ntep 0.1.0

Neuro‑symbolic Tool Composition via Spherical Embeddings

NTEP – Expanded Exposition

Neuro‑symbolic Tool Composition via Spherical Embeddings
Author: J. Roberto Jiménez ✉️ [e-mail]
DOI: https://doi.org/10.5281/zenodo.18398188

NTEP: Neural Tool Embeddings Protocol

What if your AI agent tried to email a file before encrypting it? NTEP builds a mathematical guarantee to prevent that.

As AI agents take on more complex tasks, tool chaining becomes critical. But how do you ensure an AI doesn't execute a sequence of tools in a disastrously wrong order? Traditional methods rely on post-hoc checks, which can fail.

NTEP (Neural Tool Embeddings Protocol) is a framework that builds safety directly into the AI's "thinking." It embeds tools in a vector space where their geometric relationships encode their operational dependencies.

🔑 Key Features

- Soundness-by-Construction: Any valid chain decoded from the embeddings is automatically correct. No more guessing.
- Monotonicity Guarantee: Our central theorem proves the AI is always geometrically "nudged" toward the correct next step.
- Risk Awareness: A quantum-inspired "consciousness" measure quantifies uncertainty and flags risky chains before execution.
- Built for Scale: Proven generalization bounds and efficient retrieval make it viable for enterprise applications.

🚀 Quick Start

pip install ntep

from ntep import NTEPDecoder, Tool, DependencyGraph

# 1. Define your tools and their dependencies
tools = [
    Tool(id="resize", name="Resize Image"),
    Tool(id="encrypt", name="Encrypt File"),
    Tool(id="package", name="Package for Delivery"),
    Tool(id="email", name="Send via Email")
]
dependencies = {
    "resize": ["package"],
    "encrypt": ["package"],
    "package": ["email"]
}
dep_graph = DependencyGraph(dependencies)

# 2. Load a pre-trained decoder (or train your own)
# This loads the tool embeddings and dependency graph
decoder = NTEPDecoder.from_pretrained("ntep/base-document-pipeline")

# 3. Decode a valid chain from a prompt
prompt_embedding = decoder.encode_prompt("prepare the confidential image for delivery")
chain = decoder.decode(prompt_embedding)

print(f"Valid tool chain: {' -> '.join(chain)}")
# > Valid tool chain: resize -> encrypt -> package -> email

---

📖 Documentation

The core technical narrative lives in docs/axioms.pdf. It expands on:

Symbolic definitions (𝒯, ℰ, ≤, γ, δ…)
Category‑theoretic construction (𝒫, metric enrichment, embedding functor)
Monotonicity Theorem with full proof sketch
Consciousness measure built from density‑matrix entropy
Generalization bounds, geometric properties, and generalization guarantees

Contributing

Feel free to open issues or pull requests if you spot a typo, need a visualisation, or want to extend the framework (e.g., stochastic composition, higher‑order interactions).

---

📜 Citation

If you use NTEP in research, please cite the Zenodo record:

@misc{jimenez2023ntep,
  author       = {Jiménez, J. Roberto},
  title        = {{NTEP}: Expanded Exposition},
  year         = {2026},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.18398188},
  url          = {https://doi.org/10.5281/zenodo.18398188}
}

📧 Contact

For questions or collaboration proposals, reach out to J. Roberto Jiménez at [e-mail].

---

If you would like to support the development of these resources, consider contributing towards helping me get some gear for continued improvement or simply treating me to a coffee. Your support means a lot!" buy me a coffee