RFP Guidelines

Neural-symbolic DNN architectures

Complete & Awarded
  • Type SingularityNET RFP
  • Total RFP Funding $160,000 USD
  • Proposals 17
  • Awarded Projects 1
author-img
SingularityNET
Apr. 14, 2025

This RFP invites proposals to explore and demonstrate the use of neural-symbolic deep neural networks (DNNs), such as PyNeuraLogic and Kolmogorov Arnold Networks (KANs), for experiential learning and/or higher-order reasoning. The goal is to investigate how these architectures can embed logic rules derived from experiential systems like AIRIS or user-supplied higher-order logic, and apply them to improve reasoning in graph neural networks (GNNs), LLMs, or other DNNs. Bids are expected to range from $40,000 - $100,000.

Proposal Description

Our Team

Keyvan M. Sadeghi: AI/AGI researcher (OpenCog familiarity), leads MCP & MeTTa design.

Julien Serbanescu (Co-lead): AI/ML & RAG expert (U Guelph), contributes to MCP-LLM integration, "MeTTa digest tool" design & RAG comparison.

Farhoud Mojahedzadeh: MeTTa specialist (DeepFunding Unsupervised Learning proposal contributor), focuses on `metta_reasoner` core & Atomspace optimization.

Team combines symbolic AI, ML, LLM augmentation, & Python/MeTTa dev skills.

Company Name (if applicable)

https://vhybZ.com

Project details

This project proposes the development of a Model Context Protocol (MCP) agent framework, a novel approach to significantly enhance the reasoning capabilities, interpretability, and experiential learning potential of Large Language Models (LLMs) and Vision Language Models (VLMs), directly addressing critical AGI research challenges. Our central aim aligns with the SingularityNET RFP's call to explore neural-symbolic DNN architectures, embedding logic rules for improved higher-order reasoning and experiential learning, with a clear view towards integration into the PRIMUS cognitive architecture within the Hyperon AGI framework.

The core problem we address is the inherent limitation of current LLMs/VLMs in performing robust, verifiable logical inference—such as multi-hop reasoning, causal understanding, or consistent rule application—and their difficulties with output consistency, often leading to contradictions and hallucinations. Their statistical nature, while powerful for pattern matching, struggles with efficient learning from sparse symbolic knowledge crucial for dynamic AGI, and their "black box" characteristic impedes explainability and trust. Standard LLM augmentation techniques, including tool-use frameworks like LangChain or LlamaIndex which often rely on less nuanced triggering, or generic RAG, fall short of providing the deep semantic integration needed for complex task delegation or logically-grounded context injection. The development of advanced AGI systems like PRIMUS, incorporating components such as PLN, MOSES, and ECAN, necessitates a more principled integration of symbolic reasoning with neural learning to overcome these bottlenecks.

Our solution, the Model Context Protocol (MCP), is an intelligent orchestration layer enabling LLMs to seamlessly collaborate with specialized symbolic and neural-symbolic tools. MCP employs "tool contracts"—rich, structured JSON descriptors detailing a tool's function, input/output types (e.g., MeTTa expressions, natural language, graph data), semantic capabilities (e.g., "symbolic_deduction," "graph_rule_learning," "causal_inference"), and semantic trigger conditions. This allows MCP's dispatcher logic to proactively select and invoke the most appropriate tool based on a deep contextual understanding of the ongoing prompt or problem, moving beyond simple keyword matching to more accurately align tasks with tool strengths.

Two primary neural-symbolic integrations are planned within MCP:

1.  **MeTTa-Enhanced Reasoning:** The MeTTa symbolic language, a cornerstone of OpenCog Hyperon, will function as a powerful, verifiable reasoning engine. MCP will facilitate LLM-to-MeTTa translation, for which we will explore prompt chaining, few-shot examples, and potentially fine-tuning smaller LLMs for converting natural language or intermediate LLM representations into precise MeTTa expressions. The MeTTa engine then performs formal symbolic inference, providing results that augment LLM output and support higher-order reasoning. A key innovation is the **"MeTTa Digest Tool."** This tool creates an evolving symbolic knowledge base within Atomspace from LLM interactions (prompt/response pairs, validated facts, user feedback). When a new query arises, it utilizes MeTTa's graph traversal and unification capabilities to find semantically relevant historical interactions or explicitly stored rules, providing "short-circuited," highly relevant contextual information to the LLM. This addresses issues of LLM "dumbness" and consistency more effectively than standard RAG by leveraging dynamic, symbolic memory.

2.  **PyNeuraLogic for Experiential Rule Embedding (AIRIS-Inspired):** PyNeuraLogic, combining differentiable logic programming with DNNs (like GNNs), will embed and apply rules derived from experiential learning. We will leverage a unique, **existing resource**: the "vhybZ" mobile VLM application (github.com/vhybzOS/RN-vhybZ). This app, with its "vibe coding" workflow where users iteratively refine VLM-generated artifacts (code, images), is independently funded and developed by PI Keyvan Sadeghi, offering a continuous, real-world data stream. User interactions (prompts, VLM outputs, selected refinements), captured as structured MCP tool calls or observed events, will be analyzed by an AIRIS-like process to synthesize symbolic rules. For instance, observing a user consistently transforming VLM output `A` into `B` in context `C` could yield a PyNeuraLogic-compatible rule like `preference(context(C), output(B)) :- observed_transform(input(A), output(B), user_action(accept_B))`. These learned behavioral rules will be embedded using PyNeuraLogic into a neural component accessible via MCP, enabling adaptive system suggestions.

The utility lies in delivering a functional MCP framework for advanced LLM augmentation, a Proof-of-Concept demonstrating empirically enhanced LLM reasoning and experiential adaptation via MeTTa and PyNeuraLogic, and a clear methodology for AIRIS-inspired rule generation from real user data, offering actionable insights for PRIMUS. Primary performance indicators will include: improvements in LLM reasoning accuracy (e.g., on deductive benchmarks like LogicQA) and output consistency compared to baseline LLMs; qualitative and quantitative measures of interpretability via MeTTa reasoning traces and the relevance of context provided by the "MeTTa Digest Tool"; task completion rates for multi-step problems; and the efficacy (e.g., predictive accuracy of learned rules, processing efficiency) of the AIRIS-PyNeuraLogic pipeline.

This holistic approach, combining MCP's sophisticated orchestration with the deep symbolic integration of MeTTa and the practical experiential learning pipeline of "vhybZ"/AIRIS/PyNeuraLogic, forms our core USP. The project is led by **Keyvan M. Sadeghi (PI)** (MSc AI Distinction; AGI Researcher, Hong Kong PolyU with publications in cognitive architectures; extensive practical experience at Functionland as Co-founder/CEO architecting decentralized AI, deploying edge LLMs, RAG, and agentic frameworks, securing $500k+ crowdfunding & $1.1M seed). **Julien Serbanescu (Co-lead)** (AI/ML Researcher, U Guelph; RAG expert) will be vital for benchmarking, MCP-LLM interface design, and "MeTTa Digest Tool" evaluation. **Farhoud Mojahedzadeh** (MeTTa specialist; DeepFunding "Unsupervised Learning" proposal contributor) will provide core expertise for the `metta_reasoner` and Atomspace optimization. The team, supported by two software engineers for PoC development, combines essential skills in symbolic AI, ML, LLM augmentation, MeTTa, and Python.

All software developed under this grant (MCP framework, MeTTa scripts, PyNeuraLogic components) and documentation will be released under a permissive open-source license (e.g., MIT or Apache 2.0). Pre-existing components (base LLMs, PyNeuraLogic library, "vhybZ" app) retain original licenses. This research directly contributes to the RFP's goals by advancing practical neural-symbolic application for future AGI systems.

Proposal Video

Not Avaliable Yet

Check back later during the Feedback & Selection period for the RFP that is proposal is applied to.

  • Total Milestones

    3

  • Total Budget

    $40,000 USD

  • Last Updated

    27 May 2025

Milestone 1 - MCP Design NeSy Plan AIRIS Study

Description

This foundational phase focuses on the detailed architecture of the Model Context Protocol (MCP) and the integration strategy for neural-symbolic components. We will define MCP's core: tool contracts (for MeTTa reasoner PyNeuraLogic GNN component) semantic dispatching logic and context management. A comprehensive research plan for LLM-to-MeTTa translation techniques will be established. We will design the specific PyNeuraLogic component including how it will represent and process structured rules. A critical sub-task is the AIRIS applicability study: defining data structures and processes to capture "vibe coding" interactions from the "vhybZ" mobile app (user prompts VLM outputs refinement actions) as suitable input for an AIRIS-like rule generation pipeline. This includes specifying how these learned experiential rules can be translated into formats usable by PyNeuraLogic (e.g. graph structures relational logic) and/or directly by MeTTa. The plan will detail how these NeSy elements will interact within MCP to enhance LLM capabilities in reasoning and learning from experience. We will also outline the agile breakdown of tasks for subsequent milestones risk assessment and a preliminary framework design for the overall system.

Deliverables

1. Detailed Research Plan & MCP Architecture Document (PDF): * Comprehensive MCP design: protocol specifications tool contract formats semantic dispatcher logic. * Strategy for LLM-to-MeTTa translation. * Detailed design of the PyNeuraLogic component and its rule representation. * Agile breakdown of tasks for M2 & M3 with timelines. * Initial overall framework design sketch. 2. AIRIS Integration Feasibility Report & Data Specification (LaTeX): * Analysis of AIRIS applicability to "vhybZ" app data. * Detailed specification for capturing and structuring "vibe coding" interactions (prompts VLM outputs user actions) for input into an AIRIS-like rule generation process. * Proposed formats for representing AIRIS-generated rules for PyNeuraLogic and/or MeTTa. 3. Initial MeTTa & PyNeuraLogic Setup Report: Document confirming setup of development environments for MeTTa and PyNeuraLogic including library installations and basic "hello world" style tests.

Budget

$8,000 USD

Success Criterion

1. SingularityNET approval of the Research Plan and MCP Architecture, confirming methodological soundness and alignment with RFP objectives (NeSy integration, PRIMUS compatibility). 2. AIRIS Integration Feasibility Report demonstrates a clear, viable pathway for generating symbolic rules from "vhybZ" app interactions and representing them for PyNeuraLogic/MeTTa. Data specification is complete. 3. Design for the PyNeuraLogic component is well-defined and suitable for the types of rules expected from the AIRIS-like process. 4. LLM-to-MeTTa translation strategy is clearly outlined with identified techniques for development. 5. Development environments for MeTTa and PyNeuraLogic are operational.

Milestone 2 - MCP Core NeSy Tools AIRIS PoC

Description

This phase focuses on the initial development and implementation of the core MCP framework and its key neural-symbolic tool integrations. We will build the MCP dispatcher logic tool registration system and the communication protocols based on the M1 design. The `metta_reasoner` tool will be implemented including the initial version of the LLM-to-MeTTa translation module enabling basic symbolic queries to be processed. The selected PyNeuraLogic component for handling structured/experiential rules will be developed. A significant effort will be dedicated to the AIRIS-inspired Proof-of-Concept: developing scripts and methods to process sample (anonymized if necessary) "vibe coding" interaction logs from the "vhybZ" app to generate a preliminary set of symbolic rules. We will then implement the pipeline to embed these generated rules using the PyNeuraLogic component or represent them directly in MeTTa. Preliminary testing of individual components (MCP dispatcher MeTTa reasoner PyNeuraLogic module with AIRIS rules) and their basic integration via MCP will be conducted. The "MeTTa digest tool" concept will be prototyped for capturing interaction data.

Deliverables

1. **MCP Framework v0.1 Codebase (e.g. Git repository):** * Core MCP dispatcher tool registration and communication modules. * Implemented `metta_reasoner` tool with initial LLM-to-MeTTa translation capability. * Implemented PyNeuraLogic component for rule embedding. 2. **AIRIS-Inspired Rule Generation & Embedding PoC:** * Scripts/tools for processing sample "vhybZ" interaction data to generate symbolic rules. * Demonstration of these rules being embedded via PyNeuraLogic or loaded into MeTTa. * Sample set of generated rules. 3. **Initial Testing Results & Analysis Report (PDF):** * Results from unit tests of MCP components MeTTa reasoner and PyNeuraLogic module. * Report on the AIRIS rule generation PoC performance and quality of initial rules. * Analysis against standard benchmarks (if applicable at this stage) or qualitative assessment of functionality. * Draft implementation of the "MeTTa digest tool" for capturing interactions.

Budget

$16,000 USD

Success Criterion

1. MCP framework v0.1 successfully dispatches tasks to the `metta_reasoner` and the PyNeuraLogic component based on defined tool contracts and semantic triggers. 2. The `metta_reasoner` can process basic symbolic queries translated from pseudo-natural language. 3. The AIRIS-inspired PoC successfully generates a preliminary set of symbolic rules from sample "vhybZ" app data, and these rules are demonstrably usable by the PyNeuraLogic component or MeTTa. 4. Initial testing results show functional correctness of individual components and basic integration. 5. "MeTTa digest tool" prototype successfully captures interaction KPs in MeTTa. 6. Clear progress towards addressing RFP's functional requirements for NeSy DNN integration and logic rule inclusion.

Milestone 3 - Full PoC Demo Eval & Final Report

Description

This concluding phase will deliver a fully integrated Proof-of-Concept (PoC) system conduct comprehensive evaluations and produce all final documentation. The PoC will showcase an LLM/VLM utilizing the MCP framework to dynamically call upon the `metta_reasoner` (for higher-order reasoning) and the PyNeuraLogic component (processing AIRIS-inspired rules from "vhybZ" app data) to solve complex tasks or enhance its responses. This demonstration will highlight improvements in reasoning consistency and adaptation based on experiential learning. We will conduct a thorough evaluation using the performance metrics defined in M1 comparing the MCP-augmented LLM against baseline LLM performance. The functionality and impact of the "MeTTa digest tool" on context relevance and knowledge evolution will be assessed. This phase includes detailed analysis of how the system addresses human interpretability learning from small data bridging data-driven/symbolic reasoning and applicability to structured learning domains. All code will be finalized and documented. The final report will encompass the project's research design implementation evaluation and recommendations for future PRIMUS integration and further research.

Deliverables

1. **Final Comprehensive Report (PDF):** * Complete project overview: architecture methodologies implementation details. * Detailed performance analysis based on defined metrics including comparison with baselines. * Evaluation of how the system addresses RFP criteria: interpretability small data learning bridging NeSy gap structured learning. * Analysis of the "MeTTa digest tool" and AIRIS-PyNeuraLogic pipeline. * Recommendations for PRIMUS integration and future research directions. 2. **Fully Integrated Proof-of-Concept (PoC) System & Demonstration:** * Complete documented codebase (Python MeTTa scripts) for the MCP framework MeTTa reasoner PyNeuraLogic component AIRIS rule generation pipeline and "MeTTa digest tool." * A live or recorded demonstration of the PoC system showcasing an end-to-end LLM task enhanced by MCP-orchestrated NeSy components. 3. **Complete Project Documentation:** System architecture diagrams user/developer guides for the framework and tools and setup instructions. 4. (If applicable) Websites or other dissemination materials.

Budget

$16,000 USD

Success Criterion

1. Submission of a high-quality Final Report that comprehensively details all research, development, and evaluation activities, and directly addresses all key RFP objectives and evaluation criteria. 2. Successful demonstration of the fully integrated PoC system, clearly showing an LLM/VLM leveraging MeTTa and PyNeuraLogic (with AIRIS-derived rules) via MCP to achieve improved reasoning, consistency, or experiential adaptation on defined tasks. 3. The PoC effectively illustrates the core neural-symbolic concepts explored (logic rule embedding, reasoning over DNN outputs, experiential learning). 4. All code is well-documented, functional, and reproducible as per RFP requirements. 5. The project provides clear evidence and analysis regarding improvements in interpretability, learning from small data, and bridging the data-driven/symbolic gap. 6. Actionable recommendations for PRIMUS integration are provided. 7. All deliverables are accepted by SingularityNET.

Join the Discussion (0)

Expert Ratings

Reviews & Ratings

    No Reviews Avaliable

    Check back later by refreshing the page.

Welcome to our website!

Nice to meet you! If you have any question about our services, feel free to contact us.