UI/UX: RTL layout, dark mode, stable header logo, chat/input RTL, scrollbar, focus states; agent: retrieve→generate→end, final answer separator

.gitignore

@@ -0,0 +1,30 @@
+# Secrets – never commit to cloud
+.env
+.env.local
+.env.*.local
+*.token
+.secrets
+hf_token*
+
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+venv/
+.venv/
+env/
+
+# IDE / OS
+.idea/
+.vscode/
+.DS_Store
+*.swp
+*.swo
+
+# Build / cache
+*.egg-info/
+dist/
+build/
+.cache/

agent.py

@@ -1,6 +1,6 @@
 """
-LangGraph agent: orchestrates RAG pipeline and verifies answer quality.
-If the planned answer is not good, the agent directs the pipeline toward correction (retry with feedback).
+LangGraph agent: orchestrates RAG pipeline (retrieve → generate → end).
+Single LLM call so the user always gets a final verbal answer that aggregates the retrieved context.
 """
 
 from typing import Optional, List, TypedDict, Literal
@@ -17,16 +17,12 @@ class AgentState(TypedDict, total=False):
     user_prompt: Optional[str]
     steps_log: List[str]
     draft_answer: Optional[str]
-    is_good: bool
     feedback: Optional[str]
     iteration: int
 
 
-MAX_REFINE_ITERATIONS = 2
-
-
 def build_agent_graph(engine: RAGEngine):
-    """Build the LangGraph: retrieve -> generate -> evaluate -> (end | generate with feedback)."""
+    """Build the LangGraph: retrieve → generate → end (one LLM call for final answer)."""
 
     def retrieve(state: AgentState) -> dict:
         """Run RAG up to (not including) LLM. Fill refusal or prompts + steps_log."""
@@ -63,76 +59,19 @@ def build_agent_graph(engine: RAGEngine):
         steps_log.append("✅ Draft generated")
         return {"draft_answer": draft, "steps_log": steps_log}
 
-    def evaluate(state: AgentState) -> dict:
-        """Check if the answer is good. Set is_good and optionally feedback for refinement."""
-        query = state["query"]
-        draft = state.get("draft_answer") or ""
-        steps_log = list(state.get("steps_log") or [])
-        iteration = state.get("iteration", 0)
-
-        # Error / timeout / rate limit responses are not "good" but we don't refine them
-        if draft.startswith("⚠️") or draft.startswith("❌") or draft.startswith("⏱️"):
-            return {"is_good": True, "steps_log": steps_log}  # Treat as final
-
-        steps_log.append("🔍 Evaluating answer quality...")
-        engine.configure_api(state["api_key"])
-        eval_prompt = f"""You are a quality checker. Given the user question and the assistant's answer, decide if the answer is good.
-
-User question: {query[:300]}
-
-Assistant answer: {draft[:1500]}
-
-Reply with exactly one of:
-- YES
-- NO: <one short line explaining what to improve>
-
-Reply:"""
-        models = ["gemini-2.0-flash", "gemini-1.5-flash"]
-        try:
-            raw = engine._call_api_with_backoff(
-                "You reply only with YES or NO: <feedback>. No other text.",
-                eval_prompt,
-                models,
-            )
-        except Exception:
-            raw = "YES"
-        raw = (raw or "").strip().upper()
-        is_good = raw.startswith("YES") or "NO" not in raw[:10]
-        feedback = ""
-        if not is_good and "NO" in raw:
-            idx = raw.find(":")
-            if idx != -1:
-                feedback = raw[idx + 1 :].strip()[:200]
-            else:
-                feedback = "Improve relevance and completeness."
-
-        steps_log.append("✅ Good" if is_good else f"⚠️ Needs improvement: {feedback[:60]}...")
-        return {
-            "is_good": is_good,
-            "feedback": feedback if not is_good else None,
-            "iteration": iteration + 1,
-            "steps_log": steps_log,
-        }
-
     def route_after_retrieve(state: AgentState) -> Literal["end", "generate"]:
         if state.get("refusal"):
             return "end"
         return "generate"
 
-    def route_after_evaluate(state: AgentState) -> Literal["end", "generate"]:
-        if state.get("is_good") or (state.get("iteration") or 0) >= MAX_REFINE_ITERATIONS:
-            return "end"
-        return "generate"
-
+    # Pipeline: retrieve → generate → end (no evaluate/refine – one LLM call so user always gets final answer)
     workflow = StateGraph(AgentState)
     workflow.add_node("retrieve", retrieve)
     workflow.add_node("generate", generate)
-    workflow.add_node("evaluate", evaluate)
 
     workflow.set_entry_point("retrieve")
     workflow.add_conditional_edges("retrieve", route_after_retrieve, {"end": END, "generate": "generate"})
-    workflow.add_edge("generate", "evaluate")
-    workflow.add_conditional_edges("evaluate", route_after_evaluate, {"end": END, "generate": "generate"})
+    workflow.add_edge("generate", END)
 
     return workflow.compile()
 
@@ -156,12 +95,11 @@ def run_stream(engine: RAGEngine, graph, query: str, api_key: str):
             text = f"{text}\n\n{body}"
         yield text
 
-    # Final state: ensure we have a visible answer
+    # Final state: ensure user always sees the verbal answer (aggregated from pipeline)
     final_answer = (last_state.get("refusal") or last_state.get("draft_answer") or "").strip()
     steps_log = list(last_state.get("steps_log") or [])
     steps_log.append("✅ Done")
 
-    # If no answer at all, show a clear fallback (user expects a verbal, detailed response)
     if not final_answer:
         final_answer = (
             "לא התקבלה תשובה מהמודל. ייתכן שנחסמה או שהבקשה ארכה מדי. "
@@ -173,4 +111,6 @@ def run_stream(engine: RAGEngine, graph, query: str, api_key: str):
         engine.response_cache[cache_key] = final_answer
     engine._maintain_conversation_history(query, final_answer)
 
-    yield f"{chr(10).join(steps_log)}\n\n{final_answer}"
+    # One final yield: steps + clear separator + the verbal answer (so user always sees it)
+    steps_text = chr(10).join(steps_log)
+    yield f"{steps_text}\n\n--- התשובה ---\n\n{final_answer}"

app.py

@@ -102,7 +102,7 @@ def create_comparison_start():
     """Generate a comparison question template"""
     return "Compare between the cars: "
 
-# Premium, inviting UI – automotive vibe with warmth
+# Premium UI – RTL-first, dark mode, automotive vibe
 custom_css = """
 @import url('https://fonts.googleapis.com/css2?family=Outfit:wght@400;500;600;700;800&family=Heebo:wght@400;500;600;700&display=swap');
 
@@ -127,17 +127,40 @@ custom_css = """
     --font-body: 'Heebo', -apple-system, sans-serif;
 }
 
-body.dark, [data-theme="dark"], .dark {
-    --primary: #ea580c;
-    --primary-hover: #f97316;
-    --primary-soft: rgba(234, 88, 12, 0.2);
+/* Dark mode – full coverage so UI stays readable */
+body.dark, [data-theme="dark"], .dark, html.dark {
+    --primary: #f97316;
+    --primary-hover: #fb923c;
+    --primary-soft: rgba(249, 115, 22, 0.2);
+    --accent: #2dd4bf;
     --text-primary: #fafaf9;
     --text-secondary: #d6d3d1;
     --text-muted: #a8a29e;
-    --bg-app: linear-gradient(160deg, #1c1917 0%, #292524 100%);
-    --bg-surface: #292524;
-    --bg-chat: #1c1917;
+    --bg-app: #0c0a09;
+    --bg-surface: #1c1917;
+    --bg-chat: #141211;
     --border: #44403c;
+    --shadow-sm: 0 1px 2px rgba(0,0,0,0.3);
+    --shadow-md: 0 4px 12px rgba(0,0,0,0.4);
+    --shadow-lg: 0 12px 40px rgba(0,0,0,0.5);
+}
+
+body.dark .header-section, .dark .header-section,
+body.dark .chat-container, .dark .chat-container {
+    background: var(--bg-surface) !important;
+    border-color: var(--border) !important;
+}
+
+body.dark .gradio-chatbot, .dark .gradio-chatbot,
+body.dark [data-testid="chatbot"], .dark [data-testid="chatbot"] {
+    background: var(--bg-chat) !important;
+}
+
+body.dark .gr-box, .dark .gr-box,
+body.dark .gr-textbox textarea, .dark .gr-textbox textarea {
+    background: var(--bg-surface) !important;
+    border-color: var(--border) !important;
+    color: var(--text-primary) !important;
 }
 
 * {
@@ -153,9 +176,11 @@ body {
     max-width: 900px !important;
     margin: 0 auto !important;
     padding: 24px 16px 40px !important;
+    direction: rtl !important;
+    text-align: right !important;
 }
 
-/* Header – hero style */
+/* Header – logo area stable (no disappearing image) */
 .header-section {
     background: var(--bg-surface) !important;
     border-radius: var(--radius) !important;
@@ -164,6 +189,23 @@ body {
     text-align: center !important;
     box-shadow: var(--shadow-md) !important;
     border: 1px solid var(--border) !important;
+    direction: rtl !important;
+}
+
+.header-logo {
+    display: inline-flex !important;
+    align-items: center !important;
+    justify-content: center !important;
+    width: 64px !important;
+    height: 64px !important;
+    min-width: 64px !important;
+    min-height: 64px !important;
+    font-size: 2.5rem !important;
+    line-height: 1 !important;
+    background: var(--primary-soft) !important;
+    border-radius: var(--radius) !important;
+    margin-bottom: 12px !important;
+    border: 1px solid var(--border) !important;
 }
 
 .header-section h1 {
@@ -199,13 +241,15 @@ body {
     color: var(--text-muted) !important;
 }
 
-/* Chat container – card with shadow */
+/* Chat – RTL layout */
 .chat-container {
     background: var(--bg-surface) !important;
     border: 1px solid var(--border) !important;
     border-radius: var(--radius) !important;
     overflow: hidden !important;
     box-shadow: var(--shadow-lg) !important;
+    direction: rtl !important;
+    text-align: right !important;
 }
 
 .gradio-chatbot {
@@ -213,13 +257,34 @@ body {
     padding: 24px !important;
     height: 520px !important;
     border-radius: 0 !important;
+    direction: rtl !important;
 }
 
 [data-testid="chatbot"] {
     background: var(--bg-chat) !important;
+    direction: rtl !important;
+}
+
+/* Scrollbar – subtle in light/dark */
+.gradio-chatbot .overflow-y-auto,
+[data-testid="chatbot"] .overflow-y-auto {
+    direction: rtl !important;
+}
+
+.gradio-chatbot::-webkit-scrollbar, [data-testid="chatbot"]::-webkit-scrollbar {
+    width: 8px !important;
+}
+
+.gradio-chatbot::-webkit-scrollbar-track, [data-testid="chatbot"]::-webkit-scrollbar-track {
+    background: var(--bg-chat) !important;
+}
+
+.gradio-chatbot::-webkit-scrollbar-thumb, [data-testid="chatbot"]::-webkit-scrollbar-thumb {
+    background: var(--border) !important;
+    border-radius: 4px !important;
 }
 
-/* Message bubbles */
+/* Message bubbles – RTL */
 .message {
     padding: 14px 0 !important;
     margin: 0 !important;
@@ -229,23 +294,24 @@ body {
     border: none !important;
     background: transparent !important;
     border-radius: 0 !important;
+    direction: rtl !important;
+    text-align: right !important;
 }
 
 .message.user {
-    margin: 12px 0 !important;
+    margin: 12px 0 12px auto !important;
     padding: 14px 18px !important;
     text-align: right !important;
     background: linear-gradient(135deg, var(--primary-soft) 0%, rgba(13, 148, 136, 0.08) 100%) !important;
-    border-radius: var(--radius-sm)  var(--radius-sm) 4px var(--radius-sm) !important;
-    border: 1px solid rgba(194, 65, 12, 0.15) !important;
+    border-radius: var(--radius-sm) 4px var(--radius-sm) var(--radius-sm) !important;
+    border: 1px solid rgba(194, 65, 12, 0.2) !important;
     max-width: 85% !important;
-    margin-left: auto !important;
 }
 
 .message.assistant {
-    margin: 12px 0 !important;
+    margin: 12px auto 12px 0 !important;
     padding: 16px 18px !important;
-    text-align: left !important;
+    text-align: right !important;
     background: var(--bg-surface) !important;
     border-radius: 4px var(--radius-sm) var(--radius-sm) var(--radius-sm) !important;
     border: 1px solid var(--border) !important;
@@ -258,6 +324,8 @@ body {
     background: transparent !important;
     font-size: 15px !important;
     line-height: 1.7 !important;
+    direction: rtl !important;
+    text-align: right !important;
 }
 
 .message a {
@@ -277,14 +345,25 @@ body {
     font-family: var(--font-head) !important;
 }
 
-/* Input area – pill style */
+/* Separator "--- התשובה ---" inside message */
+.message mark, .message [data-sep] {
+    display: block !important;
+    margin: 12px 0 !important;
+    padding: 8px 0 !important;
+    color: var(--primary) !important;
+    font-weight: 600 !important;
+    border-bottom: 1px solid var(--border) !important;
+}
+
+/* Input – RTL pill */
 .gr-box.gr-input-box {
     background: var(--bg-surface) !important;
     border: 1px solid var(--border) !important;
     border-radius: 999px !important;
     margin: 12px 16px 16px !important;
-    padding: 8px 8px 8px 20px !important;
+    padding: 8px 20px 8px 8px !important;
     box-shadow: var(--shadow-sm) !important;
+    direction: rtl !important;
 }
 
 .gr-textbox textarea {
@@ -295,6 +374,8 @@ body {
     font-size: 15px !important;
     min-height: 48px !important;
     border-radius: 0 !important;
+    direction: rtl !important;
+    text-align: right !important;
 }
 
 .gr-textbox textarea:focus {
@@ -306,7 +387,7 @@ body {
     color: var(--text-muted) !important;
 }
 
-/* Primary button – Send */
+/* Buttons */
 .gr-button.primary {
     background: linear-gradient(135deg, var(--primary) 0%, var(--primary-hover) 100%) !important;
     color: white !important;
@@ -323,9 +404,18 @@ body {
     box-shadow: 0 4px 14px rgba(194, 65, 12, 0.4) !important;
 }
 
-/* Example chips */
-.gr-examples .gr-sample-button,
-.gr-sample-button {
+.gr-button:focus-visible, .gr-textbox textarea:focus-visible {
+    outline: 2px solid var(--primary) !important;
+    outline-offset: 2px !important;
+}
+
+/* Example chips – RTL */
+.gr-examples, .gr-samples {
+    direction: rtl !important;
+    text-align: right !important;
+}
+
+.gr-examples .gr-sample-button, .gr-sample-button {
     background: var(--bg-surface) !important;
     border: 1px solid var(--border) !important;
     color: var(--text-primary) !important;
@@ -333,36 +423,35 @@ body {
     padding: 10px 18px !important;
     font-size: 0.9rem !important;
     transition: all 0.2s ease !important;
+    direction: rtl !important;
 }
 
-.gr-examples .gr-sample-button:hover,
-.gr-sample-button:hover {
+.gr-examples .gr-sample-button:hover, .gr-sample-button:hover {
     border-color: var(--primary) !important;
     color: var(--primary) !important;
     background: var(--primary-soft) !important;
 }
 
-/* Stop button */
-.gr-button-secondary,
-.gr-button:not(.primary) {
+.gr-button-secondary, .gr-button:not(.primary) {
     border-radius: var(--radius-sm) !important;
     font-weight: 500 !important;
 }
 
-/* Tables in answers */
+/* Tables – RTL */
 table {
     border-collapse: collapse !important;
     width: 100% !important;
     margin: 14px 0 !important;
     border-radius: var(--radius-sm) !important;
     overflow: hidden !important;
+    direction: rtl !important;
 }
 
 table th {
     background: var(--primary) !important;
     color: white !important;
     padding: 12px 14px !important;
-    text-align: left !important;
+    text-align: right !important;
     font-weight: 600 !important;
 }
 
@@ -370,6 +459,7 @@ table td {
     padding: 12px 14px !important;
     border-bottom: 1px solid var(--border) !important;
     color: var(--text-primary) !important;
+    text-align: right !important;
 }
 
 /* Code */
@@ -388,6 +478,8 @@ pre {
     padding: 14px !important;
     border-radius: var(--radius-sm) !important;
     overflow-x: auto !important;
+    direction: ltr !important;
+    text-align: left !important;
 }
 
 pre code {
@@ -420,8 +512,9 @@ with gr.Blocks(theme=theme, css=custom_css, title="CarsRUS – השוואת רכ
     
     with gr.Column(elem_classes="header-section"):
         gr.HTML("""
-        <div>
-            <h1>🚗 CarsRUS</h1>
+        <div dir="rtl" style="text-align: center;">
+            <div class="header-logo" aria-hidden="true">🚗</div>
+            <h1>CarsRUS</h1>
             <p class="header-tagline">השוואת רכבים והמלצות מבוססות כתבות מ־auto.co.il</p>
             <p class="header-note">
                 שאל על דגם בודד או בקש השוואה בין שנ�� דגמים. התשובות מבוססות רק על כתבות מבסיס הידע המקומי.

rag_engine.py

@@ -793,10 +793,8 @@ class RAGEngine:
         steps_log.append("🔍 Normalizing car names...")
         canonical = self._normalize_car_name(query)
         if canonical:
-            steps_log.append(f"✅ Recognized canonical id: {canonical}")
             search_query = canonical
         else:
-            steps_log.append("ℹ️ No canonical car found; using full query for search")
             search_query = query
 
         is_comparison = self._is_comparison_question(query)
@@ -806,6 +804,20 @@ class RAGEngine:
             steps_log.append("📋 Detected: single-model question (rule-based)")
 
         ordered_supported = self._get_ordered_supported_canonicals_in_text(query)
+        # Show user which cars were identified (for comparison: both; for single: one)
+        if is_comparison:
+            if len(ordered_supported) >= 2:
+                names = ", ".join(self.CANONICAL_TO_DISPLAY.get(c, c) for c in ordered_supported[:2])
+                steps_log.append(f"✅ זיהוי דגמים להשוואה: {names}")
+            elif len(ordered_supported) == 1:
+                one_display = self.CANONICAL_TO_DISPLAY.get(ordered_supported[0], ordered_supported[0])
+                steps_log.append(f"✅ זיהוי דגם אחד (השני לא ברשימה): {one_display}")
+        else:
+            if canonical:
+                steps_log.append(f"✅ Recognized canonical id: {canonical}")
+            elif not ordered_supported:
+                steps_log.append("ℹ️ No canonical car found; using full query for search")
+
         if is_comparison:
             if len(ordered_supported) == 0:
                 return (self._unsupported_car_refusal(query, is_comparison=True), None, None, steps_log)
@@ -921,10 +933,8 @@ Based on the context above, provide a clear answer that aggregates the informati
         processing_steps.append("🔍 Normalizing car names...")
         canonical = self._normalize_car_name(query)
         if canonical:
-            processing_steps.append(f"✅ Recognized canonical id: {canonical}")
             search_query = canonical
         else:
-            processing_steps.append("ℹ️ No canonical car found; using full query for search")
             search_query = query
 
         # עצה 7: זיהוי שאלות השוואתיות
@@ -932,6 +942,19 @@ Based on the context above, provide a clear answer that aggregates the informati
 
         # Policy guard: do not recommend models without auto.co.il articles in our KB
         ordered_supported = self._get_ordered_supported_canonicals_in_text(query)
+        # Show user which cars were identified (for comparison: both; for single: one)
+        if is_comparison:
+            if len(ordered_supported) >= 2:
+                names = ", ".join(self.CANONICAL_TO_DISPLAY.get(c, c) for c in ordered_supported[:2])
+                processing_steps.append(f"✅ זיהוי דגמים להשוואה: {names}")
+            elif len(ordered_supported) == 1:
+                one_display = self.CANONICAL_TO_DISPLAY.get(ordered_supported[0], ordered_supported[0])
+                processing_steps.append(f"✅ זיהוי דגם אחד (השני לא ברשימה): {one_display}")
+        else:
+            if canonical:
+                processing_steps.append(f"✅ Recognized canonical id: {canonical}")
+            else:
+                processing_steps.append("ℹ️ No canonical car found; using full query for search")
         if is_comparison:
             if len(ordered_supported) == 0:
                 return self._unsupported_car_refusal(query, is_comparison=True)
@@ -1072,12 +1095,9 @@ Based on the context above, provide a clear answer that aggregates the informati
 
         canonical = self._normalize_car_name(query)
         if canonical:
-            processing_steps.append(f"✅ Recognized canonical id: {canonical}")
             search_query = canonical
         else:
-            processing_steps.append("ℹ️ No canonical car found; using full query for search")
             search_query = query
-        yield steps_text()
 
         # --- Step 2: Question type (rule-based regex/keywords, no LLM) ---
         is_comparison = self._is_comparison_question(query)
@@ -1088,6 +1108,21 @@ Based on the context above, provide a clear answer that aggregates the informati
         yield steps_text()
 
         ordered_supported = self._get_ordered_supported_canonicals_in_text(query)
+        # Show user which cars were identified (for comparison: both; for single: one)
+        if is_comparison:
+            if len(ordered_supported) >= 2:
+                names = ", ".join(self.CANONICAL_TO_DISPLAY.get(c, c) for c in ordered_supported[:2])
+                processing_steps.append(f"✅ זיהוי דגמים להשוואה: {names}")
+            elif len(ordered_supported) == 1:
+                one_display = self.CANONICAL_TO_DISPLAY.get(ordered_supported[0], ordered_supported[0])
+                processing_steps.append(f"✅ זיהוי דגם אחד (השני לא ברשימה): {one_display}")
+        else:
+            if canonical:
+                processing_steps.append(f"✅ Recognized canonical id: {canonical}")
+            else:
+                processing_steps.append("ℹ️ No canonical car found; using full query for search")
+        yield steps_text()
+
         if is_comparison:
             if len(ordered_supported) == 0:
                 yield self._unsupported_car_refusal(query, is_comparison=True)