Create app.py

app.py

@@ -0,0 +1,730 @@
+import os
+import json
+import shutil
+import gradio as gr
+import tempfile
+from datetime import datetime
+from typing import List, Dict, Any, Optional
+from pytube import YouTube
+from pathlib import Path
+import re
+from moviepy.editor import VideoFileClip
+
+# --- Agent Imports & Safe Fallbacks ---
+try:
+    from alz_companion.agent import (
+        bootstrap_vectorstore, make_rag_chain, answer_query, synthesize_tts,
+        transcribe_audio, detect_tags_from_query, describe_image, build_or_load_vectorstore,
+        _default_embeddings
+    )
+    from alz_companion.prompts import BEHAVIOUR_TAGS, EMOTION_STYLES
+    from langchain.schema import Document
+    from langchain_community.vectorstores import FAISS
+    AGENT_OK = True
+except Exception as e:
+    AGENT_OK = False
+    # Define all fallback functions and classes
+    def bootstrap_vectorstore(sample_paths=None, index_path="data/"): return object()
+    def build_or_load_vectorstore(docs, index_path, is_personal=False): return object()
+    def make_rag_chain(vs_general, vs_personal, **kwargs): return lambda q, **k: {"answer": f"(Demo) You asked: {q}", "sources": []}
+    def answer_query(chain, q, **kwargs): return chain(q, **kwargs)
+    def synthesize_tts(text: str, lang: str = "en"): return None
+    def transcribe_audio(filepath: str, lang: str = "en"): return "This is a transcribed message."
+    def detect_tags_from_query(query: str, behavior_options: list, emotion_options: list, topic_options: list, context_options: list, **kwargs): return {"detected_behaviors": [], "detected_emotion": "None", "detected_topic": "None", "detected_contexts": []}
+    def describe_image(image_path: str): return "This is a description of an image."
+    def _default_embeddings(): return None
+    class Document:
+        def __init__(self, page_content, metadata):
+            self.page_content = page_content
+            self.metadata = metadata
+    class FAISS:
+        def __init__(self):
+            self.docstore = type('obj', (object,), {'_dict': {}})()
+    BEHAVIOUR_TAGS = {"None": []}
+    EMOTION_STYLES = {"None": {}}
+    print(f"WARNING: Could not import from alz_companion ({e}). Running in UI-only demo mode.")
+
+# --- Centralized Configuration ---
+CONFIG = {
+    "themes": ["All", "The Father", "Still Alice", "Away from Her", "Alive Inside", "General Caregiving"],
+    "roles": ["patient", "caregiver"],
+    "behavior_tags": ["None"] + list(BEHAVIOUR_TAGS.keys()),
+    "emotion_tags": ["None"] + list(EMOTION_STYLES.keys()),
+    "topic_tags": [
+        "None", "caregiving_advice", "medical_fact", "personal_story", "research_update", 
+        "treatment_option:home_safety", "treatment_option:long_term_care", "treatment_option:music_therapy", 
+        "treatment_option:reassurance", "treatment_option:routine_structuring", "treatment_option:validation_therapy"
+    ],
+    "context_tags": [
+        "None", "disease_stage_mild", "disease_stage_moderate", "disease_stage_advanced", 
+        "disease_stage_unspecified", "interaction_mode_one_to_one", "interaction_mode_small_group", 
+        "interaction_mode_group_activity", "relationship_family", "relationship_spouse", 
+        "relationship_staff_or_caregiver", "relationship_unspecified", "setting_home_or_community", 
+        "setting_care_home", "setting_clinic_or_hospital"
+    ],
+    "languages": {"English": "en", "Chinese": "zh", "Malay": "ms", "French": "fr", "Spanish": "es"},
+    "tones": ["warm", "neutral", "formal", "playful"]
+}
+
+# --- File Management & Vector Store Logic ---
+INDEX_BASE = os.getenv('INDEX_BASE', 'data')
+UPLOADS_BASE = os.path.join(INDEX_BASE, "uploads")
+PERSONAL_INDEX_PATH = os.path.join(INDEX_BASE, "personal_faiss_index")
+os.makedirs(UPLOADS_BASE, exist_ok=True)
+os.makedirs(os.path.dirname(PERSONAL_INDEX_PATH), exist_ok=True)
+THEME_PATHS = {t: os.path.join(INDEX_BASE, f"faiss_index_{t.replace(' ', '').lower()}") for t in CONFIG["themes"]}
+
+# Global variables
+vectorstores = {}
+personal_vectorstore = None
+test_fixtures = []
+example_retriever = None
+
+def canonical_theme(tk: str) -> str: return tk if tk in CONFIG["themes"] else "All"
+def theme_upload_dir(theme: str) -> str:
+    p = os.path.join(UPLOADS_BASE, f"theme_{canonical_theme(theme).replace(' ', '').lower()}")
+    os.makedirs(p, exist_ok=True)
+    return p
+def load_manifest(theme: str) -> Dict[str, Any]:
+    p = os.path.join(theme_upload_dir(theme), "manifest.json")
+    if os.path.exists(p):
+        try:
+            with open(p, "r", encoding="utf-8") as f: return json.load(f)
+        except Exception: pass
+    return {"files": {}}
+def save_manifest(theme: str, man: Dict[str, Any]):
+    with open(os.path.join(theme_upload_dir(theme), "manifest.json"), "w", encoding="utf-8") as f: json.dump(man, f, indent=2)
+def list_theme_files(theme: str) -> List[tuple[str, bool]]:
+    man = load_manifest(theme)
+    base = theme_upload_dir(theme)
+    found = [(n, bool(e)) for n, e in man.get("files", {}).items() if os.path.exists(os.path.join(base, n))]
+    existing = {n for n, e in found}
+    for name in sorted(os.listdir(base)):
+        if name not in existing and os.path.isfile(os.path.join(base, name)): found.append((name, False))
+    man["files"] = dict(found)
+    save_manifest(theme, man)
+    return found
+def copy_into_theme(theme: str, src_path: str) -> str:
+    fname = os.path.basename(src_path)
+    dest = os.path.join(theme_upload_dir(theme), fname)
+    shutil.copy2(src_path, dest)
+    return dest
+def seed_files_into_theme(theme: str):
+    SEED_FILES = [
+        ("sample_data/caregiving_tips.txt", True),
+        ("sample_data/the_father_segments_enriched_harmonized_plus.jsonl", True),
+        ("sample_data/still_alice_enriched_harmonized_plus.jsonl", True),
+        ("sample_data/away_from_her_enriched_harmonized_plus.jsonl", True),
+        ("sample_data/alive_inside_enriched_harmonized.jsonl", True)
+    ]
+    man, changed = load_manifest(theme), False
+    for path, enable in SEED_FILES:
+        if not os.path.exists(path): continue
+        fname = os.path.basename(path)
+        if not os.path.exists(os.path.join(theme_upload_dir(theme), fname)):
+            copy_into_theme(theme, path)
+            man["files"][fname] = bool(enable)
+            changed = True
+    if changed: save_manifest(theme, man)
+
+def ensure_index(theme='All'):
+    theme = canonical_theme(theme)
+    if theme in vectorstores: return vectorstores[theme]
+    upload_dir = theme_upload_dir(theme)
+    enabled_files = [os.path.join(upload_dir, n) for n, enabled in list_theme_files(theme) if enabled]
+    index_path = THEME_PATHS.get(theme)
+    vectorstores[theme] = bootstrap_vectorstore(sample_paths=enabled_files, index_path=index_path)
+    return vectorstores[theme]
+
+# --- Gradio Callbacks ---
+def collect_settings(*args):
+    keys = ["role", "patient_name", "caregiver_name", "tone", "language", "tts_lang", "temperature", "behaviour_tag", "emotion_tag", "topic_tag", "active_theme", "tts_on", "debug_mode"]
+    return dict(zip(keys, args))
+
+def parse_and_tag_entries(text_content: str, source: str) -> List[Document]:
+    separator_pattern = r'\n(?:---|--|-|-\*-|-\.-)\n'
+    entries = re.split(separator_pattern, text_content)
+    docs_to_add = []
+    
+    for entry in entries:
+        if not entry.strip(): continue
+        title = "Untitled Text Entry"
+        content = entry.strip()
+
+        lines = entry.strip().split('\n')
+        if lines and "title:" in lines[0].lower():
+            title_line = lines[0].split(':', 1)
+            title = title_line[1].strip() if len(title_line) > 1 else "Untitled"
+            content_part = "\n".join(lines[1:])
+            if "content:" in content_part.lower():
+                content = content_part.split(':', 1)[1].strip()
+            else:
+                content = content_part.strip()
+
+        full_content = f"Title: {title}\n\nContent: {content}"
+        print(f"  - Parsing entry: '{title}'")
+        behavior_options = CONFIG.get("behavior_tags", [])
+        emotion_options = CONFIG.get("emotion_tags", [])
+        topic_options = CONFIG.get("topic_tags", [])
+        context_options = CONFIG.get("context_tags", [])
+
+        detected_tags = detect_tags_from_query(
+            content, 
+            behavior_options=behavior_options, emotion_options=emotion_options,
+            topic_options=topic_options, context_options=context_options,
+            example_retriever=example_retriever
+        )
+        
+        metadata = {"source": source, "title": title}
+        detected_behaviors = detected_tags.get("detected_behaviors", [])
+        if detected_behaviors:
+             metadata["behaviors"] = [b.lower() for b in detected_behaviors]
+        if detected_tags.get("detected_emotion") != "None":
+            metadata["emotion"] = detected_tags.get("detected_emotion").lower()
+        detected_topics = detected_tags.get("detected_topic")
+        if detected_topics and detected_topics != "None":
+            metadata["topic_tags"] = [detected_topics.lower()]
+        detected_contexts = detected_tags.get("detected_contexts", [])
+        if detected_contexts:
+            metadata["context_tags"] = [c.lower() for c in detected_contexts]
+
+        docs_to_add.append(Document(page_content=full_content, metadata=metadata))
+    return docs_to_add
+
+def handle_add_knowledge(title, text_input, file_input, image_input, yt_url):
+    global personal_vectorstore
+    docs_to_add = []
+    if text_input and text_input.strip():
+        docs_to_add = parse_and_tag_entries(f"Title: {title}\n\nContent: {text_input}", "Text Input")
+    elif file_input:
+        content_source = os.path.basename(file_input)
+        path_for_transcription = file_input
+        temp_audio_file_path = None
+        if file_input.lower().endswith('.mov'):
+            print(f"Detected .mov file. Converting {file_input} to audio...")
+            try:
+                fd, temp_mp3_path = tempfile.mkstemp(suffix=".mp3")
+                os.close(fd)
+                with VideoFileClip(file_input) as video_clip:
+                    audio_clip = video_clip.audio
+                    audio_clip.write_audiofile(temp_mp3_path, logger=None)
+                path_for_transcription = temp_mp3_path
+                temp_audio_file_path = temp_mp3_path
+                print("Conversion successful.")
+            except Exception as e:
+                return f"Error converting .mov file: {e}"
+        
+        if file_input.lower().endswith('.txt'):
+            with open(file_input, 'r', encoding='utf-8') as f:
+                file_content = f.read()
+            docs_to_add = parse_and_tag_entries(file_content, content_source)
+        else:
+            final_title = title.strip() if title and title.strip() else "Audio/Video Note"
+            content_text = transcribe_audio(path_for_transcription)
+            full_content = f"Title: {final_title}\n\nContent: {content_text}"
+            docs_to_add = parse_and_tag_entries(full_content, content_source)
+        
+        if temp_audio_file_path:
+            os.remove(temp_audio_file_path)
+    elif image_input:
+        final_title = title.strip() if title and title.strip() else "Image Note"
+        content_text = describe_image(image_input)
+        full_content = f"Title: {final_title}\n\nContent: {content_text}"
+        docs_to_add = parse_and_tag_entries(full_content, "Image Input")
+    elif yt_url and ("youtube.com" in yt_url or "youtu.be" in yt_url):
+        try:
+            yt = YouTube(yt_url)
+            video_title = yt.title
+            final_title = title.strip() if title and title.strip() else video_title
+            audio_stream = yt.streams.get_audio_only()
+            with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as temp_audio_file:
+                audio_stream.download(filename=temp_audio_file.name)
+                temp_audio_path = temp_audio_file.name
+            content_text = transcribe_audio(temp_audio_path)
+            content_source = f"YouTube: {video_title}"
+            os.remove(temp_audio_path)
+            full_content = f"Title: {final_title}\n\nContent: {content_text}"
+            docs_to_add = parse_and_tag_entries(full_content, content_source)
+        except Exception as e:
+            return f"Error processing YouTube link: {e}"
+    else:
+        return "Please provide a title and content, or another input source."
+
+    if not docs_to_add:
+        return "No processable content found to add."
+    
+    if personal_vectorstore is None:
+        personal_vectorstore = build_or_load_vectorstore(docs_to_add, PERSONAL_INDEX_PATH, is_personal=True)
+    else:
+        personal_vectorstore.add_documents(docs_to_add)
+    
+    personal_vectorstore.save_local(PERSONAL_INDEX_PATH)
+    return f"Successfully added {len(docs_to_add)} new memory/memories."
+
+def save_chat_to_memory(chat_history):
+    global personal_vectorstore
+    if not chat_history: return "Nothing to save."
+    formatted_chat = []
+    for message in chat_history:
+        role = "User" if message["role"] == "user" else "Assistant"
+        content = message["content"].strip()
+        if content.startswith("*(Auto-detected context:") or content.startswith("*(Sources used:"): continue
+        formatted_chat.append(f"{role}: {content}")
+    conversation_text = "\n".join(formatted_chat)
+    if not conversation_text: return "No conversation content to save."
+    timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+    title = f"Conversation from {timestamp}"
+    full_content = f"Title: {title}\n\nContent:\n{conversation_text}"
+    doc_to_add = Document(page_content=full_content, metadata={"source": "Saved Chat", "title": title})
+    if personal_vectorstore is None:
+        personal_vectorstore = build_or_load_vectorstore([doc_to_add], PERSONAL_INDEX_PATH, is_personal=True)
+    else:
+        personal_vectorstore.add_documents([doc_to_add])
+    personal_vectorstore.save_local(PERSONAL_INDEX_PATH)
+    print(f"Saved conversation to long-term memory.")
+    return f"Conversation from {timestamp} saved successfully!"
+
+def list_personal_memories():
+    global personal_vectorstore
+    if personal_vectorstore is None or not hasattr(personal_vectorstore.docstore, '_dict') or not personal_vectorstore.docstore._dict:
+        return gr.update(value=[["No memories to display", "", ""]]), gr.update(choices=["No memories to select"], value=None)
+    docs = list(personal_vectorstore.docstore._dict.values())
+    dataframe_data = [[doc.metadata.get('title', 'Untitled'), doc.metadata.get('source', 'Unknown'), doc.page_content] for doc in docs]
+    dropdown_choices = [doc.page_content for doc in docs]
+    return gr.update(value=dataframe_data), gr.update(choices=dropdown_choices)
+
+def delete_personal_memory(memory_to_delete):
+    global personal_vectorstore
+    if personal_vectorstore is None or not memory_to_delete:
+        return "Knowledge base is empty or no memory selected."
+    all_docs = list(personal_vectorstore.docstore._dict.values())
+    docs_to_keep = [doc for doc in all_docs if doc.page_content != memory_to_delete]
+    if len(all_docs) == len(docs_to_keep):
+        return "Error: Could not find the selected memory to delete."
+    print(f"Deleting memory. {len(docs_to_keep)} memories remaining.")
+    if not docs_to_keep:
+        if os.path.isdir(PERSONAL_INDEX_PATH):
+            shutil.rmtree(PERSONAL_INDEX_PATH)
+        personal_vectorstore = build_or_load_vectorstore([], PERSONAL_INDEX_PATH, is_personal=True)
+    else:
+        new_vs = FAISS.from_documents(docs_to_keep, _default_embeddings())
+        new_vs.save_local(PERSONAL_INDEX_PATH)
+        personal_vectorstore = new_vs
+    return "Successfully deleted memory. The list will now refresh."
+
+def chat_fn(user_text, audio_file, settings, chat_history):
+    global personal_vectorstore, example_retriever
+    question = (user_text or "").strip()
+    if audio_file and not question:
+        try:
+            voice_lang_name = settings.get("tts_lang", "English")
+            voice_lang_code = CONFIG["languages"].get(voice_lang_name, "en")
+            question = transcribe_audio(audio_file, lang=voice_lang_code)
+        except Exception as e:
+            err_msg = f"Audio Error: {e}" if settings.get("debug_mode") else "Sorry, I couldn't understand the audio."
+            chat_history.append({"role": "assistant", "content": err_msg})
+            return "", None, chat_history
+    if not question:
+        return "", None, chat_history
+            
+    chat_history.append({"role": "user", "content": question})
+    
+    manual_behavior_tag = settings.get("behaviour_tag", "None")
+    manual_emotion_tag = settings.get("emotion_tag", "None")
+    manual_topic_tag = settings.get("topic_tag", "None")
+
+    scenario_tag = None
+    emotion_tag = None
+    topic_tag = None
+    
+    if manual_behavior_tag != "None" or manual_emotion_tag != "None" or manual_topic_tag != "None":
+        print("Manual tags detected, skipping auto-detection.")
+        scenario_tag = [manual_behavior_tag] if manual_behavior_tag != "None" else []
+        emotion_tag = manual_emotion_tag
+        topic_tag = manual_topic_tag
+    else:
+        print("No manual tags set, running auto-detection...")
+        behavior_options = CONFIG.get("behavior_tags", [])
+        emotion_options = CONFIG.get("emotion_tags", [])
+        topic_options = CONFIG.get("topic_tags", [])
+        context_options = CONFIG.get("context_tags", [])
+        
+        detected_tags = detect_tags_from_query(
+            question, 
+            behavior_options=behavior_options, emotion_options=emotion_options, 
+            topic_options=topic_options, context_options=context_options,
+            example_retriever=example_retriever
+        )
+        
+        scenario_tag = detected_tags.get("detected_behaviors", [])
+        emotion_tag = detected_tags.get("detected_emotion", "None")
+        topic_tag = detected_tags.get("detected_topic", "None")
+        
+        detected_parts = []
+        if scenario_tag: detected_parts.append(f"Behaviors=`{', '.join(scenario_tag)}`")
+        if emotion_tag and emotion_tag != "None": detected_parts.append(f"Emotion=`{emotion_tag}`")
+        if topic_tag and topic_tag != "None": detected_parts.append(f"Topic=`{topic_tag}`")
+        if detected_parts and settings.get("debug_mode"):
+            detected_msg = f"*(Auto-detected context: {', '.join(detected_parts)})*"
+            chat_history.append({"role": "assistant", "content": detected_msg})
+
+    active_theme = settings.get("active_theme", "All")
+    vs_general = ensure_index(active_theme)
+    if personal_vectorstore is None:
+        personal_vectorstore = build_or_load_vectorstore([], PERSONAL_INDEX_PATH, is_personal=True)
+    
+    rag_chain_settings = {
+        "role": settings.get("role"), "temperature": settings.get("temperature"), "language": settings.get("language"),
+        "patient_name": settings.get("patient_name"), "caregiver_name": settings.get("caregiver_name"), "tone": settings.get("tone"),
+    }
+    
+    chain = make_rag_chain(vs_general, personal_vectorstore, **rag_chain_settings)
+    
+    final_scenario_tag = scenario_tag if scenario_tag else None
+    final_emotion_tag = emotion_tag if emotion_tag != "None" else None
+    final_topic_tag = topic_tag if topic_tag != "None" else None
+
+    # Determine history length to exclude debug messages
+    history_offset = 0
+    if chat_history:
+        for msg in reversed(chat_history):
+            if msg["role"] == "user": break
+            history_offset += 1
+    simple_history = chat_history[:-history_offset] if history_offset > 0 else []
+
+    response = answer_query(chain, question, chat_history=simple_history, scenario_tag=final_scenario_tag, emotion_tag=final_emotion_tag, topic_tag=final_topic_tag)
+    
+    answer = response.get("answer", "[No answer found]")
+    chat_history.append({"role": "assistant", "content": answer})
+    
+    if settings.get("debug_mode") and response.get("sources"):
+        sources = response.get("sources", [])
+        valid_sources = [s for s in sources if s and s not in ["unknown", "placeholder", "General Knowledge"]]
+        if valid_sources:
+            source_msg = f"*(Sources used: {', '.join(valid_sources)})*"
+            chat_history.append({"role": "assistant", "content": source_msg})
+    
+    audio_out = None
+    if settings.get("tts_on") and answer:
+        tts_lang_code = CONFIG["languages"].get(settings.get("tts_lang"), "en")
+        audio_out = synthesize_tts(answer, lang=tts_lang_code)
+        
+    from gradio import update
+    return "", (update(value=audio_out, visible=bool(audio_out))), chat_history
+
+def upload_knowledge(files, current_theme):
+    if not files: return "No files were selected to upload."
+    added = 0
+    for f in files:
+        try:
+            copy_into_theme(current_theme, f.name); added += 1
+        except Exception as e: print(f"Error uploading file {f.name}: {e}")
+    if added > 0 and current_theme in vectorstores: del vectorstores[current_theme]
+    return f"Uploaded {added} file(s). Refreshing file list..."
+def save_file_selection(current_theme, enabled_files):
+    man = load_manifest(current_theme)
+    for fname in man['files']: man['files'][fname] = fname in enabled_files
+    save_manifest(current_theme, man)
+    if current_theme in vectorstores: del vectorstores[current_theme]
+    return f"Settings saved. Index for theme '{current_theme}' will rebuild on the next query."
+def refresh_file_list_ui(current_theme):
+    files = list_theme_files(current_theme)
+    enabled = [f for f, en in files if en]
+    msg = f"Found {len(files)} file(s). {len(enabled)} enabled."
+    return gr.update(choices=[f for f, _ in files], value=enabled), msg
+def auto_setup_on_load(current_theme):
+    theme_dir = theme_upload_dir(current_theme)
+    if not os.listdir(theme_dir):
+        print("First-time setup: Auto-seeding sample data...")
+        seed_files_into_theme(current_theme)
+    all_settings = collect_settings("caregiver", "", "", "warm", "English", "English", 0.7, "None", "None", "None", "All", True, False)
+    files_ui, status_msg = refresh_file_list_ui(current_theme)
+    return all_settings, files_ui, status_msg
+
+def setup_example_retriever():
+    global example_retriever
+    print("Setting up example retriever...")
+    fixtures_path = "conversation_test_fixtures.jsonl"
+    if not os.path.exists(fixtures_path):
+        print(f"WARNING: Test fixtures file not found at {fixtures_path}")
+        return
+    example_docs = []
+    with open(fixtures_path, "r", encoding="utf-8") as f:
+        for line in f:
+            data = json.loads(line)
+            doc = Document(page_content=data["turns"][0]["text"], metadata={"full_fixture": data})
+            example_docs.append(doc)
+    if example_docs:
+        example_vs = FAISS.from_documents(example_docs, _default_embeddings())
+        example_retriever = example_vs.as_retriever(search_kwargs={"k": 3})
+        print(f"Example retriever created with {len(example_docs)} examples.")
+
+def load_test_fixtures():
+    global test_fixtures
+    test_fixtures = []
+    try:
+        script_dir = os.path.dirname(os.path.abspath(__file__))
+        fixtures_path = os.path.join(script_dir, "conversation_test_fixtures.jsonl")
+        if not os.path.exists(fixtures_path):
+            print("WARNING: Test fixtures file not found.")
+            return gr.update(choices=[])
+        with open(fixtures_path, "r", encoding="utf-8") as f:
+            for line in f:
+                test_fixtures.append(json.loads(line))
+        test_titles = [fixture["title"] for fixture in test_fixtures]
+        return gr.update(choices=test_titles)
+    except Exception as e:
+        print(f"UNEXPECTED ERROR during file loading: {e}")
+        return gr.update(choices=[])
+
+def run_nlu_test(test_title: str):
+    print("\n--- RUNNING NLU TEST (Flexible Grading) ---")
+    if not test_title or not test_fixtures:
+        return "Please select a test case.", None
+    selected_fixture = next((f for f in test_fixtures if f["title"] == test_title), None)
+    if not selected_fixture:
+        return f"Error: Could not find test case titled '{test_title}'.", None
+    user_query = selected_fixture["turns"][0]["text"]
+    expected_results = selected_fixture["expected"]
+    print(f"Test Case: '{test_title}'")
+    print(f"User Query: '{user_query}'")
+    behavior_options = CONFIG.get("behavior_tags", [])
+    emotion_options = CONFIG.get("emotion_tags", [])
+    topic_options = CONFIG.get("topic_tags", [])
+    context_options = CONFIG.get("context_tags", [])
+    actual_results_raw = detect_tags_from_query(
+        user_query,
+        behavior_options=behavior_options, emotion_options=emotion_options,
+        topic_options=topic_options, context_options=context_options,
+        example_retriever=example_retriever
+    )
+    print(f"\nRAW NLU RESULTS from detect_tags_from_query:\n{actual_results_raw}\n")
+    actual_results = {
+        "emotion": [actual_results_raw.get("detected_emotion")],
+        "behaviors": actual_results_raw.get("detected_behaviors", []),
+        "topic_tags": [actual_results_raw.get("detected_topic")],
+        "context_tags": actual_results_raw.get("detected_contexts", [])
+    }
+    pass_count, total_count = 0, 0
+    comparison_data = []
+    all_keys = set(expected_results.keys()) | set(actual_results.keys())
+    print("--- COMPARING RESULTS ---")
+    for key in sorted(list(all_keys)):
+        expected_set = set(expected_results.get(key, []))
+        actual_set = set(a for a in actual_results.get(key, []) if a and a != "None")
+        if not expected_set: continue
+        total_count += 1
+        is_pass = len(expected_set.intersection(actual_set)) > 0
+        print(f"Category: '{key}'", f"  - Expected Set: {expected_set}", f"  - Actual Set  : {actual_set}", f"  - Logic       : len(intersection) > 0", f"  - Result      : {is_pass}", sep="\n")
+        if is_pass: pass_count += 1
+        comparison_data.append([
+            test_title, key, ", ".join(sorted(list(expected_set))),
+            ", ".join(sorted(list(actual_set))) if actual_set else "None",
+            "✅ Pass" if is_pass else "❌ Fail"
+        ])
+    status = f"## Test Result: {pass_count} / {total_count} Categories Passed"
+    print(f"Final Status: {pass_count}/{total_count} passed.")
+    print("--- TEST COMPLETE ---\n")
+    return status, comparison_data
+
+def run_all_nlu_tests():
+    if not test_fixtures:
+        return "Test fixtures are not loaded. Please check your JSONL file.", None
+    print("\n--- RUNNING ALL NLU TESTS ---")
+    all_results_data = []
+    total_pass_count, total_categories_count = 0, 0
+    for fixture in test_fixtures:
+        user_query = fixture["turns"][0]["text"]
+        expected_results = fixture["expected"]
+        test_title = fixture["title"]
+        print(f"\n-- Testing Case: {test_title} --")
+        behavior_options = CONFIG.get("behavior_tags", [])
+        emotion_options = CONFIG.get("emotion_tags", [])
+        topic_options = CONFIG.get("topic_tags", [])
+        context_options = CONFIG.get("context_tags", [])
+        actual_results_raw = detect_tags_from_query(
+            user_query,
+            behavior_options=behavior_options, emotion_options=emotion_options,
+            topic_options=topic_options, context_options=context_options,
+            example_retriever=example_retriever
+        )
+        actual_results = {
+            "emotion": [actual_results_raw.get("detected_emotion")],
+            "behaviors": actual_results_raw.get("detected_behaviors", []),
+            "topic_tags": [actual_results_raw.get("detected_topic")],
+            "context_tags": actual_results_raw.get("detected_contexts", [])
+        }
+        all_keys = set(expected_results.keys())
+        for key in sorted(list(all_keys)):
+            expected_set = set(expected_results.get(key, []))
+            actual_set = set(a for a in actual_results.get(key, []) if a and a != "None")
+            if not expected_set: continue
+            total_categories_count += 1
+            is_pass = len(expected_set.intersection(actual_set)) > 0
+            if is_pass: total_pass_count += 1
+            all_results_data.append([
+                test_title, key, ", ".join(sorted(list(expected_set))),
+                ", ".join(sorted(list(actual_set))) if actual_set else "None",
+                "✅ Pass" if is_pass else "❌ Fail"
+            ])
+    pass_rate = (total_pass_count / total_categories_count * 100) if total_categories_count > 0 else 0
+    summary_status = f"## Batch Test Complete\n- **Overall Score:** {total_pass_count} / {total_categories_count} Categories Passed\n- **Pass Rate:** {pass_rate:.2f}%"
+    print(summary_status)
+    print("--- BATCH TEST COMPLETE ---")
+    return summary_status, all_results_data
+
+def run_all_startup_tasks(current_theme):
+    """A single wrapper function to handle all application startup tasks."""
+    print("--- Running all startup tasks ---")
+    settings, files_ui, status_msg = auto_setup_on_load(current_theme)
+    setup_example_retriever()
+    test_case_choices = load_test_fixtures()
+    return settings, files_ui, status_msg, test_case_choices
+
+
+# --- UI Definition ---
+CSS = ".gradio-container { font-size: 14px; } #chatbot { min-height: 250px; } #audio_out audio { max-height: 40px; } #audio_in audio { max-height: 40px; padding: 0; }"
+with gr.Blocks(theme=gr.themes.Soft(), css=CSS) as demo:
+    settings_state = gr.State({})
+    with gr.Tab("Chat"):
+        user_text = gr.Textbox(show_label=False, placeholder="Type your message here...")
+        audio_in = gr.Audio(sources=["microphone"], type="filepath", label="Voice Input", elem_id="audio_in")
+        with gr.Row():
+            submit_btn = gr.Button("Send", variant="primary")
+            save_btn = gr.Button("Save to Memory")
+            clear_btn = gr.Button("Clear")
+        chat_status = gr.Markdown()
+        audio_out = gr.Audio(label="Response Audio", autoplay=True, visible=True, elem_id="audio_out")
+        chatbot = gr.Chatbot(elem_id="chatbot", label="Conversation", type="messages")
+    with gr.Tab("Personalize"):
+        with gr.Accordion("Add to Personal Knowledge Base", open=True):
+            gr.Markdown("Add personal notes, memories, or descriptions. A descriptive title helps the AI find memories more accurately.")
+            personal_title = gr.Textbox(label="Title / Entry Name", placeholder="e.g., 'Dad's favorite songs'")
+            personal_text = gr.Textbox(lines=5, label="Text Content (or use file upload)", placeholder="Type or paste text here. Use '—' on a new line to separate multiple entries.")
+            personal_file = gr.File(label="Upload Audio/Video/Text File")
+            personal_image = gr.Image(type="filepath", label="Upload Image")
+            personal_yt_url = gr.Textbox(label="Or, provide a YouTube URL", placeholder="Paste a YouTube link here...")
+            with gr.Row():
+                personal_add_btn = gr.Button("Add Knowledge to Memory", variant="primary")
+            personal_status = gr.Markdown()
+        with gr.Accordion("Manage Personal Knowledge", open=False):
+            personal_memory_display = gr.DataFrame(headers=["Title", "Source", "Content"], label="Saved Personal Memories", interactive=False, row_count=(5, "dynamic"))
+            with gr.Row():
+                personal_refresh_btn = gr.Button("Refresh Memories")
+            with gr.Row():
+                personal_delete_selector = gr.Dropdown(label="Select a memory to delete (by its full content)", scale=3, interactive=True)
+                personal_delete_btn = gr.Button("Delete Selected Memory", variant="stop", scale=1)
+            personal_delete_status = gr.Markdown()
+    with gr.Tab("Testing"):
+        gr.Markdown("## NLU Context Detection Tests")
+        gr.Markdown("Select a single test case to run, or run all tests at once for a full report.")
+        with gr.Row():
+            test_case_dropdown = gr.Dropdown(label="Select Single Test Case", scale=2)
+            run_single_test_btn = gr.Button("Run Single Test", variant="secondary", scale=1)
+            run_all_tests_btn = gr.Button("Run All Tests", variant="primary", scale=1)
+        test_status_md = gr.Markdown("### Please select and run a test.")
+        test_results_df = gr.DataFrame(
+            label="Test Results Comparison",
+            headers=["Test Case", "Category", "Expected", "Actual", "Result"],
+            interactive=False, row_count=(20, "dynamic")
+        )
+    with gr.Tab("Settings"):
+        with gr.Group():
+            gr.Markdown("## Conversation & Persona Settings")
+            with gr.Row():
+                role = gr.Radio(CONFIG["roles"], value="caregiver", label="Your Role")
+                temperature = gr.Slider(0.0, 1.2, value=0.7, step=0.1, label="Creativity")
+                tone = gr.Dropdown(CONFIG["tones"], value="warm", label="Response Tone")
+            with gr.Row():
+                patient_name = gr.Textbox(label="Patient's Name", placeholder="e.g., 'Dad' or 'John'")
+                caregiver_name = gr.Textbox(label="Caregiver's Name", placeholder="e.g., 'me' or 'Jane'")
+                behaviour_tag = gr.Dropdown(CONFIG["behavior_tags"], value="None", label="Behaviour Filter (Manual Override)")
+                emotion_tag = gr.Dropdown(CONFIG["emotion_tags"], value="None", label="Emotion Filter (Manual Override)")
+                topic_tag = gr.Dropdown(CONFIG["topic_tags"], value="None", label="Topic Tag Filter (Manual Override)")
+        with gr.Accordion("Language, Voice & Debugging", open=False):
+            language = gr.Dropdown(list(CONFIG["languages"].keys()), value="English", label="Response Language")
+            tts_lang = gr.Dropdown(list(CONFIG["languages"].keys()), value="English", label="Voice Language")
+            tts_on = gr.Checkbox(True, label="Enable Voice Response (TTS)")
+            debug_mode = gr.Checkbox(False, label="Show Debug Info")
+        gr.Markdown("--- \n ## General Knowledge Base Management")
+        active_theme = gr.Radio(CONFIG["themes"], value="All", label="Active Knowledge Theme")
+        with gr.Row():
+            with gr.Column(scale=1):
+                files_in = gr.File(file_count="multiple", file_types=[".jsonl", ".txt"], label="Upload Knowledge Files")
+                upload_btn = gr.Button("Upload to Theme", variant="secondary")
+                seed_btn = gr.Button("Import Sample Data", variant="secondary")
+            with gr.Column(scale=2):
+                mgmt_status = gr.Markdown()
+                files_box = gr.CheckboxGroup(choices=[], label="Enable Files for the Selected Theme")
+                with gr.Row():
+                    save_files_btn = gr.Button("Save Selection", variant="primary")
+                    refresh_btn = gr.Button("Refresh List")
+        
+    # --- Event Wiring ---
+    all_settings_components = [role, patient_name, caregiver_name, tone, language, tts_lang, temperature, behaviour_tag, emotion_tag, topic_tag, active_theme, tts_on, debug_mode]
+    for component in all_settings_components:
+        component.change(fn=collect_settings, inputs=all_settings_components, outputs=settings_state)
+    
+    submit_btn.click(fn=chat_fn, inputs=[user_text, audio_in, settings_state, chatbot], outputs=[user_text, audio_out, chatbot])
+    save_btn.click(fn=save_chat_to_memory, inputs=[chatbot], outputs=[chat_status])
+    clear_btn.click(lambda: (None, None, [], None, "", ""), outputs=[user_text, audio_out, chatbot, audio_in, user_text, chat_status])
+    
+    personal_add_btn.click(
+        fn=handle_add_knowledge,
+        inputs=[personal_title, personal_text, personal_file, personal_image, personal_yt_url], 
+        outputs=[personal_status]
+    ).then(
+        lambda: (None, None, None, None, None), 
+        outputs=[personal_title, personal_text, personal_file, personal_image, personal_yt_url]
+    )
+    
+    personal_refresh_btn.click(fn=list_personal_memories, inputs=None, outputs=[personal_memory_display, personal_delete_selector])
+    personal_delete_btn.click(fn=delete_personal_memory, inputs=[personal_delete_selector], outputs=[personal_delete_status]).then(fn=list_personal_memories, inputs=None, outputs=[personal_memory_display, personal_delete_selector])
+    
+    upload_btn.click(upload_knowledge, inputs=[files_in, active_theme], outputs=[mgmt_status]).then(refresh_file_list_ui, inputs=[active_theme], outputs=[files_box, mgmt_status])
+    save_files_btn.click(save_file_selection, inputs=[active_theme, files_box], outputs=[mgmt_status])
+    seed_btn.click(seed_files_into_theme, inputs=[active_theme]).then(refresh_file_list_ui, inputs=[active_theme], outputs=[files_box, mgmt_status])
+    refresh_btn.click(refresh_file_list_ui, inputs=[active_theme], outputs=[files_box, mgmt_status])
+    active_theme.change(refresh_file_list_ui, inputs=[active_theme], outputs=[files_box, mgmt_status])
+    
+    # Wiring for the Testing Tab
+    run_single_test_btn.click(
+        fn=run_nlu_test,
+        inputs=[test_case_dropdown],
+        outputs=[test_status_md, test_results_df]
+    )
+    run_all_tests_btn.click(
+        fn=run_all_nlu_tests,
+        inputs=None,
+        outputs=[test_status_md, test_results_df]
+    )
+    
+    # Single, unified startup event
+    demo.load(
+        fn=run_all_startup_tasks,
+        inputs=[active_theme],
+        outputs=[settings_state, files_box, mgmt_status, test_case_dropdown]
+    )
+
+# --- Startup Logic ---
+def pre_load_indexes():
+    global personal_vectorstore
+    print("Pre-loading all knowledge base indexes at startup...")
+    for theme in CONFIG["themes"]:
+        print(f"  - Loading general index for theme: '{theme}'")
+        try:
+            ensure_index(theme)
+            print(f"    ...'{theme}' theme loaded successfully.")
+        except Exception as e:
+            print(f"    ...Error loading theme '{theme}': {e}")
+    print("  - Loading personal knowledge index...")
+    try:
+        personal_vectorstore = build_or_load_vectorstore([], PERSONAL_INDEX_PATH, is_personal=True)
+        print("    ...Personal knowledge loaded successfully.")
+    except Exception as e:
+        print(f"    ...Error loading personal knowledge: {e}")
+    print("All indexes loaded. Application is ready.")
+
+if __name__ == "__main__":
+    pre_load_indexes()
+    demo.queue().launch(debug=True)