ai-mini-box-llm

LLM plugin for ai-mini-box-core — adds natural language processing: message classification, response drafting, entity extraction, and RAG.

Features

• Classify messages by topic (Prices, Order, Complaint, Schedule, Other)
• Draft auto-responses to customer messages
• Extract entities (phone, name, address, date, order ID)
• RAG — Retrieval-Augmented Generation from Knowledge Base
• Two providers: local (GGUF via llama-cpp-python) or remote (OpenAI API)
• CLI commands for status, testing, model download, KB indexing

Installation

pip install ai-mini-box-llm[local]   # local inference
pip install ai-mini-box-llm[remote]  # OpenAI API

Configuration

Config file: data/llm_config.json (auto-created with defaults):

{
  "provider": "local",
  "model_path": "data/models/Phi-3-mini-q4.gguf",
  "n_ctx": 4096,
  "n_threads": 4,
  "rag_enabled": false,
  "rag_top_k": 3
}

Usage

# Show status
ai-mini-box llm status

# Classify a message
ai-mini-box llm classify "сколько стоит доставка?"

# Generate a draft response
ai-mini-box llm draft "хочу заказать пиццу" --topic "Заказ"

# Extract entities
ai-mini-box llm extract "Позвоните Ивану по телефону +7 123 456 78 90"

# Download a model
ai-mini-box llm download-model Qwen/Qwen2.5-0.5B-Instruct-GGUF:q4_0

# Rebuild RAG index
ai-mini-box llm ingest-kb

Architecture

The plugin registers an LlmService implementation via the generic service registry:

from ai_mini_box.core.services.registry import register_service, get_service

# Registration (done by plugin automatically)
register_service("llm", MyLlmService())

# Usage by any other module
llm = get_service("llm")
if llm:
    topic = llm.classify("How much?")

When the plugin is not installed, get_service("llm") returns None, and the core falls back to built-in heuristics (KeywordClassifier, regex extraction).

Requirements

• Python 3.12+
• ai-mini-box-core >= 5.0.0
• Optional: llama-cpp-python, openai, huggingface-hub