aions-llm 1.0.7

Actions and Interface Object Notation: An open standard for LLM tools.

AIONS: Actions and Interface Object Notation

AIONS (pronounced IONS) is an open-standard, text-based data format engineered specifically for Large Language Model (LLM) agent architectures. By decoupling tool definitions from application logic, AIONS provides a portable, JSON-like notation while retaining the full execution capabilities of native Python.

Installation

pip install aions-llm

Architecture and Rationale

Traditional AI tool registration often pollutes backend Python code with extensive prompt strings, interface definitions, and repetitive boilerplate. AIONS modernizes this workflow through the following architectural principles:

• Logic Decoupling: Isolates LLM prompts (descriptions) and tool mappings into external .aion configuration files.
• Dynamic Evaluation: Safely parses and binds native Python lambdas and complex conditional logic directly from the notation into memory.
• Strict Validation: Enforces a rigid property schema during the parsing phase to prevent runtime failures, "ghost tools," or broken agent execution.
• Zero-Map Architecture: Eliminates the need to maintain manual dictionaries in the application layer. If a tool is defined in the .aion file, the framework autonomously mounts it to your Agent.

---

Specification and Constraints

To maintain standard integrity, every .aion file must adhere strictly to the following validation rules:

1. The Array Constraint

An AIONS definition must always manifest as a root-level array, enclosed in square brackets [ ]. Single-tool definitions must still reside within this array.

2. The Assignment Operator

Properties are assigned exclusively using the "Action-Link" operator: -->.

• Valid: name --> "AuthModule"
• Invalid: name: "AuthModule" or name = "AuthModule"

3. Strict Property Isolation

AIONS enforces a strict, closed-property ecosystem. If the parser encounters a top-level key not present in the Approved Registry, it will raise an AIONPropertyError.

• Approved Registry: name, function, description, args_schema, link.

4. The Function/Link Dependency Rule

Every valid AION element must possess at least one executable or referential parameter. An element must declare a function, a link (to feed public documentation to the AI), or both. If neither property is present, the parser will fail.

5. The Interface Block

When utilizing the function property, the value must be a raw string representing the executable (e.g., a lambda or function name), immediately followed by an Interface Block { } that maps the inputs and outputs.

• Inputs must follow the sequential arg-N pattern.
• Outputs must follow the sequential return-N pattern.

---

Syntax Variations

The following examples demonstrate the flexibility of the Function/Link Dependency Rule. An element can act as a direct execution tool, a documentation reference, or a hybrid of both.

Variation A: Function-Only Execution

Used when the agent needs to directly execute backend logic without requiring external documentation.

[
  {
    name --> "SendOutput",
    function --> "lambda user: send_output(user)" --> {
        arg-1 --> "string (user identifier)",
        return-1 --> "dict (execution status)"
    },
    args_schema --> "SendOutputSchema",
    description --> "Executes the output transmission to the specified user."
  }
]

Variation B: Link-Only Reference

Used when the agent needs access to external API documentation or a web resource, but no direct backend Python execution is required. Note: The framework automatically binds a fallback function that returns this URL to the LLM.

[
  {
    name --> "FetchAbcDocs",
    link --> "[https://api.Abc.com/docs](https://api.Abc.com/docs)",
    description --> "Retrieves the public documentation for the Abc API to understand endpoint structures."
  }
]

Variation C: Hybrid Execution & Reference

Used for complex tools where the agent can execute the function, but is also provided a link to the relevant documentation to understand the broader context of the action.

[
  {
    name --> "AdminSignup",
    function --> "lambda params: admin_signup(*params.split('|'))" --> {
        arg-1 --> "string (pipe-separated user data)",
        return-1 --> "dict (new user profile)"
    },
    link --> "[https://api.Abc.com/docs/admin_signup](https://api.Abc.com/docs/admin_signup)",
    description --> "Creates a new admin account. Refer to the provided documentation link for strict password policies."
  }
]

---

Implementation Guide

1. Recommended Directory Structure

To ensure clean modularity, isolate your notation files in a dedicated directory.

project_root/
├── aion_tools/         # Directory for AIONS notation files
│   ├── auth.aion
│   └── users.aion
├── schemas.py          # Pydantic validation models
├── api_functions.py    # Core application logic
└── agent_factory.py    # LangChain entry point

2. LangChain Integration

AIONS is engineered to natively compile into LangChain Tool objects. By passing globals() into the execution context, the parser autonomously binds the strings in your .aion files to the functions and schemas residing in your application's memory.

from aions import AIONS
import api_functions
from schemas import SendOutputSchema

# Initialize the AIONS parser with the local execution context
tool_registry = AIONS.get_langchain_tools(
    source_path="aion_tools/", 
    context=globals()
)

# The resulting list can be injected directly into a LangChain Agent
# agent = initialize_agent(tools=tool_registry, llm=model, ...)

---

Exception Reference

The AIONS parser is designed to fail fast. Strict validation ensures the LLM is never provided with malformed tool schemas.

Exception Class	Trigger Condition
AIONPropertyError	Raised when an unauthorized key is declared (e.g., using desc instead of the approved description property).
AIONParseError	Raised upon detecting syntax violations, missing array brackets, failure to meet the Function/Link Dependency Rule, or if a declared function/schema does not exist in the provided execution context.

---

API Reference

The AIONS class exposes several static methods to accommodate various system architectures:

• AIONS.get_langchain_tools(source_path, context): High-level factory method that parses files and compiles them directly into LangChain Tool instances.
• AIONS.load_dir(dirpath, context): Scans a target directory and compiles all contained .aion files into a unified Python dictionary list.
• AIONS.loads(aion_text, context): Parses a raw AIONS string directly from memory, bypassing the file system.
• AIONS.dumps(tools_list): Serializes an existing list of Python tool dictionaries back into the standard AIONS text format.

---

Developed by Sourav Modak