LLM Tools

aria.llmtools is ARIA’s provider and routing layer for LLM inference across online and local backends. It is also used by verification workflows such as the LLM4Inv engine, but it is no longer nested under aria.efmc.

Package layout

The current package structure is:

aria/llmtools/
├── __init__.py
├── client.py
├── local_client.py
├── routing.py
├── tooling.py
├── core/
└── providers/

Key modules:

client.py: routed LLM client
local_client.py: local-provider client helpers
routing.py: provider/model resolution
tooling.py: abstract base classes for LLM-backed tools
core/: shared types, logging, retry logic, and client flow
providers/: online and local provider adapters

Public API

The package exports:

from aria.llmtools import LLM, LLMLocal, Logger, resolve_provider

Example

from aria.llmtools import LLM, Logger

logger = Logger()
llm = LLM(
    model_name="gpt-4o-mini",
    logger=logger,
    temperature=0.1,
)

The concrete provider is selected by resolve_provider() based on the model name and optional provider hint.

LLM-backed tool base classes

aria.llmtools.tooling defines reusable abstractions for higher-level tools:

from aria.llmtools.tooling import LLMTool, LLMToolInput, LLMToolOutput

These classes provide:

a typed input/output boundary for tool wrappers
retry-aware invocation through the routed LLM client
prompt construction and response parsing hooks
simple caching keyed by tool input

Relationship to EFMC

Verification engines may consume aria.llmtools, but the LLM infrastructure is shared repo-wide and should be documented at the top-level package location.