app_kit/llm.py

import os
import logging
from pathlib import Path

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

_model = None

def get_llm():
    global _model
    if _model is not None:
        return _model
    
    try:
        from llama_cpp import Llama
    except ImportError:
        raise RuntimeError("llama-cpp-python is required for model inference. Please install it.")

    import os
    fallback_paths = [
        Path("/data/MiniCPM5-1B-Q4_K_M.gguf"),
        Path(__file__).resolve().parent.parent / "models" / "MiniCPM5-1B-Q4_K_M.gguf"
    ]
    model_path = fallback_paths[1]
    for p in fallback_paths:
        if p.exists():
            model_path = p
            break
    if not model_path.exists():
        raise RuntimeError(f"Model not found at {model_path}. Model inference requires a valid model.")
    
    logger.info(f"Loading model from {model_path}")
    _model = Llama(model_path=str(model_path), n_ctx=2048, n_threads=4, verbose=False)
    return _model

def generate_repair_checklist(symptom: str, equipment_type: str, location: str, notes: str, photo_caption: str, references: list[str], insufficient: bool) -> tuple[str, dict]:
    llm = get_llm()
    
    refs_text = "\n\n".join([f"Reference {i+1}:\n{ref}" for i, ref in enumerate(references)]) if references else "None available."
    
    if insufficient:
        prompt = f"""You are an expert field repair assistant. The user provided insufficient information to make a confident diagnosis.

Equipment: {equipment_type}
Location: {location}
Symptom: {symptom}
Notes: {notes}
Photo Context: {photo_caption}

Explain that there is insufficient evidence and suggest next steps to diagnose the issue (e.g. check meter, fault code, inspect obvious hazards). Start with a safety reminder.
Response:"""
    else:
        prompt = f"""You are an expert field repair assistant. Based on the symptom, equipment type, location, notes, and the provided manual references, create a short, actionable checklist for diagnosing and fixing the issue. Always start with a safety reminder.
Do NOT use hallucinated information. Rely strictly on the references provided.

Equipment: {equipment_type}
Location: {location}
Symptom: {symptom}
Notes: {notes}
Photo Context: {photo_caption}

References:
{refs_text}

Provide the response as a bulleted checklist starting with "Safety reminder: "
Checklist:"""

    response = llm(
        prompt,
        max_tokens=300,
        temperature=0.3,
        stop=["\n\n\n"]
    )
    
    text = response['choices'][0]['text'].strip()
    stats = {
        "model_id": "nvidia/NeMoTRON-3-Nano-4B-Instruct",
        "adapter": "llama_cpp",
        "prompt_tokens": response['usage']['prompt_tokens'],
        "completion_tokens": response['usage']['completion_tokens'],
        "total_tokens": response['usage']['total_tokens']
    }
    logger.info(f"Inference complete. Stats: {stats}")
    
    return text, stats