accuralai-core

accuralai-core is the orchestration nucleus for the AccuralAI open-source ecosystem. It provides an async pipeline that coordinates canonicalization, caching, routing, backend invocation, validation, and post-processing for local LLM text generation workflows.

This package exposes both a Python API and CLI surface while remaining transport-agnostic. Concrete functionality (canonicalizers, caches, routers, backends, validators) is supplied by sibling packages or third-party plugins via entry-point discovery.

Quick start

Install with the canonicalizer plugin for a fully working local pipeline:

pip install accuralai-core accuralai-canonicalize

Generate completions via the CLI:

accuralai-core generate \
  --prompt " Summarize   this text.  " \
  --system-prompt "You are a precise assistant." \
  --tag demo --metadata topic=news --param temperature=0.3

The accuralai-canonicalize plugin normalizes the prompt, merges metadata defaults, and creates deterministic cache keys before the request is routed to the configured backend.

Add accuralai-cache to enable TTL-aware in-memory caching:

pip install accuralai-cache

With caching enabled, repeat invocations are served instantly until the configured TTL expires.

See plan/accuralai-core-spec.md for the full architectural specification guiding implementation.

Interactive CLI

Launch the Codex-style REPL by invoking accuralai (or accuralai-cli) with no subcommand:

accuralai --config ~/.accuralai/core.toml

The shell remains active until /exit, and supports / commands for adjusting settings on the fly (/help, /backend, /model, /meta, /history, /save, /tool ..., etc.). Plain text input triggers pipeline runs using the current session defaults, and multi-line prompts are accepted by entering """ on an empty line. Tools can be listed via /tool list, executed manually with /tool run ..., and made available to the model for function calling via /tool enable <name>.