SynaLinks/synalinks

From idea to production in just few lines: Graph-Based Programmable Neuro-Symbolic LM Framework - a production-first LM framework built with decade old Deep Learning best practices

What it solves

Synalinks is a neuro-symbolic framework designed to simplify the creation, training, and deployment of advanced language model (LM) applications. It removes the boilerplate associated with building complex systems like RAGs, autonomous agents, and self-evolving reasoning systems, providing a clean, declarative API similar to Keras for deep learning.

How it works

The framework uses a modular composition approach where developers compose Modules (similar to deep learning layers) into Programs. It leverages a neuro-symbolic approach, combining logic, structure, and language models.

Key technical features include:

  • In-Context Reinforcement Learning: Optimizes prompts and trainable variables using a .compile() / .fit() / .evaluate() / .predict() API without altering model weights.
  • Embedded Runtimes: Includes a container-free, pure-Python sandbox for agents to run untrusted code safely and embedded graph/SQL databases for knowledge bases.
  • Model Agnostic: Integrates with LiteLLM to allow seamless switching between numerous providers (OpenAI, Anthropic, Gemini, etc.) and supports multi-objective model selection for cost/quality balance.
  • Structured Outputs: Uses Pydantic-based data models to ensure constrained JSON outputs for correctness.

Who it’s for

  • AI Developers: To build production-grade LM apps without boilerplate.
  • AI Researchers: To prototype neuro-symbolic and RL-in-context systems quickly.
  • Data Scientists: To integrate LM workflows with APIs and databases.
  • Students and Hobbyists: To learn AI composition through an intuitive framework.

Highlights

  • Container-free Sandbox: Safe, isolated runtime for agents that requires no Docker or external services.
  • Embedded Databases: Built-in support for graph-based RAG and SQL knowledge bases without separate servers.
  • Keras-like API: Familiar deep learning workflow for optimizing prompts and variables via in-context RL.
  • Production Ready: Built-in observability via MLflow, async/parallel execution, and deployment options for FastAPI or FastMCP.