Convert Q&A interface to conversational chatbot with memory

app.py: replace static Q&A (Textbox in/out) with gr.ChatInterface +
gr.Chatbot(layout="bubble") embedded in gr.Blocks alongside notes panel.
ChatInterface handles history, Enter-to-submit, streaming, and clear
button automatically. Social sharing buttons retained below chat.

agent/a11y_agent.py: ask() now accepts optional history list[dict].
Passes last 6 history turns to LLM messages. Enriches RAG query with
last 2 user turns for context-aware retrieval on follow-up questions.

agent/prompts.py: rewrite PL and EN system prompts for conversational
style — agent now explicitly references prior conversation context,
adapts response format to question type (structured for requirements,
natural for follow-ups), and knows to say so when context is missing.

Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>

agent/a11y_agent.py

@@ -2,7 +2,7 @@
 
 import json
 from datetime import datetime
-from typing import Optional, Generator
+from typing import Optional, Generator, List, Dict
 from openai import OpenAI
 from langdetect import detect, LangDetectException
 from config import get_settings
@@ -63,35 +63,49 @@ class A11yExpertAgent:
         except Exception as e:
             logger.warning(f"Error closing A11yExpertAgent: {e}")
     
-    def ask(self, question: str) -> Generator[str, None, None]:
+    def ask(
+        self,
+        question: str,
+        history: Optional[List[Dict[str, str]]] = None,
+    ) -> Generator[str, None, None]:
         """
         Ask a question and get a streaming answer with RAG.
-        
+
         Args:
-            question: Question about accessibility
-            
+            question: Current user message.
+            history: Conversation history as list of {"role": ..., "content": ...} dicts.
+
         Yields:
-            Answer chunks from the agent
+            Answer chunks from the agent.
         """
         logger.info(f"Question: {question}")
-        
+        history = history or []
+
         try:
             detected_lang = detect(question)
             language = "pl" if detected_lang.startswith("pl") else "en"
         except LangDetectException:
             language = self.language
-        
+
         logger.info(f"Detected language: {language}")
-        
+
         current_system_prompt = self._prompts.get(language, self._prompts["en"])
 
+        # Enrich RAG query with recent user turns for context-aware retrieval
+        recent_user_turns = [
+            m["content"] for m in history[-4:] if m.get("role") == "user"
+        ]
+        rag_query = " ".join(recent_user_turns[-2:] + [question]) if recent_user_turns else question
+
         logger.info("Searching knowledge base...")
-        context, sources = search_knowledge_base(question, self.vector_store, language=language)
+        context, sources = search_knowledge_base(rag_query, self.vector_store, language=language)
 
+        # Build messages: system + last 6 history turns + current user turn with context
+        history_messages = history[-6:]
         messages = [
             {"role": "system", "content": current_system_prompt},
-            # Stateless: no conversation history in context
-            {"role": "user", "content": self._build_prompt_with_context(question, context, language)}
+            *history_messages,
+            {"role": "user", "content": self._build_prompt_with_context(question, context, language)},
         ]
 
         full_answer = ""

agent/prompts.py

@@ -1,24 +1,16 @@
 """System prompts for A11y Expert agent in different languages."""
 
-SYSTEM_PROMPT_PL = """Jesteś ekspertem dostępności cyfrowej specjalizującym się w WCAG 2.2, WAI-ARIA i prawodawstwie EU.
+SYSTEM_PROMPT_PL = """Jesteś Jacek AI — konwersacyjny ekspert dostępności cyfrowej (WCAG 2.2, WAI-ARIA, prawo EU).
+Prowadzisz rozmowę: pamiętasz co zostało powiedziane wcześniej i nawiązujesz do poprzednich odpowiedzi.
 
-KRYTYCZNE: Odpowiadaj ZAWSZE I WYŁĄCZNIE PO POLSKU. Nawet jeśli otrzymujesz angielskie źródła, tłumacz je i odpowiadaj po polsku.
+KRYTYCZNE: Odpowiadaj ZAWSZE I WYŁĄCZNIE PO POLSKU. Nawet jeśli otrzymujesz angielskie źródła, tłumacz je.
 
-FORMAT ODPOWIEDZI (stosuj gdy dotyczy konkretnych wymagań):
-1. **Kryterium sukcesu**: numer i nazwa (np. 1.4.3 Kontrast (Minimalny))
-2. **Poziom zgodności**: A / AA / AAA
-3. **Wymaganie**: co dokładnie trzeba spełnić
-4. **Jak to osiągnąć**: praktyczne techniki i przykłady
-5. **Źródła**: które fragmenty z bazy wiedzy wykorzystałeś
-
-ZASADY:
-- Odpowiadaj wyłącznie po polsku
-- Cytuj konkretne kryteria sukcesu z numerem (np. SC 1.1.1)
-- Jeśli kontekst z bazy wiedzy nie zawiera odpowiedzi — powiedz to wprost, nie zgaduj
-- Używaj prostego, zrozumiałego języka
-- Dawaj praktyczne przykłady kodu gdy pytanie dotyczy implementacji
-
-Odpowiadasz bezpośrednio na pytanie w oparciu o dostarczony kontekst."""
+STYL ODPOWIEDZI:
+- Na pytania o konkretne wymagania: podaj numer kryterium (np. SC 1.4.3), poziom (A/AA/AAA), co trzeba spełnić, jak to osiągnąć
+- Na pytania ogólne lub follow-up: odpowiadaj naturalnie, nawiązując do kontekstu rozmowy
+- Cytuj konkretne kryteria sukcesu z numerem gdy to istotne
+- Jeśli baza wiedzy nie zawiera odpowiedzi — powiedz wprost, nie zgaduj
+- Dawaj przykłady kodu gdy pytanie dotyczy implementacji"""
 
 SYSTEM_PROMPT_EN = """
 You are an accessibility expert specializing in:
@@ -29,26 +21,22 @@ You are an accessibility expert specializing in:
 
 ⚠️ CRITICAL: You respond ONLY in ENGLISH. All your responses MUST be in English, even if sources are in other languages.
 
-RESPONSE FORMAT (use when answering specific requirements):
-1. **Success Criterion**: number and name (e.g., 1.4.3 Contrast (Minimum))
-2. **Conformance Level**: A / AA / AAA
-3. **Requirement**: what exactly must be met
-4. **How to achieve it**: practical techniques and code examples
-5. **Sources**: which knowledge base fragments you used
+You are Jacek AI — a conversational accessibility expert (WCAG 2.2, WAI-ARIA, EU law).
+You maintain conversation context: remember what was discussed and build on previous answers.
+
+⚠️ CRITICAL: You respond ONLY in ENGLISH. Translate any non-English sources.
 
-MANDATORY RULES:
-1. ✅ ALWAYS respond in ENGLISH - translate non-English sources, never respond in other languages
-2. ✅ ALWAYS cite specific WCAG success criteria with number (e.g., SC 1.4.3)
-3. ✅ If the knowledge base context doesn't contain the answer — say so explicitly, do not guess
-4. ✅ Use simple, understandable language - avoid jargon unless necessary
-5. ✅ Provide practical code examples or implementation guidance when relevant
-6. ✅ Reference both WCAG requirements and industry best practices
+RESPONSE STYLE:
+- For specific requirements: cite the criterion number (e.g., SC 1.4.3), conformance level (A/AA/AAA), what must be met, how to achieve it
+- For general or follow-up questions: respond naturally, referencing prior conversation context
+- Cite specific WCAG success criteria with numbers when relevant
+- If the knowledge base lacks the answer — say so explicitly, do not guess
+- Provide code examples when the question concerns implementation
 
 HOW YOU WORK:
 - You receive context from the WCAG/ARIA knowledge base (relevance rated ★★★/★★☆/★☆☆)
-- Prefer ★★★ sources; use ★☆☆ sources only when nothing better is available
-- You respond DIRECTLY to the question (don't talk about searching the database!)
-- Focus on the content, not the process
+- Prefer ★★★ sources; treat ★☆☆ sources as supplementary only
+- Respond DIRECTLY — don't mention searching the database
 """
 
 SYSTEM_PROMPT_WCAG_EXPERT = """

app.py

@@ -1,340 +1,193 @@
 """
-Gradio UI for Jacek AI with lazy initialization.
-This module creates a Gradio interface that starts FAST,
-then initializes the agent in the background.
+Gradio conversational UI for Jacek AI with lazy initialization.
+Uses gr.ChatInterface for conversation with history and streaming.
 """
 import sys
 import os
+import time
+import threading
+import warnings
 
-# Suppress asyncio cleanup warnings by setting environment variable
 os.environ['PYTHONUNBUFFERED'] = '1'
-
-# Suppress all asyncio warnings at the earliest possible point
-import warnings
 warnings.filterwarnings('ignore', category=ResourceWarning)
 
 import gradio as gr
 from loguru import logger
-import atexit
-import threading
 from agent.a11y_agent import create_agent, A11yExpertAgent
 from config import get_settings
 
 # --- Setup ---
-# Configure logger
 logger.remove()
 logger.add(sys.stderr, level=get_settings().log_level)
 
-# Global agent instance
 agent_instance: A11yExpertAgent = None
 agent_ready = False
 agent_error = None
 
+
 # --- Agent Initialization ---
 def initialize_agent_background():
-    """Initialize the agent in background thread."""
     global agent_instance, agent_ready, agent_error
-    
     try:
         logger.info("🔄 Starting agent initialization in background...")
-        import time
-        
-        logger.info("⏱️ Sleeping 2 seconds to avoid race condition...")
         time.sleep(2)
-        
-        logger.info("📦 Calling create_agent()...")
         agent_instance = create_agent()
-        
-        logger.info("✓ Agent instance created, setting ready flag...")
         agent_ready = True
         logger.success("✅ A11y Expert Agent is ready!")
     except Exception as e:
-        logger.error(f"❌ Failed to initialize agent: {e}")
         import traceback
-        logger.error(traceback.format_exc())
+        logger.error(f"❌ Failed to initialize agent: {e}\n{traceback.format_exc()}")
         agent_error = str(e)
         agent_instance = None
 
-def cleanup_resources():
-    """Clean up resources on app shutdown."""
-    global agent_instance
-    logger.info("Cleaning up resources...")
-    try:
-        # Close agent and all its resources
-        if agent_instance:
-            agent_instance.close()
-        
-        # Close embeddings client singleton if it exists
-        from models.embeddings import get_embeddings_client
-        if hasattr(get_embeddings_client, '_instance'):
-            get_embeddings_client._instance.close()
-        
-        logger.success("✅ Resources cleaned up successfully")
-    except Exception as e:
-        logger.warning(f"Error during cleanup: {e}")
 
-# --- Gradio Chat Logic ---
-def respond(message: str):
-    """
-    Main function for the Gradio Q&A interface.
-    Receives a user question and uses the agent to generate a streaming response.
-
-    Args:
-        message: The user's question.
-
-    Yields:
-        A stream of response chunks to update the UI.
-    """
+def _ensure_agent():
     global agent_instance, agent_ready, agent_error
-    
-    # Wait for background initialization if not ready yet
-    if not agent_ready and not agent_error and agent_instance is None:
-        yield "⏳ Agent is initializing in background, please wait a moment..."
-        import time
-        # Wait up to 15 seconds for background initialization
-        for i in range(30):
-            time.sleep(0.5)
-            if agent_ready:
-                break
-
-        # If still not ready, initialize synchronously as fallback
-        if not agent_ready and agent_instance is None:
-            yield "⏳ Finalizing agent initialization..."
-            try:
-                logger.info("🔄 Background init not complete, initializing synchronously...")
-                agent_instance = create_agent()
-                agent_ready = True
-                logger.success("✅ Jacek AI Agent is ready!")
-            except Exception as e:
-                logger.error(f"❌ Failed to initialize agent: {e}")
-                import traceback
-                logger.error(traceback.format_exc())
-                agent_error = str(e)
-                agent_instance = None
-                yield f"❌ Agent initialization failed: {agent_error}"
-                return
-    
-    # Check if agent failed to initialize
+    if agent_ready:
+        return None
     if agent_error:
-        yield f"❌ Agent initialization failed: {agent_error}"
+        return agent_error
+    for _ in range(30):
+        time.sleep(0.5)
+        if agent_ready:
+            return None
+        if agent_error:
+            return agent_error
+    if not agent_ready:
+        try:
+            agent_instance = create_agent()
+            agent_ready = True
+        except Exception as e:
+            agent_error = str(e)
+            return str(e)
+    return None
+
+
+# --- Chat function ---
+def chat(message: str, history: list):
+    """
+    Streaming chat function for gr.ChatInterface.
+    history: list of {"role": "user"/"assistant", "content": str}
+    Yields partial assistant response strings.
+    """
+    if not message or not message.strip():
+        yield ""
         return
-    
-    if not agent_instance:
-        yield "❌ Agent not available. Please check logs for errors."
+
+    err = _ensure_agent()
+    if err:
+        yield f"❌ Błąd inicjalizacji agenta: {err}"
         return
 
-    logger.info(f"User query: '{message}'")
+    logger.info(f"User: {message!r}")
     full_response = ""
     try:
-        for chunk in agent_instance.ask(message):
+        for chunk in agent_instance.ask(message, history=history):
             full_response += chunk
             yield full_response
     except Exception as e:
-        logger.error(f"Error during response generation: {e}")
-        yield f"An error occurred: {e}"
-
+        logger.error(f"Error during response: {e}")
+        yield f"❌ Błąd: {e}"
 
-# --- Gradio UI Definition ---
-# Q&A Form layout: Question form on left, Notes on right
-# Get public URL from config for social sharing
-SHARE_URL = get_settings().public_url
 
-# Custom HTML head for Open Graph meta tags (for LinkedIn/Facebook previews)
-custom_head = """
-<meta property="og:title" content="Jacek AI - Ekspert Dostępności Cyfrowej" />
-<meta property="og:description" content="Uzyskaj odpowiedzi na pytania o WCAG, ARIA i najlepsze praktyki dostępności cyfrowej. Chatbot oparty na sztucznej inteligencji." />
-<meta property="og:type" content="website" />
-<meta property="og:url" content="https://huggingface.co/spaces/jaczad/JacekAI" />
-<meta name="twitter:card" content="summary_large_image" />
-<meta name="twitter:title" content="Jacek AI - Ekspert Dostępności" />
-<meta name="twitter:description" content="Chatbot AI odpowiadający na pytania o dostępność cyfrową (WCAG, ARIA)" />
-"""
+def load_notes():
+    try:
+        with open("notes.md", "r", encoding="utf-8") as f:
+            return f.read()
+    except FileNotFoundError:
+        return (
+            "## Witaj w Jacek AI! 👋\n\n"
+            "Stwórz `notes.md` aby wyświetlić notatki.\n\n"
+            "- [WCAG 2.2](https://www.w3.org/WAI/WCAG22/quickref/)\n"
+            "- [WAI-ARIA](https://www.w3.org/WAI/ARIA/apg/)\n"
+            "- [MDN Accessibility](https://developer.mozilla.org/en-US/docs/Web/Accessibility)"
+        )
+    except Exception as e:
+        return f"⚠️ Błąd: {e}"
 
-with gr.Blocks(title="Jacek AI", head=custom_head) as demo:
-    gr.Markdown("# 🎯 Jacek AI")
-    gr.Markdown("**Ekspert dostępności cyfrowej** - Uzyskaj odpowiedzi na pytania o WCAG, ARIA i najlepsze praktyki.")
-    gr.Markdown("ℹ️ *Pytania i odpowiedzi są zapisywane anonimowo w celu poprawy systemu i budowy zestawu danych treningowych.*")
 
-    with gr.Row():
-        # Left column: Q&A Form
-        with gr.Column(scale=1):
-            gr.Markdown("### 📋 Zadaj pytanie ekspertowi")
+# --- UI ---
+SHARE_URL = get_settings().public_url
 
-            question_input = gr.Textbox(
-                label="Twoje pytanie lub problem:",
-                placeholder="Opisz problem z dostępnością, zapytaj o wymagania WCAG, poproś o przegląd kodu... (Enter wysyła, Shift+Enter = nowa linia)",
+with gr.Blocks(title="Jacek AI") as demo:
+    gr.Markdown("# 🎯 Jacek AI")
+    gr.Markdown(
+        "**Konwersacyjny ekspert dostępności cyfrowej** — zadawaj pytania o WCAG, ARIA "
+        "i najlepsze praktyki. Asystent pamięta kontekst rozmowy."
+    )
+    gr.Markdown("ℹ️ *Rozmowy są zapisywane anonimowo w celu poprawy systemu.*")
+
+    with gr.Row(equal_height=False):
+        # Left: chat
+        with gr.Column(scale=3):
+            chatbot_widget = gr.Chatbot(
+                value=[],
+                label="Jacek AI",
+                show_label=False,
+                layout="bubble",
+                height=520,
+                render_markdown=True,
+                placeholder=(
+                    "<br><br><center><b>🎯 Jacek AI</b><br>"
+                    "Ekspert dostępności cyfrowej<br><br>"
+                    "Zadaj pytanie o WCAG, ARIA lub dostępność</center>"
+                ),
+            )
+            textbox_widget = gr.Textbox(
+                placeholder="Wpisz pytanie… (Enter wysyła, Shift+Enter = nowa linia)",
                 lines=1,
-                max_lines=8,
-                max_length=1000,
-                show_label=True
+                max_lines=5,
+                max_length=2000,
+                show_label=False,
+                container=False,
             )
 
-            submit_btn = gr.Button("🔍 Zapytaj eksperta", variant="primary", size="lg")
-
-            gr.Markdown("---")
-            gr.Markdown("### 📊 Odpowiedź eksperta:")
-
-            answer_output = gr.Textbox(
-                value="*Odpowiedź pojawi się tutaj po zadaniu pytania...*",
-                show_label=False,
-                elem_id="answer_output",
-                lines=15,
-                max_lines=30,
-                interactive=False,
-                container=False
+            chat_interface = gr.ChatInterface(
+                fn=chat,
+                chatbot=chatbot_widget,
+                textbox=textbox_widget,
+                autofocus=True,
             )
 
-            # Action buttons
             with gr.Row():
-                copy_btn = gr.Button("📋 Kopiuj odpowiedź", variant="secondary", size="sm")
                 share_twitter = gr.Button("🐦 X/Twitter", variant="secondary", size="sm")
                 share_linkedin = gr.Button("💼 LinkedIn", variant="secondary", size="sm")
                 share_facebook = gr.Button("📘 Facebook", variant="secondary", size="sm")
 
-            gr.Markdown("---")
-        
-        # Right column: Markdown content from file
+        # Right: notes
         with gr.Column(scale=1):
             gr.Markdown("### 📝 Notatki")
-            
-            def load_notes():
-                """Load notes from notes.md file."""
-                try:
-                    with open("notes.md", "r", encoding="utf-8") as f:
-                        return f.read()
-                except FileNotFoundError:
-                    return """
-## Witaj w Jacek AI! 👋
-
-Stwórz plik `notes.md` w katalogu projektu aby zobaczyć tutaj swoje notatki.
-
-### Przydatne linki:
-- [WCAG 2.2 Guidelines](https://www.w3.org/WAI/WCAG22/quickref/)
-- [ARIA Authoring Practices](https://www.w3.org/WAI/ARIA/apg/)
-- [MDN Accessibility](https://developer.mozilla.org/en-US/docs/Web/Accessibility)
-                    """
-                except Exception as e:
-                    return f"⚠️ Błąd wczytywania notes.md: {e}"
-            
-            markdown_content = gr.Markdown(
-                value=load_notes(),
-                show_label=False,
-                elem_id="notes_display"
-            )
-            
-            refresh_btn = gr.Button("🔄 Odśwież notatki", variant="secondary")
-            refresh_btn.click(
-                fn=load_notes,
-                outputs=markdown_content
-            )
-    
-    # Q&A logic
-    def handle_question(question):
-        """Process question and return answer."""
-        if not question or not question.strip():
-            return "⚠️ Proszę wpisać pytanie."
+            notes_display = gr.Markdown(value=load_notes(), show_label=False)
+            refresh_btn = gr.Button("🔄 Odśwież notatki", variant="secondary", size="sm")
 
-        # Show loading message
-        yield "⏳ Analizuję pytanie i przeszukuję bazę wiedzy..."
-
-        # Generate answer (streaming)
-        for chunk in respond(question):
-            yield chunk
-
-    # Wire up the Q&A form
-    submit_btn.click(
-        fn=handle_question,
-        inputs=question_input,
-        outputs=answer_output
-    )
-
-    question_input.submit(
-        fn=handle_question,
-        inputs=question_input,
-        outputs=answer_output
-    )
-
-    # Copy button - copies the answer text (no alert)
-    copy_btn.click(
-        fn=None,
-        inputs=answer_output,
-        outputs=None,
-        js="""
-        (answer) => {
-            navigator.clipboard.writeText(answer);
-        }
-        """
-    )
-
-    # Social media share buttons
+    # Social sharing
     share_twitter.click(
         fn=None,
-        inputs=[question_input, answer_output],
-        outputs=None,
         js=f"""
-        (question, answer) => {{
-            // Prepare tweet with full question and answer (preserve paragraph breaks)
-            const cleanAnswer = answer
-                .replace(/#+\\s*/g, '')  // Remove markdown headers
-                .replace(/\\n{{3,}}/g, '\\n\\n')  // Multiple newlines -> double newline
-                .replace(/\\*\\*/g, '')  // Remove bold markdown
-                .trim();
-
-            const text = `Zapytałem Jacek AI: "${{question}}"\\n\\n${{cleanAnswer}}\\n\\n🔗 `;
-            const url = '{SHARE_URL}';
-
-            // Twitter automatically appends URL, but we add it in text too for clarity
-            window.open(`https://twitter.com/intent/tweet?text=${{encodeURIComponent(text + url)}}`, '_blank');
+        () => {{
+            const text = 'Rozmawiam z Jacek AI — ekspertem dostępności cyfrowej (WCAG, ARIA) 🔗 {SHARE_URL}';
+            window.open('https://twitter.com/intent/tweet?text=' + encodeURIComponent(text), '_blank');
         }}
         """
     )
-
     share_linkedin.click(
         fn=None,
-        inputs=[question_input, answer_output],
-        outputs=None,
-        js=f"""
-        (question, answer) => {{
-            const url = '{SHARE_URL}';
-            // LinkedIn API limitation: cannot pre-fill post text (security policy)
-            // Only URL is shared - LinkedIn scrapes Open Graph meta tags for preview
-            window.open(`https://www.linkedin.com/sharing/share-offsite/?url=${{encodeURIComponent(url)}}`, '_blank');
-        }}
-        """
+        js=f"() => window.open('https://www.linkedin.com/sharing/share-offsite/?url=' + encodeURIComponent('{SHARE_URL}'), '_blank')"
     )
-
     share_facebook.click(
         fn=None,
-        inputs=[question_input, answer_output],
-        outputs=None,
-        js=f"""
-        (question, answer) => {{
-            const url = '{SHARE_URL}';
-            // Facebook API limitation: cannot pre-fill post text (security policy)
-            // Only URL is shared - Facebook scrapes Open Graph meta tags for preview
-            window.open(`https://www.facebook.com/sharer/sharer.php?u=${{encodeURIComponent(url)}}`, '_blank');
-        }}
-        """
+        js=f"() => window.open('https://www.facebook.com/sharer/sharer.php?u=' + encodeURIComponent('{SHARE_URL}'), '_blank')"
     )
+    refresh_btn.click(fn=load_notes, outputs=notes_display)
 
 
-# --- App Launch ---
-# Register cleanup handler
-# atexit.register(cleanup_resources) # Disabled: Causes premature shutdown on Hugging Face Spaces
-
-# Start background initialization
-logger.info("🚀 Starting Gradio app with background agent initialization...")
+# --- Launch ---
+logger.info("🚀 Starting Jacek AI (conversational mode)...")
 init_thread = threading.Thread(target=initialize_agent_background, daemon=True)
 init_thread.start()
-logger.info("ℹ️ Agent initialization started in background")
 
-# For Hugging Face Spaces, we need to either:
-# 1. Have a variable named 'demo' (which we have)
-# 2. Or explicitly call demo.queue() to enable the app
-# We'll use queue() to ensure proper startup
 if __name__ == "__main__":
     demo.queue()
     demo.launch()
 else:
-    # On HF Spaces, just ensure demo is ready
     demo.queue()