Update CrossTalk AI with premium UI

README.md

@@ -14,7 +14,7 @@ models:
 
 This Space runs the full CrossTalk AI trained hybrid retrieval system.
 
-It downloads the full trained artifacts from:
+It downloads trained artifacts from:
 
 `ElMETRICO/crosstalk-ai-full-artifacts`
 
@@ -25,3 +25,5 @@ It downloads the full trained artifacts from:
 - fine-tuned multilingual E5 semantic fallback
 - FAISS vector search
 - confidence-aware safe output handling
+
+Semantic fallback outputs are candidate suggestions only, not verified translations.

app.py

@@ -10,21 +10,13 @@ from sentence_transformers import SentenceTransformer
 
 MODEL_REPO_ID = "ElMETRICO/crosstalk-ai-full-artifacts"
 
-print("Downloading CrossTalk AI full artifacts from:", MODEL_REPO_ID)
-
-artifact_dir = Path(snapshot_download(
-    repo_id=MODEL_REPO_ID,
-    repo_type="model"
-))
+print("Downloading artifacts from:", MODEL_REPO_ID)
+artifact_dir = Path(snapshot_download(repo_id=MODEL_REPO_ID, repo_type="model"))
 
 MODEL_DIR = artifact_dir / "model" / "e5_lexical_contrastive_finetuned"
 INDEX_PATH = artifact_dir / "artifacts" / "trained_e5_source_to_meaning.index"
 DF_PATH = artifact_dir / "artifacts" / "trained_e5_source_to_meaning_df.csv"
 
-print("Model path:", MODEL_DIR)
-print("Index path:", INDEX_PATH)
-print("Data path:", DF_PATH)
-
 print("Loading fine-tuned E5 model...")
 model = SentenceTransformer(str(MODEL_DIR))
 model.max_seq_length = 128
@@ -32,7 +24,7 @@ model.max_seq_length = 128
 print("Loading FAISS index...")
 source_index = faiss.read_index(str(INDEX_PATH))
 
-print("Loading source dataframe...")
+print("Loading dataframe...")
 source_df = pd.read_csv(DF_PATH, encoding="utf-8-sig")
 print("Rows loaded:", len(source_df))
 
@@ -45,9 +37,7 @@ def clean_text(x):
 
 
 def normalize_lookup_text(x):
-    x = clean_text(x).lower()
-    x = re.sub(r"\s+", " ", x).strip()
-    return x
+    return clean_text(x).lower()
 
 
 def remove_parentheses_text(x):
@@ -59,9 +49,7 @@ def remove_parentheses_text(x):
 
 def count_source_files(x):
     x = "" if pd.isna(x) else str(x)
-    if not x.strip():
-        return 0
-    return len([p for p in x.split("||") if p.strip()])
+    return len([p for p in x.split("||") if p.strip()]) if x.strip() else 0
 
 
 def add_quality_score(df):
@@ -92,7 +80,7 @@ def add_quality_score(df):
 
 source_df = add_quality_score(source_df)
 
-for col in ["language", "source_text", "english_meaning", "bangla_meaning"]:
+for col in ["language", "source_text", "english_meaning", "bangla_meaning", "part_of_speech"]:
     if col not in source_df.columns:
         source_df[col] = ""
     source_df[col] = source_df[col].apply(clean_text)
@@ -103,7 +91,6 @@ source_df["base_source"] = source_df["source_text"].apply(remove_parentheses_tex
 
 def format_verified_result(df, query, method, top_k=10):
     result = df.copy()
-
     result = result.sort_values(
         by=["quality_score", "duplicate_count"],
         ascending=[False, False]
@@ -134,7 +121,6 @@ def format_verified_result(df, query, method, top_k=10):
         result["confidence"] = "high"
 
     result["note"] = "Verified dictionary match."
-
     return result
 
 
@@ -167,9 +153,7 @@ def trained_semantic_fallback(query, top_k=5, search_k_per_language=20):
                 continue
 
             if idx not in best_by_index or score > best_by_index[idx]["score"]:
-                best_by_index[idx] = {
-                    "score": score
-                }
+                best_by_index[idx] = {"score": score}
 
     ranked = sorted(
         best_by_index.items(),
@@ -196,7 +180,7 @@ def trained_semantic_fallback(query, top_k=5, search_k_per_language=20):
             return "low_trained_semantic_candidate"
 
     result["confidence"] = result["score"].apply(label_score)
-    result["note"] = "No verified dictionary match. Semantic candidate only; not a confirmed translation."
+    result["note"] = "Semantic candidate only; not a confirmed translation."
 
     keep_cols = [
         "query",
@@ -224,7 +208,7 @@ def safe_search(query):
     query = clean_text(query)
 
     if not query:
-        return "Please enter a word.", pd.DataFrame()
+        return "⚠️ Please enter a word.", "No input provided.", pd.DataFrame()
 
     q_norm = normalize_lookup_text(query)
     q_base = remove_parentheses_text(query)
@@ -233,62 +217,261 @@ def safe_search(query):
 
     if len(exact) > 0:
         result = format_verified_result(exact, query, "hybrid_exact_source_match", top_k=10)
-        return "Verified dictionary match found.", result
+        return (
+            "✅ Verified dictionary match found.",
+            f"High-confidence verified dictionary output for **{query}**.",
+            result
+        )
 
     base = source_df[source_df["base_source"] == q_base].copy()
 
     if len(base) > 0:
         result = format_verified_result(base, query, "hybrid_base_form_match", top_k=10)
-        return "Verified base-form dictionary match found.", result
+        return (
+            "✅ Verified base-form match found.",
+            f"High-confidence base-form dictionary output for **{query}**.",
+            result
+        )
 
     semantic = trained_semantic_fallback(query, top_k=5, search_k_per_language=20)
 
     if len(semantic) == 0:
-        return "No verified dictionary match or semantic candidate found.", pd.DataFrame()
+        return (
+            "❌ No match found.",
+            "No verified dictionary match or semantic candidate was found.",
+            pd.DataFrame()
+        )
 
     strong = semantic[semantic["score"] >= 0.88].copy()
 
     if len(strong) > 0:
-        return "No verified dictionary match found. Showing trained semantic candidates only.", strong
-
-    return "No verified dictionary match found. Semantic scores are below safe threshold; no translation is claimed.", semantic
+        return (
+            "🧠 Semantic candidates found.",
+            "No verified dictionary match was found. Showing fine-tuned E5 semantic candidates only. These are suggestions, not confirmed translations.",
+            strong
+        )
+
+    return (
+        "⚠️ Low-confidence semantic candidates.",
+        "No verified dictionary match was found. Semantic scores are below the safe verification threshold, so no translation is claimed.",
+        semantic
+    )
 
 
-description = """
-CrossTalk AI Full Deployment
+custom_css = """
+.gradio-container {
+    background:
+        radial-gradient(circle at 15% 5%, rgba(139,92,246,0.32), transparent 28%),
+        radial-gradient(circle at 90% 10%, rgba(168,85,247,0.22), transparent 26%),
+        linear-gradient(180deg, #07020F 0%, #090716 50%, #07020F 100%) !important;
+    color: #F8F5FF !important;
+    font-family: Inter, ui-sans-serif, system-ui, sans-serif !important;
+}
+
+#shell {
+    max-width: 1280px;
+    margin: auto;
+}
+
+.hero {
+    background: linear-gradient(135deg, rgba(18,16,38,0.96), rgba(37,22,78,0.82));
+    border: 1px solid rgba(255,255,255,0.10);
+    border-radius: 30px;
+    padding: 34px;
+    margin-bottom: 22px;
+    box-shadow: 0 26px 80px rgba(0,0,0,0.42);
+}
+
+.badge {
+    display: inline-block;
+    padding: 8px 13px;
+    border-radius: 999px;
+    background: rgba(139,92,246,0.16);
+    border: 1px solid rgba(139,92,246,0.36);
+    color: #EDE7FF;
+    font-weight: 800;
+    font-size: 13px;
+    margin-bottom: 14px;
+}
+
+.title {
+    font-size: 48px;
+    line-height: 1.02;
+    font-weight: 950;
+    letter-spacing: -0.05em;
+    margin: 0;
+    background: linear-gradient(135deg, #FFFFFF, #D9CCFF, #A78BFA);
+    -webkit-background-clip: text;
+    -webkit-text-fill-color: transparent;
+}
+
+.subtitle {
+    max-width: 980px;
+    margin-top: 15px;
+    color: #BDB2DE;
+    font-size: 16px;
+    line-height: 1.7;
+}
+
+.metrics {
+    display: grid;
+    grid-template-columns: repeat(4, 1fr);
+    gap: 14px;
+    margin-top: 24px;
+}
+
+.metric {
+    padding: 17px;
+    border-radius: 20px;
+    background: rgba(255,255,255,0.055);
+    border: 1px solid rgba(255,255,255,0.09);
+}
+
+.metric strong {
+    display: block;
+    color: #FFFFFF;
+    font-size: 25px;
+    font-weight: 950;
+}
+
+.metric span {
+    color: #B8ADD8;
+    font-size: 12px;
+}
+
+.panel {
+    background: rgba(18,16,38,0.78) !important;
+    border: 1px solid rgba(255,255,255,0.10) !important;
+    border-radius: 24px !important;
+    padding: 18px !important;
+    box-shadow: 0 20px 55px rgba(0,0,0,0.30);
+}
+
+button.primary-btn {
+    background: linear-gradient(135deg, #8B5CF6, #A855F7) !important;
+    border: none !important;
+    border-radius: 16px !important;
+    color: white !important;
+    font-weight: 900 !important;
+    min-height: 48px !important;
+}
+
+button.secondary-btn {
+    background: rgba(255,255,255,0.08) !important;
+    border: 1px solid rgba(255,255,255,0.12) !important;
+    border-radius: 16px !important;
+    color: white !important;
+    font-weight: 800 !important;
+    min-height: 48px !important;
+}
+
+textarea, input {
+    background: rgba(5,4,14,0.78) !important;
+    color: white !important;
+    border: 1px solid rgba(255,255,255,0.14) !important;
+    border-radius: 16px !important;
+}
+
+.footer {
+    text-align: center;
+    color: #A99FD0;
+    font-size: 13px;
+    line-height: 1.6;
+    margin-top: 18px;
+}
+
+@media (max-width: 900px) {
+    .title { font-size: 34px; }
+    .metrics { grid-template-columns: repeat(2, 1fr); }
+}
+"""
 
-This is the full trained hybrid lexical retrieval system:
-1. exact dictionary matching
-2. base-form matching
-3. fine-tuned multilingual E5 semantic fallback
-4. FAISS vector search
-5. confidence-aware safe output handling
 
-Semantic fallback results are candidate suggestions only, not verified translations.
-"""
+with gr.Blocks(
+    css=custom_css,
+    theme=gr.themes.Soft(
+        primary_hue="violet",
+        secondary_hue="purple",
+        neutral_hue="slate"
+    )
+) as demo:
+
+    with gr.Column(elem_id="shell"):
+        gr.HTML(
+            """
+            <div class="hero">
+                <div class="badge">🌐 Full trained hybrid retrieval system</div>
+                <h1 class="title">CrossTalk AI</h1>
+                <div class="subtitle">
+                    A confidence-aware lexical retrieval and translation support system for low-resource ethnic languages of Bangladesh.
+                    It combines exact dictionary matching, base-form matching, fine-tuned multilingual E5 semantic fallback,
+                    FAISS vector search, and safe output handling.
+                </div>
+                <div class="metrics">
+                    <div class="metric"><strong>12</strong><span>Ethnic language groups</span></div>
+                    <div class="metric"><strong>70K+</strong><span>Indexed source entries</span></div>
+                    <div class="metric"><strong>E5</strong><span>Fine-tuned semantic retriever</span></div>
+                    <div class="metric"><strong>FAISS</strong><span>Vector search backend</span></div>
+                </div>
+            </div>
+            """
+        )
+
+        with gr.Row():
+            with gr.Column(scale=5, elem_classes=["panel"]):
+                gr.Markdown("### Search source word")
+                query = gr.Textbox(
+                    label="Input word",
+                    placeholder="Try: kəkhyáŋ, Hula, Aina, aam, bajaoo",
+                    lines=1
+                )
+
+                with gr.Row():
+                    clear_btn = gr.Button("Clear", elem_classes=["secondary-btn"])
+                    submit_btn = gr.Button("Search", elem_classes=["primary-btn"])
+
+                gr.Examples(
+                    examples=[
+                        ["kəkhyáŋ"],
+                        ["Hula"],
+                        ["Aina"],
+                        ["aam"],
+                        ["bajaoo"],
+                        ["unknown tribal word"]
+                    ],
+                    inputs=query,
+                    label="Quick examples"
+                )
+
+            with gr.Column(scale=7, elem_classes=["panel"]):
+                gr.Markdown("### Retrieval output")
+                status = gr.Textbox(label="System status", lines=1)
+                summary = gr.Markdown(
+                    value="Enter a source word to retrieve verified dictionary matches or semantic candidates."
+                )
+                results = gr.Dataframe(
+                    label="Results",
+                    interactive=False,
+                    wrap=True
+                )
+
+        gr.HTML(
+            """
+            <div class="footer">
+                Exact and base-form matches are verified dictionary outputs.
+                Fine-tuned semantic fallback results are candidate suggestions only and should not be treated as confirmed translations.
+            </div>
+            """
+        )
+
+    submit_btn.click(fn=safe_search, inputs=query, outputs=[status, summary, results])
+    query.submit(fn=safe_search, inputs=query, outputs=[status, summary, results])
+    clear_btn.click(
+        fn=lambda: ("", "Ready.", "Enter a source word to retrieve verified dictionary matches or semantic candidates.", pd.DataFrame()),
+        inputs=None,
+        outputs=[query, status, summary, results]
+    )
 
-demo = gr.Interface(
-    fn=safe_search,
-    inputs=gr.Textbox(
-        label="Enter source / ethnic word",
-        placeholder="Example: kəkhyáŋ, Hula, Aina"
-    ),
-    outputs=[
-        gr.Textbox(label="System Message"),
-        gr.Dataframe(label="Results")
-    ],
-    title="CrossTalk AI Full",
-    description=description,
-    examples=[
-        ["kəkhyáŋ"],
-        ["Hula"],
-        ["Aina"],
-        ["aam"],
-        ["bajaoo"],
-        ["unknown tribal word"]
-    ],
-    flagging_mode="never"
-)
 
 if __name__ == "__main__":
     demo.launch()