🧠 SL5 Aura Hybrid Mode: Local LLM Integration

Status: Experimental / Stable Technology: Ollama (Llama 3.2) + Python Subprocess Privacy: 100% Offline

The Concept: “Architect & Intern”

Traditionally, Aura relies on deterministic rules (RegEx) – fast, precise, and predictable. This is the “Architect”. However, sometimes the user wants to ask something “fuzzy” or creative, like “Tell me a joke” or “Summarize this text”.

This is where the Local LLM Plugin comes in (the “Intern”):

  1. Aura (RegEx) first checks all strict commands (“Turn on lights”, “Open App”).

  2. If nothing matches AND/ OR a specific trigger word (e.g., “Aura …”) is detected, the fallback rule activates.

  3. The text is sent to a local AI model (Ollama).

  4. The response is sanitized and output via TTS or text typing.

🛠 Prerequisites

The plugin requires a running instance of Ollama operating locally on the machine.

# Installation (Arch/Manjaro)
sudo pacman -S ollama
sudo systemctl enable --now ollama

# Download model (Llama 3.2 3B - only ~2GB, very fast)
ollama run llama3.2

📂 Structure & Load Order

The plugin is intentionally placed in the folder z_fallback_llm. Since Aura loads plugins alphabetically, this naming ensures that the LLM rule is loaded last. It serves as a “safety net” for unrecognized commands.

Path: config/maps/plugins/z_fallback_llm/de-DE/

1. The Map (FUZZY_MAP_pre.py)

We use a high score (100) and a trigger word to force Aura to hand over control to the script.

import re
from pathlib import Path
CONFIG_DIR = Path(__file__).parent

FUZZY_MAP_pre = [
    # Trigger: "Aura" + any text
    ('ask_ollama', r'^\s*(Aura|Aurora|Laura)\s+(.*)$', 100, {
        'command_flags': re.IGNORECASE,
        # 'skip_list': ['LanguageTool'], # Optional: Performance boost
        'on_match_exec': [CONFIG_DIR / 'ask_ollama.py']
    }),
]

2. The Handler (ask_ollama.py)

This script communicates with the Ollama CLI. Important: It contains a clean_text_for_typing function. Raw LLM outputs often contain emojis (😂, 🚀) or special characters that can crash tools like xdotool or legacy TTS systems.

# Snippet from ask_ollama.py
def execute(match_data):
    # ... (Regex group extraction) ...
    
    # System prompt for short answers
    system_instruction = "Answer in German. Max 2 sentences. No emojis."
    
    # Subprocess call (blocks briefly, note the timeout!)
    cmd = ["ollama", "run", "llama3.2", full_prompt]
    result = subprocess.run(cmd, capture_output=True, ...)

    # IMPORTANT: Sanitize output for system stability
    return clean_text_for_typing(result.stdout)

⚙️ Customization Options

Changing the Trigger

Modify the RegEx in FUZZY_MAP_pre.py if you don’t want to use “Aura” as the wake word.

  • Example for a true Catch-All (everything Aura doesn’t know): r'^(.*)$' (Caution: Adjust the score!)

Swapping the Model

You can easily swap the model in ask_ollama.py (e.g., to mistral for more complex logic, though it requires more RAM).

cmd = ["ollama", "run", "mistral", full_prompt]

System Prompt (Persona)

You can give Aura a personality by adjusting the system_instruction:

“You are a sarcastic assistant from a Sci-Fi movie.”

⚠️ Known Limitations

  1. Latency: The very first request after boot might take 1-3 seconds as the model loads into RAM. Subsequent requests are faster.

  2. Conflicts: If the RegEx is too broad (.*) without a proper folder structure, it might swallow standard commands. The alphabetical ordering (z_...) is essential.

  3. Hardware: Requires approx. 2GB of free RAM for Llama 3.2.