from synalinks-skills898
Keras-inspired framework for building structured, neuro-symbolic LLM programs with DataModel schemas, modular Programs, and training/optimization tools.
Synalinks provides a Keras-like API for constructing structured LLM applications: define DataModel schemas, compose Modules and Programs, run Generators and Decisions, and train with optimizer wrappers. It focuses on structured JSON I/O, parallel branching patterns, and optimizer-driven prompt/example evolution for reliable outputs. Concrete use case: build a QA program that defines input/output schemas, runs multiple generator branches for self-consistency, and merges results with guards and rewards.
Use Synalinks when you need predictable, schema-driven LLM outputs (RAG, function-calling, or tool-enabled agents), or when you want to treat LLM pipelines like trainable programs (optimizers, rewards, and program checkpoints). Ideal for structured-output tasks, RAG/KAG workflows, and programmatic agent tooling.
Targets LLM-based agent runtimes and tool-enabled agents (OpenAI-style chat models, Claude, Ollama/Mistral local runtimes, OpenRouter-backed models).
Synalinks is a Keras-inspired neuro-symbolic LLM framework with DataModel schemas, modular Programs, and training/optimization tools. The SKILL.md is comprehensive with detailed examples of branching, guards, agents, training, and multiple provider patches. Scripts are well-written demos but all require synalinks installed (not on PyPI by default), making them unrunnable without setup. The restricted eval in tool_agent.py is a minor security note but properly sandboxed.
synalinksnumpyWell-crafted skill with extensive documentation. The eval() in tool_agent.py is sandboxed with __builtins__=None and input validation, but still a pattern to flag. 7/8 scripts fail purely due to missing synalinks package, not code quality issues. The one succeeding script (subclassing_modules.py) is a 3-line stub.
