Upload 8 files

MODULAR_DEPLOYMENT.md

@@ -0,0 +1,399 @@
+# 🎉 VoiceNote AI - Modular Version READY!
+
+## ✅ الخطأ اتصلح + الكود أصبح منظم!
+
+---
+
+## 🔧 **المشكلة اللي كانت**
+
+```
+IndentationError: unexpected indent at line 1071
+```
+
+**السبب**: HTML زيادة في ملف واحد كبير (1100+ سطر!)
+
+---
+
+## ✅ **الحل: تقسيم الكود لملفات متعددة!**
+
+### **الهيكل الجديد:**
+
+```
+voicenote-ai/
+├── config.py              # ⚙️ كل الإعدادات
+├── gdpr_filter.py         # 🔒 GDPR anonymization
+├── models.py              # 🤖 MistralClient + ASRModel
+├── vips_classifier.py     # 📋 VIPS classification + prompts
+├── utils.py               # 🛠️ WER calculator + formatters
+├── app_modular.py         # 🎨 Gradio interface (MAIN)
+└── requirements.txt       # 📦 Dependencies
+```
+
+---
+
+## 📁 **شرح كل ملف**
+
+### **1. config.py** ⚙️ (Configuration)
+```python
+class Config:
+    ASR_MODEL = "openai/whisper-small"
+    PROMPT_TECHNIQUE = "few_shot"  # أو "chain_of_thought"
+    LLM_MODEL = "mistral-large-latest"
+    # ... إلخ
+```
+
+**الفائدة**: غيّر الإعدادات من مكان واحد!
+
+---
+
+### **2. gdpr_filter.py** 🔒 (GDPR Anonymization)
+```python
+class GDPRFilter:
+    @classmethod
+    def anonymize(cls, text: str) -> str:
+        # Anonymize personnummer, phone, email, etc.
+        ...
+```
+
+**الفائدة**: GDPR logic منفصل ومرتب!
+
+---
+
+### **3. models.py** 🤖 (AI Models)
+```python
+class MistralClient:
+    def chat(self, prompt: str) -> str:
+        # HTTP API call to Mistral
+        ...
+
+class ASRModel:
+    def transcribe(self, audio_path: str) -> str:
+        # Whisper transcription
+        ...
+```
+
+**الفائدة**: كل model في class منفصل!
+
+---
+
+### **4. vips_classifier.py** 📋 (VIPS Classification)
+```python
+class VIPSClassifier:
+    def build_prompt_few_shot(text: str) -> str:
+        # Few-shot prompting
+        ...
+    
+    def build_prompt_chain_of_thought(text: str) -> str:
+        # Chain-of-Thought prompting
+        ...
+```
+
+**الفائدة**: Prompt logic واضح ومنظم!
+
+---
+
+### **5. utils.py** 🛠️ (Utilities)
+```python
+class WERCalculator:
+    def calculate(reference, hypothesis) -> float:
+        ...
+
+def formatera_vips_html(vips: dict) -> str:
+    ...
+
+def berakna_sus(*svar) -> str:
+    ...
+```
+
+**الفائدة**: Helper functions في مكان واحد!
+
+---
+
+### **6. app_modular.py** 🎨 (Main Interface)
+```python
+from config import Config
+from models import MistralClient, ASRModel
+from vips_classifier import VIPSClassifier
+from utils import *
+
+# Initialize
+mistral = MistralClient()
+asr = ASRModel()
+vips = VIPSClassifier(mistral)
+
+# Gradio interface
+demo = gr.Blocks(...)
+demo.launch()
+```
+
+**الفائدة**: الملف الرئيسي صغير وواضح! (300 سطر بدل 1100!)
+
+---
+
+## 🚀 **كيف تستخدمه؟**
+
+### **الطريقة 1: Local Testing** 💻
+
+```bash
+# 1. ضع كل الملفات في مجلد واحد
+voicenote-ai/
+├── config.py
+├── gdpr_filter.py
+├── models.py
+├── vips_classifier.py
+├── utils.py
+├── app_modular.py
+└── requirements.txt
+
+# 2. Install dependencies
+pip install -r requirements.txt
+
+# 3. Set API key
+export MISTRAL_API_KEY="your-key-here"
+
+# 4. Run
+python app_modular.py
+```
+
+---
+
+### **الطريقة 2: HuggingFace Spaces** ☁️ (RECOMMENDED)
+
+#### **خطوات Deploy:**
+
+**1. إنشاء Space جديد**
+```
+Create Space → Name: voicenote-ai → Type: Gradio → SDK: Gradio
+```
+
+**2. رفع الملفات**
+```
+Upload ALL files to Space:
+✓ config.py
+✓ gdpr_filter.py  
+✓ models.py
+✓ vips_classifier.py
+✓ utils.py
+✓ app_modular.py (rename to app.py!)
+✓ requirements.txt
+```
+
+**⚠️ مهم جداً**: غيّر اسم `app_modular.py` إلى `app.py` قبل رفعه!
+
+**3. ضبط الإعدادات**
+```
+Settings → Secrets → Add:
+Name: MISTRAL_API_KEY
+Value: [your-key]
+```
+
+**4. اختر Hardware**
+```
+Settings → Hardware → T4 small GPU
+```
+
+**5. استنى Build**
+```
+Build time: 3-5 دقائق
+```
+
+**6. ✅ Done!**
+
+---
+
+## 🎛️ **تغيير Prompt Technique**
+
+### **في config.py** (سطر ~31):
+
+```python
+# For Few-shot Prompting:
+PROMPT_TECHNIQUE = "few_shot"
+
+# For Chain-of-Thought Prompting:
+PROMPT_TECHNIQUE = "chain_of_thought"
+```
+
+**Save → Upload → Rebuild**
+
+---
+
+## 🎯 **الميزات**
+
+### **✅ ماذا تحسّن؟**
+
+| قبل | بعد |
+|-----|-----|
+| ❌ ملف واحد 1100+ سطر | ✅ 6 ملفات منظمة |
+| ❌ صعب الصيانة | ✅ سهل الصيانة |
+| ❌ صعب فهم الكود | ✅ واضح ومنطقي |
+| ❌ IndentationErrors | ✅ Clean code! |
+
+### **✅ الميزات الجديدة**
+
+1. **Modular Design** - كل شي في مكانه!
+2. **Easy Configuration** - غيّر settings من config.py
+3. **Research Ready** - Few-shot vs CoT جاهز!
+4. **Clean Code** - Professional structure
+5. **Bug Fixed** - IndentationError اتحلّ!
+
+---
+
+## 📊 **للبحث: كيف تستخدمه؟**
+
+### **Study 1: Prompt Comparison**
+
+**Step 1: Few-shot**
+```python
+# في config.py
+PROMPT_TECHNIQUE = "few_shot"
+```
+→ Upload → Run 20 scenarios → Export CSV
+
+**Step 2: Chain-of-Thought**
+```python
+# في config.py
+PROMPT_TECHNIQUE = "chain_of_thought"
+```
+→ Upload → Run same 20 scenarios → Export CSV
+
+**Step 3: Compare**
+```
+Compare results_few_shot.csv vs results_cot.csv
+```
+
+---
+
+## 🆘 **Troubleshooting**
+
+### **Error: "cannot import config"**
+```
+✅ تأكد أن كل الملفات في نفس المجلد!
+```
+
+### **Error: "MISTRAL_API_KEY not found"**
+```
+✅ Settings → Secrets → Add MISTRAL_API_KEY
+```
+
+### **Error: "Module not found"**
+```
+✅ تأكد من requirements.txt موجود ومحدّث
+```
+
+---
+
+## 📦 **requirements.txt**
+
+```
+# Core ML/AI frameworks
+torch
+transformers
+accelerate
+safetensors
+
+# HTTP requests (for Mistral API)
+requests
+
+# Speech processing
+jiwer
+librosa
+soundfile
+
+# Web interface
+gradio
+
+# Utilities
+python-dotenv
+
+# HuggingFace Spaces
+spaces
+```
+
+---
+
+## 💡 **نصائح مهمة**
+
+### **1. استخدم الـ Modular Version!**
+أسهل بكثير للصيانة والتطوير.
+
+### **2. اختبر محلياً أولاً**
+قبل ما ترفع على HuggingFace.
+
+### **3. راجع config.py**
+قبل كل deployment - تأكد الإعدادات صح!
+
+### **4. احفظ نسخ احتياطية**
+من كل الملفات قبل التعديل.
+
+---
+
+## ✅ **Checklist للـ Deployment**
+
+### **قبل Deploy:**
+- [ ] كل الملفات جاهزة (6 files)
+- [ ] app_modular.py → app.py (renamed)
+- [ ] requirements.txt محدّث
+- [ ] config.py مضبوط صح
+
+### **أثناء Deploy:**
+- [ ] رفع كل الملفات
+- [ ] MISTRAL_API_KEY في Secrets
+- [ ] Hardware = T4 small GPU
+- [ ] Build نجح (no errors)
+
+### **بعد Deploy:**
+- [ ] Test recording
+- [ ] Test text input
+- [ ] Check VIPS output
+- [ ] Export CSV works
+- [ ] Prompt technique badge shows
+
+---
+
+## 🎓 **للتقرير**
+
+### **ذكر في Methodology:**
+
+```
+4.4 Systemarkitektur
+
+Systemet implementerades med en modulär arkitektur för att 
+underlätta underhåll och vidareutveckling:
+
+- config.py: Centraliserad konfiguration
+- gdpr_filter.py: GDPR-anonymisering
+- models.py: AI-modeller (Whisper, Mistral)
+- vips_classifier.py: VIPS-klassificering med prompttekniker
+- utils.py: Hjälpfunktioner och exportverktyg
+- app.py: Gradio-gränssnitt
+
+Denna struktur möjliggjorde enkel jämförelse mellan olika 
+prompt-tekniker genom att endast ändra en parameter i 
+konfigurationsfilen.
+```
+
+**Professionellt och vetenskapligt!** ✅
+
+---
+
+## 🎉 **Du är redo!**
+
+### **Next Steps:**
+
+1. ✅ Ladda ner alla 6 filer
+2. ✅ Byt namn app_modular.py → app.py
+3. ✅ Ladda upp på HuggingFace
+4. ✅ Lägg till MISTRAL_API_KEY
+5. ✅ Vänta på build
+6. ✅ Börja testa!
+
+---
+
+**Modular = Professional = Easy to maintain!** 💪
+
+**Lycka till med projektet!** 🎓🚀
+
+---
+
+VoiceNote AI - Modular Version  
+Högskolan Väst · 2025

app_modular.py

@@ -0,0 +1,353 @@
+"""
+VoiceNote AI — Modular Version
+================================
+Automatisk VIPS-dokumentation från svenska patientsamtal.
+
+Modular structure for easier maintenance:
+- config.py: All configuration
+- gdpr_filter.py: GDPR anonymization
+- models.py: MistralClient + ASRModel
+- vips_classifier.py: VIPS classification with prompt techniques
+- utils.py: WER calculator, formatters, export functions
+- app.py: Gradio interface (this file)
+"""
+
+import os
+import time
+import logging
+import gradio as gr
+from datetime import datetime
+
+# Import our modules
+from config import Config
+from models import MistralClient, ASRModel
+from vips_classifier import VIPSClassifier
+from utils import (
+    WERCalculator,
+    formatera_vips_html,
+    wer_badge,
+    formatera_historik_html,
+    spara_nedladdning,
+    exportera_csv,
+    berakna_sus,
+    berakna_nasa
+)
+
+# ══════════════════════════════════════════════════════════
+# LOGGING SETUP
+# ══════════════════════════════════════════════════════════
+
+logging.basicConfig(
+    level=logging.INFO,
+    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s'
+)
+logger = logging.getLogger(__name__)
+
+# ══════════════════════════════════════════════════════════
+# INITIALIZE MODELS
+# ══════════════════════════════════════════════════════════
+
+logger.info("Initializing VoiceNote AI...")
+try:
+    mistral_client = MistralClient()
+    asr_model = ASRModel()
+    vips_classifier = VIPSClassifier(mistral_client)
+    logger.info("All models loaded successfully")
+except Exception as e:
+    logger.error(f"Initialization error: {e}")
+    raise
+
+# ══════════════════════════════════════════════════════════
+# DATA STORAGE
+# ══════════════════════════════════════════════════════════
+
+historik = []
+alla_resultat = []
+sus_data = {"poang": None, "svar": []}
+nasa_data = {"poang": None, "svar": []}
+
+# ══════════════════════════════════════════════════════════
+# MAIN PROCESSING FUNCTIONS
+# ══════════════════════════════════════════════════════════
+
+def transkribera_ljud(ljud_path: str, referens_text: str):
+    """
+    Transcribe audio and build result with VIPS classification.
+    Returns: (transcription_html, vips_html, history_html, download_text)
+    """
+    try:
+        # Validate audio
+        if not ljud_path:
+            return ("❌ Ingen ljudfil",  "", "", None)
+        
+        # Transcribe
+        start_time = time.time()
+        transkription = asr_model.transcribe(ljud_path)
+        asr_tid = round(time.time() - start_time, 2)
+        
+        # Build result
+        return bygg_resultat(transkription, referens_text, asr_tid)
+        
+    except Exception as e:
+        logger.error(f"Audio transcription error: {e}")
+        return (f"❌ Fel: {str(e)}", "", "", None)
+
+
+def klassificera_text(text_input: str, referens_text: str):
+    """
+    Classify text directly without ASR.
+    Returns: (transcription_html, vips_html, history_html, download_text)
+    """
+    try:
+        if not text_input or not text_input.strip():
+            return ("❌ Ingen text angiven", "", "", None)
+        
+        return bygg_resultat(text_input.strip(), referens_text, asr_tid=None)
+        
+    except Exception as e:
+        logger.error(f"Text classification error: {e}")
+        return (f"❌ Fel: {str(e)}", "", "", None)
+
+
+def bygg_resultat(transkription: str, referens_text: str, asr_tid: float):
+    """
+    Build result with VIPS classification and WER calculation.
+    Returns: (transcription_html, vips_html, history_html, download_text)
+    """
+    try:
+        # Classify with VIPS
+        llm_start = time.time()
+        vips_text = vips_classifier.classify(transkription)
+        llm_tid = round(time.time() - llm_start, 2)
+        
+        # Calculate total time
+        total_tid = round((asr_tid or 0) + llm_tid, 2)
+        
+        # Calculate WER if reference provided
+        wer_poang = WERCalculator.calculate(referens_text, transkription) if referens_text else None
+        
+        # Parse VIPS
+        vips_dict = VIPSClassifier.parse_vips(vips_text)
+        
+        # Store result
+        post = {
+            "tid": datetime.now().strftime("%H:%M:%S"),
+            "transkription": transkription,
+            "vips": vips_text,
+            "asr_tid": asr_tid,
+            "llm_tid": llm_tid,
+            "total_tid": total_tid,
+            "wer_poang": wer_poang,
+            "referens": referens_text,
+            "prompt_technique": Config.PROMPT_TECHNIQUE,
+        }
+        historik.append(post)
+        alla_resultat.append(post)
+        
+        # Build timing badges
+        badges = []
+        if asr_tid is not None:
+            badges.append(f"<span style='background:#EFF6FF;border:1px solid #BFDBFE;color:#0369A1;padding:4px 12px;border-radius:100px;font-size:12px;font-weight:600;'>🎙️ ASR: {asr_tid}s</span>")
+        badges.append(f"<span style='background:#F5F3FF;border:1px solid #DDD6FE;color:#7C3AED;padding:4px 12px;border-radius:100px;font-size:12px;font-weight:600;'>🧠 LLM: {llm_tid}s</span>")
+        badges.append(f"<span style='background:#ECFDF5;border:1px solid #A7F3D0;color:#059669;padding:4px 12px;border-radius:100px;font-size:12px;font-weight:600;'>⚡ Total: {total_tid}s</span>")
+        
+        # Add prompt technique badge
+        technique_name = Config.get_technique_name()
+        technique_emoji = Config.get_technique_emoji()
+        badges.append(f"<span style='background:#F5F3FF;border:1px solid #DDD6FE;color:#7C3AED;padding:4px 12px;border-radius:100px;font-size:12px;font-weight:600;'>{technique_emoji} {technique_name}</span>")
+        badges.append(f"<span style='background:#ECFDF5;border:1px solid #A7F3D0;color:#059669;padding:4px 12px;border-radius:100px;font-size:12px;font-weight:600;'>🔒 GDPR ×2</span>")
+        timing = f"<div style='display:flex;gap:8px;margin-bottom:12px;flex-wrap:wrap;'>{''.join(badges)}</div>"
+        
+        # Build transcription HTML
+        transkription_html = f"""
+        {timing}{wer_badge(wer_poang)}
+        <div style='background:#F0F9FF;border:1.5px solid #BAE6FD;border-radius:12px;
+                    padding:16px;color:#0F172A;font-size:15px;line-height:1.7;'>
+            🎙️ {transkription}
+        </div>"""
+        
+        # Build download text
+        prompt_technique_full = f"{technique_name} ({Config.PROMPT_TECHNIQUE})"
+        nedladdnings_text = f"""VoiceNote AI — VIPS Journalanteckning
+Tid:       {post['tid']}
+ASR Model: OpenAI Whisper ({Config.ASR_MODEL})
+Prompt:    {prompt_technique_full}
+GDPR:      Dubbel anonymisering · Mistral AI EU-servrar
+
+ASR: {asr_tid if asr_tid else 'N/A'}s  |  LLM: {llm_tid}s  |  Total: {total_tid}s  |  WER: {wer_poang if wer_poang is not None else 'N/A'}%
+
+Transkription:
+{transkription}
+
+VIPS-Dokumentation:
+{vips_text}
+
+Referenstext (för WER):
+{referens_text if referens_text else 'Ingen referenstext angiven'}
+"""
+        
+        # Build VIPS HTML
+        vips_html = formatera_vips_html(vips_dict)
+        
+        # Build history HTML
+        historik_html = formatera_historik_html(historik)
+        
+        return (transkription_html, vips_html, historik_html, nedladdnings_text)
+        
+    except Exception as e:
+        logger.error(f"Result building error: {e}")
+        raise
+
+
+# ══════════════════════════════════════════════════════════
+# GRADIO INTERFACE
+# ══════════════════════════════════════════════════════════
+
+# CSS
+css = """
+.panel{background:#FFFFFF !important;border:1.5px solid #E2E8F0 !important;border-radius:12px !important;padding:14px !important;box-shadow:0 1px 3px rgba(0,0,0,0.06) !important;}
+.gradio-container{font-family:'Inter',-apple-system,BlinkMacSystemFont,sans-serif !important;}
+.gradio-container textarea{color:#0F172A !important;background:#F8FAFC !important;}
+.gradio-radio fieldset{border:none !important;}
+.gradio-radio label{display:inline-flex !important;align-items:center !important;justify-content:center !important;min-width:52px !important;padding:9px 20px !important;margin:4px !important;border:2px solid #CBD5E1 !important;border-radius:12px !important;background:#F8FAFC !important;color:#475569 !important;font-weight:700 !important;font-size:16px !important;cursor:pointer !important;transition:all 0.15s ease !important;}
+.gradio-radio label:hover{border-color:#0369A1 !important;background:#EFF6FF !important;color:#0369A1 !important;}
+.gradio-radio input[type="radio"]{display:none !important;}
+.gradio-radio label:has(input[type="radio"]:checked){background:#0369A1 !important;border-color:#0369A1 !important;color:#ffffff !important;box-shadow:0 4px 12px rgba(3,105,161,0.4) !important;}
+input[type="range"]{accent-color:#0369A1 !important;}
+::-webkit-scrollbar{width:5px;}::-webkit-scrollbar-thumb{background:#CBD5E1;border-radius:4px;}
+footer{display:none !important;}
+"""
+
+# SUS Questions
+sus_fragor = [
+    "Jag skulle vilja använda det här systemet regelbundet.",
+    "Jag tyckte att systemet var onödigt komplicerat.",
+    "Jag tyckte att systemet var lätt att använda.",
+    "Jag tror att jag behöver teknisk hjälp för att använda systemet.",
+    "Jag tyckte att funktionerna var väl integrerade.",
+    "Jag tyckte att det fanns för mycket inkonsistens i systemet.",
+    "De flesta skulle lära sig systemet väldigt snabbt.",
+    "Jag tyckte att systemet var väldigt krångligt.",
+    "Jag kände mig trygg när jag använde systemet.",
+    "Jag behövde lära mig mycket innan jag kunde börja.",
+]
+
+# NASA-TLX Dimensions
+nasa_dimensioner = [
+    "Mental belastning (tänka, besluta, minnas)",
+    "Fysisk belastning (fysisk ansträngning)",
+    "Tidskrav (stressad av tempot)",
+    "Prestation (Nöjd=1 → Missnöjd=20)",
+    "Frustration (irriterad, stressad)",
+    "Ansträngning (hur hårt fick du jobba)",
+]
+
+with gr.Blocks(css=css, theme=gr.themes.Base()) as demo:
+    # Build header with technique indicator
+    technique_name = Config.get_technique_name()
+    technique_emoji = Config.get_technique_emoji()
+    technique_color = Config.get_technique_color()
+    
+    gr.HTML(f"""
+    <div style='background:linear-gradient(135deg,#EFF6FF,#F0FDF4);border:1.5px solid #BFDBFE;
+                border-radius:20px;padding:28px;text-align:center;margin-bottom:20px;
+                box-shadow:0 2px 16px rgba(3,105,161,0.08);'>
+        <div style='display:inline-flex;align-items:center;gap:8px;background:#EFF6FF;
+                    border:1.5px solid #93C5FD;border-radius:100px;padding:6px 18px;margin-bottom:16px;
+                    font-size:11px;color:#0369A1;letter-spacing:0.1em;font-weight:700;'>
+            <span style='width:7px;height:7px;border-radius:50%;background:#0369A1;
+                         animation:pulse 2s infinite;display:inline-block;'></span>
+            Healthcare AI · Högskolan Väst · GDPR-compliant
+        </div>
+        <h1 style='font-size:34px;font-weight:900;margin:0 0 10px;color:#0F172A;'>VoiceNote AI 🎙️</h1>
+        <p style='color:#64748B;margin:0 0 14px;font-size:15px;'>
+            OpenAI Whisper · VIPS-dokumentation · WER · SUS · NASA-TLX
+        </p>
+        <div style='display:inline-flex;align-items:center;gap:8px;background:#F5F3FF;
+                    border:1.5px solid {technique_color}55;border-radius:100px;padding:5px 16px;
+                    font-size:12px;color:{technique_color};font-weight:700;'>
+            {technique_emoji} RESEARCH MODE: {technique_name}
+        </div>
+    </div>
+    <style>@keyframes pulse{{0%,100%{{opacity:1}}50%{{opacity:0.3}}}}</style>""")
+    
+    with gr.Tabs():
+        with gr.TabItem("🎙️ VIPS-dokumentation"):
+            with gr.Row():
+                with gr.Column(scale=1):
+                    gr.HTML("""<div style='background:#EFF6FF;border:1.5px solid #BFDBFE;border-radius:12px;padding:14px;margin-bottom:8px;'>
+                        <p style='color:#1E40AF;font-size:13px;margin:0;'>⓪ Spela in patientsamtalet. Referenstexten används för <strong>WER</strong> — lämna tomt om du inte har en.</p></div>""")
+                    ljud_input = gr.Audio(sources=["microphone"], type="filepath", label="Spela in patientsamtal", elem_classes=["panel"])
+                    referens_input = gr.Textbox(label="Referenstext (valfritt, för WER-beräkning)", placeholder="Exakt text som borde ha sagts...", lines=3, elem_classes=["panel"])
+                    
+                    with gr.Row():
+                        analysera_btn = gr.Button("① Analysera 🎙️", variant="primary", size="lg")
+                    
+                    gr.HTML("<div style='border-top:1.5px solid #E2E8F0;margin:18px 0;'></div>")
+                    gr.HTML("<div style='background:#F0FDF4;border:1.5px solid #A7F3D0;border-radius:12px;padding:14px;margin-bottom:8px;'><p style='color:#047857;font-size:13px;margin:0;'>Eller skriv in text direkt (utan ljudinspelning):</p></div>")
+                    text_input = gr.Textbox(label="Text-input (direkt VIPS-klassificering)", placeholder="Jag har ont i huvudet och tar Metoprolol...", lines=5, elem_classes=["panel"])
+                    text_btn = gr.Button("② Klassificera text 📝", size="lg")
+                
+                with gr.Column(scale=1):
+                    transkription_output = gr.HTML(label="Transkription + WER")
+                    vips_output = gr.HTML(label="VIPS")
+                    nedladdning = gr.File(label="💾 Ladda ner VIPS-anteckning (.txt)")
+            
+            gr.HTML("<div style='border-top:2px solid #E2E8F0;margin:24px 0;'></div>")
+            gr.HTML("<h3 style='color:#1E40AF;margin-bottom:12px;'>📊 Senaste analyser</h3>")
+            historik_output = gr.HTML()
+            exportera_btn = gr.Button("📥 Exportera alla analyser (.csv)")
+            csv_output = gr.File(label="CSV-export")
+        
+        with gr.TabItem("📋 SUS"):
+            gr.HTML("<div style='background:#EFF6FF;border:1.5px solid #BFDBFE;border-radius:12px;padding:16px;margin-bottom:16px;'><p style='color:#1E40AF;margin:0;font-size:14px;'><strong>System Usability Scale (SUS)</strong> — Betygsätt systemet från 1 (instämmer inte) till 5 (instämmer helt).</p></div>")
+            sus_inputs = [gr.Radio([1,2,3,4,5], label=q, elem_classes=["panel"]) for q in sus_fragor]
+            sus_btn = gr.Button("Beräkna SUS-poäng", variant="primary", size="lg")
+            sus_resultat = gr.HTML()
+        
+        with gr.TabItem("🧠 NASA-TLX"):
+            gr.HTML("<div style='background:#EFF6FF;border:1.5px solid #BFDBFE;border-radius:12px;padding:16px;margin-bottom:16px;'><p style='color:#1E40AF;margin:0;font-size:14px;'><strong>NASA Task Load Index (TLX)</strong> — Betygsätt varje dimension från 1 (låg) till 20 (hög).</p></div>")
+            nasa_inputs = [gr.Slider(1,20,step=1,label=d,elem_classes=["panel"]) for d in nasa_dimensioner]
+            nasa_btn = gr.Button("Beräkna NASA-TLX", variant="primary", size="lg")
+            nasa_resultat = gr.HTML()
+    
+    # Event handlers
+    analysera_btn.click(
+        fn=transkribera_ljud,
+        inputs=[ljud_input, referens_input],
+        outputs=[transkription_output, vips_output, historik_output, nedladdning]
+    ).then(
+        fn=spara_nedladdning,
+        inputs=[nedladdning],
+        outputs=[nedladdning]
+    )
+    
+    text_btn.click(
+        fn=klassificera_text,
+        inputs=[text_input, referens_input],
+        outputs=[transkription_output, vips_output, historik_output, nedladdning]
+    ).then(
+        fn=spara_nedladdning,
+        inputs=[nedladdning],
+        outputs=[nedladdning]
+    )
+    
+    exportera_btn.click(
+        fn=lambda: exportera_csv(alla_resultat),
+        outputs=[csv_output]
+    )
+    
+    sus_btn.click(
+        fn=berakna_sus,
+        inputs=sus_inputs,
+        outputs=[sus_resultat]
+    )
+    
+    nasa_btn.click(
+        fn=berakna_nasa,
+        inputs=nasa_inputs,
+        outputs=[nasa_resultat]
+    )
+
+if __name__ == "__main__":
+    demo.launch()

config.py

@@ -0,0 +1,108 @@
+"""
+VoiceNote AI - Configuration
+=============================
+All system configuration in one place.
+"""
+
+import os
+import tempfile
+
+
+class Config:
+    """Configuration class for VoiceNote AI"""
+    
+    # ══════════════════════════════════════════════════════════
+    # MODEL SETTINGS
+    # ══════════════════════════════════════════════════════════
+    
+    # ASR Model (Automatic Speech Recognition)
+    # Using OpenAI's official Whisper model (works with Swedish via language parameter)
+    # Options: "openai/whisper-tiny", "openai/whisper-small", "openai/whisper-medium", "openai/whisper-large-v2"
+    ASR_MODEL = "openai/whisper-small"  # Fast and good for Swedish
+    
+    # LLM Model (Large Language Model for VIPS classification)
+    LLM_MODEL = "mistral-large-latest"
+    
+    # ══════════════════════════════════════════════════════════
+    # PROMPT TECHNIQUE (for research comparison)
+    # ══════════════════════════════════════════════════════════
+    
+    # Options: "few_shot", "chain_of_thought"
+    # Change this to compare different prompt techniques
+    PROMPT_TECHNIQUE = "few_shot"
+    
+    # ══════════════════════════════════════════════════════════
+    # PROCESSING LIMITS
+    # ══════════════════════════════════════════════════════════
+    
+    MAX_AUDIO_SECONDS = 30
+    CHUNK_LENGTH_S = 30
+    STRIDE_LENGTH_S = 5
+    
+    # ══════════════════════════════════════════════════════════
+    # LLM SETTINGS
+    # ══════════════════════════════════════════════════════════
+    
+    LLM_TEMPERATURE = 0.1
+    LLM_MAX_TOKENS = 800  # Increased for Chain-of-Thought
+    
+    # ══════════════════════════════════════════════════════════
+    # WER (Word Error Rate) THRESHOLDS
+    # ══════════════════════════════════════════════════════════
+    
+    WER_EXCELLENT = 5.0   # < 5% = Excellent
+    WER_GOOD = 10.0       # 5-10% = Good
+    WER_ACCEPTABLE = 20.0 # 10-20% = Acceptable
+    
+    # ══════════════════════════════════════════════════════════
+    # USABILITY THRESHOLDS
+    # ══════════════════════════════════════════════════════════
+    
+    # SUS (System Usability Scale) - out of 100
+    SUS_EXCELLENT = 85
+    SUS_GOOD = 70
+    SUS_PASS = 50
+    
+    # NASA-TLX (Task Load Index) - out of 120 (lower is better)
+    NASA_LOW = 40
+    NASA_MEDIUM = 60
+    
+    # ══════════════════════════════════════════════════════════
+    # FILE PATHS
+    # ══════════════════════════════════════════════════════════
+    
+    TEMP_DIR = tempfile.gettempdir()
+    
+    # ══════════════════════════════════════════════════════════
+    # DISPLAY NAMES (for UI)
+    # ══════════════════════════════════════════════════════════
+    
+    @classmethod
+    def get_technique_name(cls) -> str:
+        """Get full prompt technique name"""
+        if cls.PROMPT_TECHNIQUE == "few_shot":
+            return "Few-shot Prompting"
+        elif cls.PROMPT_TECHNIQUE == "chain_of_thought":
+            return "Chain-of-Thought Prompting"
+        else:
+            return "Unknown Technique"
+    
+    @classmethod
+    def get_technique_emoji(cls) -> str:
+        """Get emoji for current prompt technique"""
+        if cls.PROMPT_TECHNIQUE == "few_shot":
+            return "📚"
+        elif cls.PROMPT_TECHNIQUE == "chain_of_thought":
+            return "🧠"
+        else:
+            return "❓"
+    
+    @classmethod
+    def get_technique_color(cls) -> str:
+        """Get color for current prompt technique"""
+        if cls.PROMPT_TECHNIQUE == "few_shot":
+            return "#7C3AED"  # Purple
+        elif cls.PROMPT_TECHNIQUE == "chain_of_thought":
+            return "#0369A1"  # Blue
+        else:
+            return "#6B7280"  # Gray

gdpr_filter.py

@@ -0,0 +1,159 @@
+"""
+VoiceNote AI - GDPR Filter
+===========================
+Anonymizes personal information from text.
+Dual-layer protection: input + output anonymization.
+"""
+
+import re
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class GDPRFilter:
+    """GDPR-compliant anonymization with dual-layer protection"""
+    
+    # Replacement tokens
+    REPLACEMENTS = {
+        'personnummer': '[PERSONNR]',
+        'telefon': '[TELEFON]',
+        'epost': '[EPOST]',
+        'datum': '[DATUM]',
+        'adress': '[ADRESS]',
+        'namn': '[NAMN]',
+    }
+    
+    # Words to exclude from name anonymization (Swedish medical/common terms)
+    EXCLUDE_WORDS = {
+        # Common Swedish words
+        'Patienten', 'Patient', 'Jag', 'Mig', 'Min', 'Mitt', 'Mina',
+        'Du', 'Dig', 'Din', 'Ditt', 'Dina',
+        'Han', 'Hon', 'Hen', 'Honom', 'Henne',
+        'Vi', 'Oss', 'Vår', 'Vårt', 'Våra',
+        'De', 'Dem', 'Deras',
+        
+        # Medical terms
+        'Läkaren', 'Läkare', 'Sjuksköterskan', 'Sköterska',
+        'Doktor', 'Doktorn', 'Sjuksköterska',
+        
+        # Time/location words that might be capitalized
+        'Sverige', 'Svensk', 'Svenska',
+        'Måndag', 'Tisdag', 'Onsdag', 'Torsdag', 'Fredag', 'Lördag', 'Söndag',
+        'Januari', 'Februari', 'Mars', 'April', 'Maj', 'Juni',
+        'Juli', 'Augusti', 'September', 'Oktober', 'November', 'December',
+        
+        # Common actions/verbs that might be capitalized
+        'Ta', 'Tar', 'Tog', 'Tagit',
+        'Gå', 'Går', 'Gick', 'Gått',
+        'Känner', 'Kände', 'Känt',
+    }
+    
+    @classmethod
+    def anonymize(cls, text: str) -> str:
+        """
+        Anonymize personal information in text.
+        Returns anonymized text with PII replaced by tokens.
+        """
+        if not text or not text.strip():
+            return text
+        
+        anonymized = text
+        stats = {}
+        
+        # 1. Personnummer (Swedish personal ID)
+        personnr_pattern = r'\b\d{6,8}[-–]\d{4}\b'
+        count = len(re.findall(personnr_pattern, anonymized))
+        if count > 0:
+            anonymized = re.sub(personnr_pattern, cls.REPLACEMENTS['personnummer'], anonymized)
+            stats['personnummer'] = count
+        
+        # 2. Phone numbers
+        telefon_patterns = [
+            r'\b0\d{1,3}[-\s]?\d{5,8}\b',  # Swedish landline/mobile
+            r'\b\+46[-\s]?\d{1,3}[-\s]?\d{5,8}\b',  # International format
+        ]
+        phone_count = 0
+        for pattern in telefon_patterns:
+            matches = re.findall(pattern, anonymized)
+            phone_count += len(matches)
+            anonymized = re.sub(pattern, cls.REPLACEMENTS['telefon'], anonymized)
+        if phone_count > 0:
+            stats['telefon'] = phone_count
+        
+        # 3. Email addresses
+        epost_pattern = r'\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b'
+        count = len(re.findall(epost_pattern, anonymized))
+        if count > 0:
+            anonymized = re.sub(epost_pattern, cls.REPLACEMENTS['epost'], anonymized)
+            stats['epost'] = count
+        
+        # 4. Dates
+        datum_patterns = [
+            r'\b\d{4}-\d{2}-\d{2}\b',  # YYYY-MM-DD
+            r'\b\d{2}/\d{2}/\d{4}\b',  # DD/MM/YYYY
+            r'\b\d{1,2}\s+(januari|februari|mars|april|maj|juni|juli|augusti|september|oktober|november|december)\s+\d{4}\b',
+        ]
+        date_count = 0
+        for pattern in datum_patterns:
+            matches = re.findall(pattern, anonymized, re.IGNORECASE)
+            date_count += len(matches)
+            anonymized = re.sub(pattern, cls.REPLACEMENTS['datum'], anonymized, flags=re.IGNORECASE)
+        if date_count > 0:
+            stats['datum'] = date_count
+        
+        # 5. Addresses (Swedish format)
+        adress_patterns = [
+            r'\b[A-ZÅÄÖ][a-zåäö]+(?:gatan|vägen|stigen|platsen|torget)\s+\d+[A-Za-z]?\b',
+            r'\b\d{3}\s?\d{2}\s+[A-ZÅÄÖ][a-zåäö]+\b',  # Postal code + city
+        ]
+        address_count = 0
+        for pattern in adress_patterns:
+            matches = re.findall(pattern, anonymized)
+            address_count += len(matches)
+            anonymized = re.sub(pattern, cls.REPLACEMENTS['adress'], anonymized)
+        if address_count > 0:
+            stats['adress'] = address_count
+        
+        # 6. Names (capitalized words, with exclusions)
+        # Only anonymize standalone capitalized words that aren't common words
+        words = anonymized.split()
+        name_count = 0
+        for i, word in enumerate(words):
+            # Check if word starts with capital and is >2 chars
+            if len(word) > 2 and word[0].isupper() and word[1:].islower():
+                # Extract just the word (remove punctuation)
+                clean_word = re.sub(r'[^\w]', '', word)
+                name = clean_word
+                
+                # Check if it's in exclusion list
+                if name not in cls.EXCLUDE_WORDS:
+                    anonymized = anonymized.replace(name, cls.REPLACEMENTS['namn'], 1)
+                    name_count += 1
+        
+        if name_count > 0:
+            stats['namn'] = name_count
+        
+        if stats:
+            logger.debug(f"GDPR filter stats: {stats}")
+        
+        return anonymized
+    
+    @classmethod
+    def validate_anonymization(cls, text: str) -> bool:
+        """
+        Check if text still contains personal information.
+        Returns True if text appears anonymized, False otherwise.
+        """
+        # Check for obvious patterns
+        dangerous_patterns = [
+            r'\b\d{6,8}[-–]\d{4}\b',  # Personnummer
+            r'\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b',  # Email
+        ]
+        
+        for pattern in dangerous_patterns:
+            if re.search(pattern, text):
+                logger.warning("Potential PII found after anonymization!")
+                return False
+        
+        return True

models.py

@@ -0,0 +1,134 @@
+"""
+VoiceNote AI - Models
+=====================
+ASR (Whisper) and LLM (Mistral) model interfaces.
+"""
+
+import os
+import torch
+import spaces
+import requests
+import logging
+from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
+from config import Config
+
+logger = logging.getLogger(__name__)
+
+
+# ══════════════════════════════════════════════════════════
+# MISTRAL AI CLIENT (HTTP API)
+# ══════════════════════════════════════════════════════════
+
+class MistralClient:
+    """Mistral AI client using direct HTTP API (more reliable than SDK)"""
+    
+    def __init__(self):
+        self.api_key = os.environ.get("MISTRAL_API_KEY", "")
+        if not self.api_key:
+            logger.warning("MISTRAL_API_KEY not found in environment")
+        
+        self.api_url = "https://api.mistral.ai/v1/chat/completions"
+        logger.info("Mistral client initialized with HTTP API")
+    
+    def chat(self, prompt: str) -> str:
+        """Send chat request to Mistral AI via HTTP"""
+        try:
+            headers = {
+                "Content-Type": "application/json",
+                "Authorization": f"Bearer {self.api_key}"
+            }
+            
+            payload = {
+                "model": Config.LLM_MODEL,
+                "messages": [{"role": "user", "content": prompt}],
+                "temperature": Config.LLM_TEMPERATURE,
+                "max_tokens": Config.LLM_MAX_TOKENS
+            }
+            
+            response = requests.post(
+                self.api_url,
+                headers=headers,
+                json=payload,
+                timeout=30
+            )
+            
+            if response.status_code != 200:
+                error_msg = f"Mistral API error: {response.status_code}"
+                logger.error(f"{error_msg} - {response.text}")
+                raise Exception(error_msg)
+            
+            result = response.json()
+            return result["choices"][0]["message"]["content"]
+            
+        except Exception as e:
+            logger.error(f"Mistral API error: {e}")
+            raise Exception(f"LLM fel: {str(e)}")
+
+
+# ══════════════════════════════════════════════════════════
+# ASR MODEL (WHISPER)
+# ══════════════════════════════════════════════════════════
+
+class ASRModel:
+    """Automatic Speech Recognition using Whisper"""
+    
+    def __init__(self):
+        logger.info(f"Loading ASR model: {Config.ASR_MODEL}")
+        
+        device = "cuda:0" if torch.cuda.is_available() else "cpu"
+        torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
+        
+        try:
+            model = AutoModelForSpeechSeq2Seq.from_pretrained(
+                Config.ASR_MODEL,
+                torch_dtype=torch_dtype,
+                low_cpu_mem_usage=True,
+                use_safetensors=True,
+            )
+            model.to(device)
+            
+            processor = AutoProcessor.from_pretrained(Config.ASR_MODEL)
+            
+            self.pipe = pipeline(
+                "automatic-speech-recognition",
+                model=model,
+                tokenizer=processor.tokenizer,
+                feature_extractor=processor.feature_extractor,
+                chunk_length_s=Config.CHUNK_LENGTH_S,
+                stride_length_s=Config.STRIDE_LENGTH_S,
+                torch_dtype=torch_dtype,
+                device=device,
+            )
+            
+            logger.info(f"ASR model loaded successfully on {device}")
+            
+        except Exception as e:
+            logger.error(f"Failed to load ASR model: {e}")
+            raise
+    
+    @spaces.GPU(duration=60)
+    def transcribe(self, audio_path: str) -> str:
+        """
+        Transcribe audio file to Swedish text.
+        Returns transcribed text.
+        """
+        try:
+            if not audio_path:
+                raise ValueError("No audio file provided")
+            
+            logger.info(f"Transcribing audio: {audio_path}")
+            
+            result = self.pipe(
+                audio_path,
+                generate_kwargs={"language": "sv"},  # Swedish language
+                return_timestamps=False,
+            )
+            
+            text = result["text"].strip()
+            logger.info(f"Transcription successful: {len(text)} characters")
+            
+            return text
+            
+        except Exception as e:
+            logger.error(f"Transcription error: {e}")
+            raise Exception(f"Taligenkänning misslyckades: {str(e)}")

requirements.txt

@@ -0,0 +1,30 @@
+# VoiceNote AI - Requirements
+# Optimized for HuggingFace Spaces
+
+# Core ML/AI frameworks
+torch
+transformers
+accelerate
+safetensors
+
+# HTTP requests (for Mistral API)
+requests
+
+# Speech processing
+jiwer
+librosa
+soundfile
+
+# Web interface
+gradio
+
+# Utilities
+python-dotenv
+
+# HuggingFace Spaces (CRITICAL!)
+spaces
+
+# Optional: Development dependencies (uncomment if needed locally)
+# pytest
+# black
+# flake8

utils.py

@@ -0,0 +1,260 @@
+"""
+VoiceNote AI - Utilities
+=========================
+WER calculator and HTML formatting functions.
+"""
+
+import os
+import csv
+import logging
+from typing import Optional, Dict
+from datetime import datetime
+from jiwer import wer as compute_wer
+from config import Config
+
+logger = logging.getLogger(__name__)
+
+
+# ══════════════════════════════════════════════════════════
+# WER CALCULATOR
+# ══════════════════════════════════════════════════════════
+
+class WERCalculator:
+    """Word Error Rate calculator with validation"""
+    
+    @staticmethod
+    def calculate(reference: str, hypothesis: str) -> Optional[float]:
+        """
+        Calculate WER score.
+        Returns percentage (0-100) or None if invalid input.
+        """
+        if not reference or not reference.strip():
+            return None
+        
+        if not hypothesis or not hypothesis.strip():
+            logger.warning("Empty hypothesis for WER calculation")
+            return 100.0  # All words are errors
+        
+        try:
+            score = compute_wer(
+                reference.lower().strip(),
+                hypothesis.lower().strip()
+            )
+            percentage = round(score * 100, 1)
+            logger.info(f"WER calculated: {percentage}%")
+            return percentage
+        except Exception as e:
+            logger.error(f"WER calculation error: {e}")
+            return None
+    
+    @staticmethod
+    def get_quality_label(wer: Optional[float]) -> str:
+        """Get quality label for WER score"""
+        if wer is None:
+            return "N/A"
+        
+        if wer < Config.WER_EXCELLENT:
+            return "Utmärkt"
+        elif wer < Config.WER_GOOD:
+            return "Bra"
+        elif wer < Config.WER_ACCEPTABLE:
+            return "Godkänd"
+        else:
+            return "Behöver förbättras"
+
+
+# ══════════════════════════════════════════════════════════
+# HTML FORMATTERS
+# ══════════════════════════════════════════════════════════
+
+def formatera_vips_html(vips: dict) -> str:
+    """Format VIPS dictionary as colored HTML"""
+    colors = {
+        'V': ('#10B981', '#ECFDF5'),  # Green
+        'I': ('#F59E0B', '#FFFBEB'),  # Amber
+        'P': ('#3B82F6', '#EFF6FF'),  # Blue
+        'S': ('#EF4444', '#FEF2F2'),  # Red
+    }
+    
+    categories = {
+        'V': 'Välbefinnande',
+        'I': 'Integritet',
+        'P': 'Prevention',
+        'S': 'Säkerhet',
+    }
+    
+    html_parts = []
+    for key in ['V', 'I', 'P', 'S']:
+        fg, bg = colors[key]
+        cat = categories[key]
+        text = vips.get(key, "Ingen relevant information.")
+        
+        html_parts.append(f"""
+        <div style='background:{bg};border-left:4px solid {fg};padding:12px 16px;margin-bottom:10px;border-radius:8px;'>
+            <div style='color:{fg};font-weight:700;font-size:13px;margin-bottom:4px;'>{key} — {cat}</div>
+            <div style='color:#1F2937;font-size:14px;line-height:1.6;'>{text}</div>
+        </div>
+        """)
+    
+    return "".join(html_parts)
+
+
+def wer_badge(wer_poang: Optional[float]) -> str:
+    """Create WER badge HTML"""
+    if wer_poang is None:
+        return ""
+    
+    if wer_poang < Config.WER_EXCELLENT:
+        color, bg = "#059669", "#ECFDF5"
+        label = "Utmärkt ✅"
+    elif wer_poang < Config.WER_GOOD:
+        color, bg = "#0369A1", "#EFF6FF"
+        label = "Bra ✅"
+    elif wer_poang < Config.WER_ACCEPTABLE:
+        color, bg = "#D97706", "#FFFBEB"
+        label = "Godkänd ⚠️"
+    else:
+        color, bg = "#DC2626", "#FEF2F2"
+        label = "Behöver förbättras ❌"
+    
+    return f"""
+    <div style='background:{bg};border:2px solid {color}55;border-radius:14px;
+                padding:20px;margin-bottom:12px;text-align:center;'>
+        <div style='font-size:42px;font-weight:900;color:{color};'>{wer_poang:.1f}%</div>
+        <div style='color:{color};font-size:16px;font-weight:700;margin-top:4px;'>WER: {label}</div>
+    </div>"""
+
+
+def formatera_historik_html(historik: list) -> str:
+    """Format history list as HTML"""
+    if not historik:
+        return "<p style='color:#64748B;text-align:center;padding:40px;'>Ingen historik ännu.</p>"
+    
+    items = []
+    for post in reversed(historik[-5:]):  # Last 5 entries
+        tid = post.get('tid', 'N/A')
+        wer = post.get('wer_poang')
+        wer_text = f"{wer:.1f}%" if wer is not None else "N/A"
+        
+        technique = post.get('prompt_technique', 'unknown')
+        technique_emoji = "📚" if technique == "few_shot" else "🧠"
+        
+        items.append(f"""
+        <div style='background:#F8FAFC;border:1.5px solid #E2E8F0;border-radius:10px;
+                    padding:12px;margin-bottom:8px;'>
+            <div style='display:flex;justify-content:space-between;align-items:center;'>
+                <span style='color:#475569;font-weight:600;'>{technique_emoji} {tid}</span>
+                <span style='color:#64748B;font-size:13px;'>WER: {wer_text}</span>
+            </div>
+        </div>
+        """)
+    
+    return "".join(items)
+
+
+# ══════════════════════════════════════════════════════════
+# FILE EXPORT FUNCTIONS
+# ══════════════════════════════════════════════════════════
+
+def spara_nedladdning(innehall: str) -> Optional[str]:
+    """Save text file for download"""
+    try:
+        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+        path = os.path.join(Config.TEMP_DIR, f"vips_anteckning_{timestamp}.txt")
+        with open(path, "w", encoding="utf-8") as f:
+            f.write(innehall)
+        logger.info(f"Saved text file: {path}")
+        return path
+    except Exception as e:
+        logger.error(f"Error saving text file: {e}")
+        return None
+
+
+def exportera_csv(alla_resultat: list) -> Optional[str]:
+    """Export all results as CSV"""
+    if not alla_resultat:
+        return None
+    try:
+        path = os.path.join(Config.TEMP_DIR, "voicenote_resultat.csv")
+        with open(path, "w", newline="", encoding="utf-8") as f:
+            fieldnames = ["tid", "prompt_technique", "asr_tid", "llm_tid", "total_tid", "wer_poang", "transkription", "referens", "vips"]
+            writer = csv.DictWriter(f, fieldnames=fieldnames, extrasaction="ignore")
+            writer.writeheader()
+            writer.writerows(alla_resultat)
+        logger.info(f"Exported CSV: {path}")
+        return path
+    except Exception as e:
+        logger.error(f"CSV export error: {e}")
+        return None
+
+
+# ══════════════════════════════════════════════════════════
+# USABILITY CALCULATORS
+# ══════════════════════════════════════════════════════════
+
+def berakna_sus(*svar) -> str:
+    """Calculate SUS score and return HTML"""
+    try:
+        pl = []
+        for i, val in enumerate(svar):
+            try:
+                v = int(val)
+            except (TypeError, ValueError):
+                return "<div style='color:#DC2626;padding:14px;background:#FEF2F2;border-radius:10px;font-weight:600;'>⚠️ Fyll i alla 10 frågor.</div>"
+            pl.append(v - 1 if i % 2 == 0 else 5 - v)
+        
+        total = sum(pl) * 2.5
+        
+        # Determine color and label
+        if total >= Config.SUS_GOOD:
+            f, b = "#059669", "#ECFDF5"
+        elif total >= Config.SUS_PASS:
+            f, b = "#D97706", "#FFFBEB"
+        else:
+            f, b = "#DC2626", "#FEF2F2"
+        
+        e = ("Utmärkt ✅" if total >= Config.SUS_EXCELLENT else 
+             "Bra ✅" if total >= Config.SUS_GOOD else 
+             "Godkänd ⚠️" if total >= Config.SUS_PASS else 
+             "Underkänd ❌")
+        
+        return f"""<div style='background:{b};border:2px solid {f}55;border-radius:14px;padding:28px;text-align:center;'>
+            <div style='font-size:60px;font-weight:900;color:{f};'>{total:.1f}</div>
+            <div style='color:{f};font-size:18px;font-weight:700;margin-top:6px;'>{e}</div>
+        </div>"""
+    except Exception as ex:
+        logger.error(f"SUS calculation error: {ex}")
+        return "<div style='color:#DC2626;'>Fel vid beräkning</div>"
+
+
+def berakna_nasa(*svar) -> str:
+    """Calculate NASA-TLX score and return HTML"""
+    try:
+        values = []
+        for val in svar:
+            try:
+                v = int(val)
+                values.append(v)
+            except (TypeError, ValueError):
+                return "<div style='color:#DC2626;padding:14px;background:#FEF2F2;border-radius:10px;font-weight:600;'>⚠️ Fyll i alla 6 dimensioner.</div>"
+        
+        total = sum(values)
+        
+        if total <= Config.NASA_LOW:
+            f, b = "#059669", "#ECFDF5"
+            e = "Låg belastning ✅"
+        elif total <= Config.NASA_MEDIUM:
+            f, b = "#D97706", "#FFFBEB"
+            e = "Medel belastning ⚠️"
+        else:
+            f, b = "#DC2626", "#FEF2F2"
+            e = "Hög belastning ❌"
+        
+        return f"""<div style='background:{b};border:2px solid {f}55;border-radius:14px;padding:28px;text-align:center;'>
+            <div style='font-size:60px;font-weight:900;color:{f};'>{total}</div>
+            <div style='color:{f};font-size:14px;margin-top:4px;'>/ 120 poäng</div>
+            <div style='color:{f};font-size:18px;font-weight:700;margin-top:10px;'>{e}</div>
+        </div>"""
+    except Exception as ex:
+        logger.error(f"NASA-TLX calculation error: {ex}")
+        return "<div style='color:#DC2626;'>Fel vid beräkning</div>"

vips_classifier.py

@@ -0,0 +1,181 @@
+"""
+VoiceNote AI - VIPS Classifier
+===============================
+VIPS classification with dual GDPR protection.
+Supports both Few-shot and Chain-of-Thought prompting.
+"""
+
+import logging
+from typing import Dict
+from config import Config
+from gdpr_filter import GDPRFilter
+from models import MistralClient
+
+logger = logging.getLogger(__name__)
+
+
+class VIPSClassifier:
+    """VIPS classification with support for different prompt techniques"""
+    
+    def __init__(self, llm_client: MistralClient):
+        self.llm = llm_client
+    
+    @staticmethod
+    def build_prompt_few_shot(text: str) -> str:
+        """
+        Few-shot prompting: Ger 2-3 konkreta exempel före den verkliga uppgiften.
+        
+        Fördelar:
+        - Tydliga mönster för modellen att följa
+        - Bättre formatering
+        - Konsekvent output
+        
+        Forskningsreferens: Brown et al. (2020) - Language Models are Few-Shot Learners
+        """
+        return f"""Du är ett dokumentationssystem för sjuksköterskor.
+Din uppgift är att klassificera patientinformation enligt VIPS-modellen.
+
+OBLIGATORISKA REGLER:
+1. Skriv ALDRIG namn, personnummer, telefonnummer eller adresser.
+2. Hitta INTE på information. Dokumentera ENBART vad patienten faktiskt säger.
+3. Lägg INTE till medicinska råd, diagnoser eller rekommendationer.
+4. Om en kategori saknar information, skriv exakt: Ingen relevant information.
+5. Svara i exakt VIPS-format — fyra rader, en per kategori, inget annat.
+
+VIPS-KATEGORIER:
+V (Välbefinnande) — Symtom, smärta, mående och känslor
+I (Integritet)    — Vanor, önskemål och preferenser
+P (Prevention)    — Förebyggande åtgärder som nämnts
+S (Säkerhet)      — Risker, mediciner och säkerhetsaspekter
+
+EXEMPEL 1:
+Samtal: "Jag har haft huvudvärk i två dagar. Jag tar Metoprolol dagligen."
+Svar:
+V: Patienten rapporterar huvudvärk sedan två dagar.
+I: Ingen relevant information.
+P: Ingen relevant information.
+S: Patienten tar Metoprolol dagligen.
+
+EXEMPEL 2:
+Samtal: "Jag sover dåligt på nätterna och känner mig orolig. Jag brukar gå på promenader varje dag."
+Svar:
+V: Patienten rapporterar sömnsvårigheter och känslor av oro.
+I: Patienten promenerar dagligen.
+P: Ingen relevant information.
+S: Ingen relevant information.
+
+EXEMPEL 3:
+Samtal: "Jag har ont i bröstet när jag går i trappor. Jag röker 10 cigaretter om dagen."
+Svar:
+V: Patienten rapporterar bröstsmärta vid ansträngning.
+I: Ingen relevant information.
+P: Ingen relevant information.
+S: Patienten röker 10 cigaretter dagligen. Risk för hjärt-kärlsjukdom.
+
+NU ÄR DET DIN TUR. Klassificera följande samtal:
+SAMTAL:
+"{text}"
+
+Klassificera i exakt VIPS-format (V:, I:, P:, S:):"""
+    
+    @staticmethod
+    def build_prompt_chain_of_thought(text: str) -> str:
+        """
+        Chain-of-Thought prompting: Be modellen att tänka steg-för-steg.
+        
+        Fördelar:
+        - Bättre resonemang
+        - Minskad hallucination
+        - Tydligare logik
+        
+        Forskningsreferens: Wei et al. (2022) - Chain-of-Thought Prompting 
+        Elicits Reasoning in Large Language Models
+        """
+        return f"""Du är ett dokumentationssystem för sjuksköterskor.
+Din uppgift är att klassificera patientinformation enligt VIPS-modellen.
+
+OBLIGATORISKA REGLER:
+1. Skriv ALDRIG namn, personnummer, telefonnummer eller adresser.
+2. Hitta INTE på information. Dokumentera ENBART vad patienten faktiskt säger.
+3. Lägg INTE till medicinska råd, diagnoser eller rekommendationer.
+4. Om en kategori saknar information, skriv exakt: Ingen relevant information.
+
+VIPS-KATEGORIER:
+V (Välbefinnande) — Symtom, smärta, mående och känslor
+I (Integritet)    — Vanor, önskemål och preferenser
+P (Prevention)    — Förebyggande åtgärder som nämnts
+S (Säkerhet)      — Risker, mediciner och säkerhetsaspekter
+
+SAMTAL:
+"{text}"
+
+STEG-FÖR-STEG ANALYS:
+Tänk igenom detta systematiskt:
+
+1. Läs igenom samtalet noggrant.
+2. Identifiera ALL information som nämnts (symtom, känslor, vanor, mediciner, risker).
+3. Sortera varje informationsdel:
+   - Handlar det om hur patienten MÅR? → V (Välbefinnande)
+   - Handlar det om patientens VANOR/PREFERENSER? → I (Integritet)
+   - Handlar det om FÖREBYGGANDE åtgärder? → P (Prevention)
+   - Handlar det om RISKER/MEDICINER/SÄKERHET? → S (Säkerhet)
+4. För varje kategori: Formulera en kort, professionell mening.
+5. Om kategorin är tom: Skriv "Ingen relevant information."
+
+Genomför analysen steg-för-steg, sedan ge ditt svar i exakt VIPS-format:
+V: [din analys]
+I: [din analys]
+P: [din analys]
+S: [din analys]"""
+    
+    def build_prompt(self, text: str) -> str:
+        """
+        Build prompt based on configured technique.
+        Used for research comparison between Few-shot and Chain-of-Thought.
+        """
+        if Config.PROMPT_TECHNIQUE == "chain_of_thought":
+            logger.info("Using Chain-of-Thought prompting")
+            return self.build_prompt_chain_of_thought(text)
+        else:  # Default to few_shot
+            logger.info("Using Few-shot prompting")
+            return self.build_prompt_few_shot(text)
+    
+    def classify(self, text: str) -> str:
+        """
+        Classify text according to VIPS model with dual GDPR protection.
+        Returns VIPS-formatted text.
+        """
+        try:
+            # Layer 1: Anonymize input
+            anonymized_input = GDPRFilter.anonymize(text)
+            logger.info("Input anonymized (Layer 1)")
+            
+            # Build prompt using selected technique
+            prompt = self.build_prompt(anonymized_input)
+            response = self.llm.chat(prompt)
+            
+            # Layer 2: Anonymize output
+            anonymized_output = GDPRFilter.anonymize(response)
+            logger.info("Output anonymized (Layer 2)")
+            
+            # Validate anonymization
+            if not GDPRFilter.validate_anonymization(anonymized_output):
+                logger.error("GDPR validation failed!")
+            
+            return anonymized_output
+        except Exception as e:
+            logger.error(f"VIPS classification error: {e}")
+            raise
+    
+    @staticmethod
+    def parse_vips(vips_text: str) -> Dict[str, str]:
+        """Parse VIPS text into dictionary"""
+        vips = {k: "Ingen relevant information." for k in ["V", "I", "P", "S"]}
+        
+        for line in vips_text.strip().split("\n"):
+            line = line.strip()
+            for key in vips:
+                if line.startswith(f"{key}:"):
+                    vips[key] = line[2:].strip()
+        
+        return vips