ChogwuCY commited on
Commit
6f71d15
·
verified ·
1 Parent(s): 08cfa45

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +344 -118
app.py CHANGED
@@ -1,142 +1,368 @@
1
- import math, re, json
 
2
  from pathlib import Path
3
- import numpy as np
4
  import gradio as gr
5
- from transformers import pipeline
6
  import joblib
 
 
 
 
 
7
 
8
- REPO = "ChogwuCY/salatus-qa-xlmrb-base" # your model on the Hub
9
-
10
- # ---- Bright UI ----
11
- BRAND_PRIMARY = "#60a5fa"; BRAND_ACCENT="#f59e0b"; BRAND_TEXT="#1f2937"; BRAND_BG="#f9fafb"; HILITE_BG="#fff3a3"
12
- CSS = f"""
13
- :root {{ --brand-primary:{BRAND_PRIMARY}; --brand-accent:{BRAND_ACCENT}; --brand-text:{BRAND_TEXT}; --brand-bg:{BRAND_BG}; --hilite-bg:{HILITE_BG}; }}
14
- .gradio-container {{ font-family: ui-sans-serif, system-ui, Segoe UI, Roboto, Helvetica, Arial; color:var(--brand-text); background:var(--brand-bg); }}
15
- #hero {{ background: linear-gradient(120deg, #93c5fd 0%, #60a5fa 55%, #3b82f6 100%); color:#fff; border-radius:14px; padding:22px; margin-bottom:12px; display:flex; align-items:center; gap:16px; }}
16
- #hero .title {{ color:#fff; font-weight:800; font-size:1.55rem; line-height:1.15; text-shadow:0 1px 2px rgba(0,0,0,.2); }}
17
- #hero .subtitle {{ color:#eef5ff; font-size:.98rem; text-shadow:0 1px 2px rgba(0,0,0,.25); }}
18
- #hero .pill {{ color:#eaf2ff; background:rgba(255,255,255,.14); border:1px solid rgba(255,255,255,.25); padding:4px 10px; border-radius:999px; font-size:.80rem; }}
19
- button, .gr-button {{ border-radius:12px !important; font-weight:700 !important; border:0 !important; background:var(--brand-accent) !important; color:#fff !important; }}
20
- button:hover, .gr-button:hover {{ filter:brightness(1.06); }}
21
- textarea, input, .gr-textbox, .gr-box {{ border-radius:12px !important; }}
22
- .markdown-body mark, mark {{ background:var(--hilite-bg); padding:0 .15em; border-radius:.2em; }}
23
- .card {{ background:#fff; border:1px solid #e5e7eb; border-radius:12px; padding:14px; }}
24
- .footer-note {{ color:#6b7280; font-size:.85rem; }}
25
- """
26
 
27
- # ---- Load QA model
28
- qa = pipeline("question-answering", model=REPO, tokenizer=REPO)
 
 
 
 
 
 
29
 
30
- # ---- Load retriever assets (only 2 files)
31
- RDIR = Path("retriever")
32
- PIPE = joblib.load(RDIR / "tfidf_pipe.joblib") # fitted (TfidfVectorizer + Normalizer)
33
- CONTEXTS = json.loads((RDIR / "contexts.json").read_text(encoding="utf-8"))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
34
 
35
- # Precompute matrix on startup (re-transform same contexts). This avoids uploading a separate X.npz.
36
- X = PIPE.transform(CONTEXTS) # rows are L2-normalized (because pipeline includes Normalizer)
37
 
 
 
 
38
  APOS = {"’": "'", "‘": "'", "“": '"', "”": '"'}
 
 
 
 
 
 
39
  def norm(s: str) -> str:
40
- s = s.replace("\u200c","").replace("\u200b","")
 
41
  for k,v in APOS.items(): s = s.replace(k,v)
42
  return re.sub(r"\s+", " ", s).strip()
43
 
44
- TITLE_LEFT = re.compile(r"(Raja|Sultan|Tun|Tuan|Puteri|Megat|Hang|Wan)\s*$", re.I)
45
- RIGHT_TITLES = (" Syah", " Shah", " Riayat Shah")
46
-
47
- def retrieve(question: str, top_k: int = 5):
48
- qv = PIPE.transform([norm(question)]) # L2-normalized
49
- sims = (X @ qv.T).toarray().ravel() # cosine = dot, since rows are normalized
50
- idx = np.argsort(-sims)[:top_k]
51
- return [(int(i), float(sims[i]), CONTEXTS[i]) for i in idx]
52
-
53
- def extract(q, c, max_len):
54
- outs = qa({"question": q, "context": c}, top_k=3, max_answer_len=int(max_len), truncation=True)
55
- if isinstance(outs, dict): outs = [outs]
56
- best = max(outs, key=lambda x: x["score"])
57
- s, e = best.get("start", -1), best.get("end", -1)
58
-
59
- # expand titles/honorifics
60
- if isinstance(s, int) and isinstance(e, int) and s>=0 and e>s:
61
- left = c[max(0, s-25):s]; m = TITLE_LEFT.search(left)
62
- if m: s -= len(m.group(0))
63
- if 0<=s<e<=len(c):
64
- for rt in RIGHT_TITLES:
65
- t = e + len(rt)
66
- if t <= len(c) and c[e:t] == rt:
67
- e = t; break
68
- if 0<=s<e<=len(c): ans = c[s:e]
69
- else: ans = best["answer"]
70
- else:
71
- ans = best["answer"]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
72
 
73
- # highlight
74
- if isinstance(s, int) and isinstance(e, int) and 0<=s<e<=len(c):
75
- highlighted = c[:s] + "<mark>" + c[s:e] + "</mark>" + c[e:]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
76
  else:
77
- highlighted = c.replace(ans, f"<mark>{ans}</mark>", 1)
 
 
 
 
 
 
 
 
 
 
78
 
79
- # friendly confidences
80
- mx = max(o["score"] for o in outs)
81
- exps = [math.exp(o["score"] - mx) for o in outs]; den = sum(exps) or 1.0
82
- tops = "\n".join([f"- {o['answer']} (confidence={(exps[i]/den):.1%})" for i,o in enumerate(outs)])
83
- return ans, highlighted, tops
84
 
85
- def ask_only(question, top_k, max_len, threshold):
86
- cands = retrieve(question, top_k=top_k)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
87
  if not cands:
88
- return "—", "", "Tiada konteks ditemui.", ""
89
- best_idx, best_sim, best_ctx = cands[0]
90
- if best_sim < threshold:
91
- return "—", "", "Konteks kurang pasti (skor rendah). Cuba jelaskan soalan.", f"Best sim: {best_sim:.3f}"
92
- ans, hl, tops = extract(question, best_ctx, max_len)
93
- md_lines = []
94
- for i, s, _ in cands[:3]:
95
- snippet = CONTEXTS[i][:120].replace("\n", " ")
96
- md_lines.append(f"- {i} (sim={s:.3f}) {snippet}…")
97
- md = "\n".join(md_lines)
98
- return ans, hl, tops, md
99
-
100
- with gr.Blocks(title="Salatus QA — Question-Only (RAG)", css=CSS, theme=gr.themes.Soft(primary_hue="slate", secondary_hue="amber")) as demo:
101
- with gr.Row():
102
- gr.HTML("""
103
  <div id="hero">
104
- <div style="font-size:1.6rem">🧭</div>
105
  <div>
106
- <div class="title">Salatus QA — Question-Only</div>
107
- <div class="subtitle">Masukkan soalan sahaja. Aplikasi memilih konteks terbaik, kemudian mengekstrak jawapan.</div>
 
 
 
108
  </div>
109
- <div style="margin-left:auto"><span class="pill">RAG Demo</span></div>
110
  </div>
111
- """)
112
-
113
- with gr.Tab("Demo"):
114
- q = gr.Textbox(label="Soalan / Question", lines=2, placeholder="Tulis soalan…")
115
- with gr.Row():
116
- top_k = gr.Slider(1, 10, value=5, step=1, label="Bilangan konteks dicadang (top_k)")
117
- max_len = gr.Slider(8, 24, value=14, step=1, label="Maks panjang jawapan (token)")
118
- thresh = gr.Slider(0.05, 0.40, value=0.12, step=0.01, label="Ambang persamaan (retrieval)")
119
- with gr.Row():
120
- btn = gr.Button("Tanya", size="lg")
121
- with gr.Row():
122
- with gr.Column(scale=1):
123
- ans = gr.Textbox(label="Jawapan terbaik", interactive=False)
124
- with gr.Column(scale=2):
125
- html = gr.HTML(label="Sorotan dalam Konteks (terpilih)")
126
- tops = gr.Textbox(label="Calon jawapan (Top-3)", interactive=False)
127
- rets = gr.Markdown(label="Konteks dicapai (Top-3)")
128
-
129
- gr.Examples(
130
- examples=[
131
- ["Siapakah yang dikahwinkan dengan Tuan Puteri Mahtabu’l-Bahri?"],
132
- ["Siapakah yang memerintahkan pihak Sultan Alau'd-Din?"],
133
- ],
134
- inputs=[q],
135
- outputs=[ans, html, tops, rets],
136
- fn=lambda Q: ask_only(Q, 5, 14, 0.12)
137
- )
138
- btn.click(ask_only, inputs=[q, top_k, max_len, thresh], outputs=[ans, html, tops, rets])
139
-
140
- gr.Markdown('<div class="footer-note">Nota: Jika skor persamaan rendah, cuba jelaskan soalan. “Maks panjang jawapan” mengawal panjang span yang diekstrak (contoh: 12–18 untuk nama panjang).</div>')
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
141
 
142
  demo.queue().launch()
 
1
+ # app.py — Salatus QA (Question-Only RAG, bilingual, bright theme)
2
+ import math, re, json, os
3
  from pathlib import Path
4
+
5
  import gradio as gr
 
6
  import joblib
7
+ import numpy as np
8
+ from sklearn.feature_extraction.text import TfidfVectorizer
9
+ from sklearn.pipeline import make_pipeline
10
+ from sklearn.preprocessing import Normalizer
11
+ from transformers import pipeline
12
 
13
+ # =========================
14
+ # CONFIG
15
+ # =========================
16
+ # Change this to your model repo if needed (Hugging Face Hub id)
17
+ HF_MODEL = os.getenv("MODEL_REPO", "ChogwuCY/salatus-qa-xlmrb-base")
 
 
 
 
 
 
 
 
 
 
 
 
 
18
 
19
+ # Colors (bright header + amber button)
20
+ CSS = """
21
+ :root {
22
+ --brand-primary: #1e3a8a; /* blue-800 */
23
+ --brand-secondary: #0ea5e9;/* sky-500 */
24
+ --brand-accent: #f59e0b; /* amber-500 */
25
+ --brand-bg: #f8fafc; /* slate-50 */
26
+ }
27
 
28
+ .gradio-container {
29
+ font-family: ui-sans-serif, system-ui, -apple-system, Segoe UI, Roboto, Helvetica, Arial;
30
+ background: var(--brand-bg);
31
+ }
32
+
33
+ #hero {
34
+ background: linear-gradient(100deg, var(--brand-primary), #2563eb 50%, var(--brand-secondary));
35
+ color: #fff;
36
+ border-radius: 16px;
37
+ padding: 18px 20px;
38
+ display: flex;
39
+ align-items: center;
40
+ gap: 14px;
41
+ }
42
+
43
+ #hero h1 { margin: 0; font-size: 1.45rem; font-weight: 800; letter-spacing: .2px; }
44
+ #hero p { margin: 2px 0 0; opacity: .95; }
45
+
46
+ .gr-button, button {
47
+ background: var(--brand-accent) !important;
48
+ color: #111 !important;
49
+ font-weight: 800 !important;
50
+ border-radius: 12px !important;
51
+ border: 0 !important;
52
+ }
53
+ .gr-button:hover { filter: brightness(1.05); }
54
+
55
+ .markdown-body mark, mark {
56
+ background: #fff59d;
57
+ padding: 0 .18em;
58
+ border-radius: .2em;
59
+ }
60
+
61
+ .card {
62
+ background: #fff;
63
+ border: 1px solid #e5e7eb;
64
+ border-radius: 12px;
65
+ padding: 12px;
66
+ }
67
 
68
+ textarea, input, .gr-textbox, .gr-box { border-radius: 12px !important; }
69
+ """
70
 
71
+ # =========================
72
+ # Utilities: normalisation / cleaning
73
+ # =========================
74
  APOS = {"’": "'", "‘": "'", "“": '"', "”": '"'}
75
+ LEAD_STOP = {"maka","oleh","akan","yang","dan","bahawa","adapun","pun","di","ke","dari"}
76
+ TAIL_STOP = {"itu","pun","lah","kah"}
77
+ NAME_PAT = re.compile(
78
+ r"(Raja|Sultan|Tun|Tuan|Puteri|Megat|Hang|Wan)\s+[A-ZĀ-ÿ][^\s,.;]{1,}(?:\s+[A-ZĀ-ÿ][^\s,.;]{1,}){0,3}"
79
+ )
80
+
81
  def norm(s: str) -> str:
82
+ if not s: return ""
83
+ s = s.replace("\u200b","").replace("\u200c","")
84
  for k,v in APOS.items(): s = s.replace(k,v)
85
  return re.sub(r"\s+", " ", s).strip()
86
 
87
+ def clean_edges(text: str) -> str:
88
+ t = text.strip(" .,:;—-")
89
+ toks = t.split()
90
+ while toks and toks[0].lower() in LEAD_STOP: toks = toks[1:]
91
+ while toks and toks[-1].lower() in TAIL_STOP: toks = toks[:-1]
92
+ return " ".join(toks)
93
+
94
+ def answer_sentence(context: str, name: str) -> str:
95
+ # Try to surface a natural clause mentioning name + dikahwinkan
96
+ m = re.search(rf"([^.]{re.escape(name)}[^.]*dikahwinkan[^.]\.)", context, flags=re.I)
97
+ if m:
98
+ s = re.sub(r"\b(Maka|oleh|akan)\b\s*", "", m.group(1), flags=re.I).strip()
99
+ return s
100
+ m2 = re.search(r"dikahwinkan dengan ([^.]*?)\.", context, flags=re.I)
101
+ if m2:
102
+ return f"{name} dikahwinkan dengan {m2.group(1)}."
103
+ return name
104
+
105
+ # =========================
106
+ # Retriever: load (or build from sources) at startup
107
+ # =========================
108
+ RDIR = Path("retriever")
109
+ RDIR.mkdir(exist_ok=True)
110
+ PIPE_PATH = RDIR / "tfidf_pipe.joblib"
111
+ CTX_PATH = RDIR / "contexts.json"
112
+
113
+ def load_or_build_retriever():
114
+ if PIPE_PATH.exists() and CTX_PATH.exists():
115
+ pipe = joblib.load(PIPE_PATH)
116
+ contexts = json.loads(CTX_PATH.read_text(encoding="utf-8"))
117
+ X = pipe.transform(contexts)
118
+ return pipe, contexts, X
119
+
120
+ # Try building from files shipped in repo root
121
+ candidates = [Path("contexts.json"), Path("contexts.jsonl"),
122
+ Path("salatus_qa.json"), Path("salatus_qa.jsonl")]
123
+ seen, contexts = set(), []
124
+ for p in candidates:
125
+ if not p.exists(): continue
126
+ if p.suffix.lower() == ".jsonl":
127
+ with p.open("r", encoding="utf-8") as f:
128
+ for line in f:
129
+ if not line.strip(): continue
130
+ obj = json.loads(line)
131
+ c = norm(obj.get("context",""))
132
+ if len(c) >= 40 and c not in seen:
133
+ seen.add(c); contexts.append(c)
134
+ else:
135
+ data = json.loads(p.read_text(encoding="utf-8"))
136
+ if isinstance(data, list):
137
+ # list of objects with context or list of strings
138
+ for ex in data:
139
+ c = norm(ex["context"] if isinstance(ex, dict) else str(ex))
140
+ if len(c) >= 40 and c not in seen:
141
+ seen.add(c); contexts.append(c)
142
+ if contexts: break
143
+
144
+ if not contexts:
145
+ # tiny fallback (so app still runs)
146
+ contexts = [
147
+ "Maka oleh Raja Aktabu'l-Ard akan Raja Suran dikahwinkan dengan anaknya Tuan Puteri Mahtabu'l-Bahri itu.",
148
+ "Hatta Sultan Alau'd-Din ... pihak Sultan Alau'd-Din, Bendahara yang memerintahkan."
149
+ ]
150
+
151
+ vec = TfidfVectorizer(analyzer="word", ngram_range=(1,2), min_df=1)
152
+ pipe = make_pipeline(vec, Normalizer(copy=False))
153
+ X = pipe.fit_transform(contexts)
154
+ # Save for next boot
155
+ joblib.dump(pipe, PIPE_PATH)
156
+ CTX_PATH.write_text(json.dumps(contexts, ensure_ascii=False, indent=2), encoding="utf-8")
157
+ return pipe, contexts, X
158
+
159
+ PIPE, CONTEXTS, X = load_or_build_retriever()
160
+
161
+ def retrieve(query: str, top_k: int = 5, threshold: float = 0.12):
162
+ qv = PIPE.transform([norm(query)])
163
+ sims = (X @ qv.T).toarray().ravel()
164
+ idx = np.argsort(-sims)[:max(1, int(top_k))]
165
+ out = [(int(i), float(sims[i]), CONTEXTS[int(i)]) for i in idx if sims[i] >= threshold]
166
+ return out
167
 
168
+ # =========================
169
+ # QA pipeline
170
+ # =========================
171
+ qa = pipeline("question-answering", model=HF_MODEL, tokenizer=HF_MODEL)
172
+
173
+ def answer_with_cleaning(q: str, contexts, max_len: int):
174
+ """
175
+ Run QA on each candidate context; pick the highest-score span.
176
+ Return (short_answer, sentence_answer, highlighted_context, top_candidates_md, retrieved_md)
177
+ """
178
+ if not contexts:
179
+ return "—", "", "", "—", "—"
180
+
181
+ best = None
182
+ cands_md_lines = []
183
+ for (i, sim, ctx) in contexts:
184
+ outs = qa({"question": q, "context": ctx}, top_k=3, max_answer_len=int(max_len), truncation=True)
185
+ if isinstance(outs, dict): outs = [outs]
186
+
187
+ # top span for this context
188
+ b = max(outs, key=lambda x: x["score"])
189
+ # accumulate candidate spans (softmax for display)
190
+ mx = max(o["score"] for o in outs)
191
+ exps = [math.exp(o["score"] - mx) for o in outs]; den = sum(exps) or 1.0
192
+ for j, o in enumerate(outs[:3]):
193
+ conf = exps[j] / den
194
+ cands_md_lines.append(f"- {o['answer']} (confidence={conf:.1%})")
195
+
196
+ # keep global best
197
+ if (best is None) or (b["score"] > best["score"]):
198
+ best = {"span": b, "ctx": ctx, "idx": i, "sim": sim}
199
+
200
+ raw = best["span"]["answer"]
201
+ s, e = best["span"].get("start", -1), best["span"].get("end", -1)
202
+
203
+ short = clean_edges(raw)
204
+ # prefer titled name near the span
205
+ if isinstance(s, int) and isinstance(e, int) and 0 <= s < e <= len(best["ctx"]):
206
+ window = best["ctx"][max(0, s-30):min(len(best["ctx"]), e+40)]
207
+ m = NAME_PAT.search(window)
208
+ if m: short = m.group(0).strip()
209
+
210
+ # natural sentence (optional)
211
+ sent = answer_sentence(best["ctx"], short) if short else ""
212
+
213
+ # highlight (avoid backslash in f-string expression)
214
+ ctx = best["ctx"]
215
+ pos = ctx.find(short) if short else -1
216
+ if pos != -1:
217
+ highlighted = ctx[:pos] + "<mark>" + short + "</mark>" + ctx[pos+len(short):]
218
  else:
219
+ # fallback to the raw span
220
+ raw_span = raw
221
+ highlighted = ctx.replace(raw_span, f"<mark>{raw_span}</mark>", 1) if raw_span else ctx
222
+
223
+ # retrieved list (first 3)
224
+ ret_lines = []
225
+ for (i, sim, _c) in contexts[:3]:
226
+ snip = _c[:120].replace("\n", " ")
227
+ ret_lines.append(f"{i} (sim={sim:.3f}) {snip}…")
228
+ retrieved_md = "\n".join(ret_lines)
229
+ cands_md = "\n".join(cands_md_lines)
230
 
231
+ return short, sent, highlighted, cands_md, retrieved_md
 
 
 
 
232
 
233
+ # =========================
234
+ # UI text (Malay / English)
235
+ # =========================
236
+ TXT = {
237
+ "ms": {
238
+ "title": "Salatus QA — Question-Only",
239
+ "subtitle": "Masukkan soalan sahaja. Aplikasi memilih konteks terbaik, kemudian mengekstrak jawapan.",
240
+ "tab_demo": "Demo",
241
+ "q": "Soalan / Question",
242
+ "topk": "Bilangan konteks dicadang (top_k)",
243
+ "maxlen": "Maks panjang jawapan (token)",
244
+ "thresh": "Ambang persamaan (retrieval)",
245
+ "ask": "Tanya",
246
+ "best": "Jawapan terbaik",
247
+ "sent": "Jawapan (ayat)",
248
+ "cands": "Calon jawapan (Top-3)",
249
+ "retr": "Konteks terpilih",
250
+ "placeholder_q": "Tulis soalan…",
251
+ "examples": "Contoh pantas",
252
+ "no_ctx": "Maaf, konteks tidak mencukupi. Cuba soalan lain atau turunkan ambang."
253
+ },
254
+ "en": {
255
+ "title": "Salatus QA — Question-Only",
256
+ "subtitle": "Ask a question only. The app retrieves the best passage and extracts the answer.",
257
+ "tab_demo": "Demo",
258
+ "q": "Question",
259
+ "topk": "Number of candidate passages (top_k)",
260
+ "maxlen": "Max answer length (tokens)",
261
+ "thresh": "Similarity threshold (retrieval)",
262
+ "ask": "Ask",
263
+ "best": "Best answer",
264
+ "sent": "Answer (sentence)",
265
+ "cands": "Top-3 candidates",
266
+ "retr": "Selected passages",
267
+ "placeholder_q": "Type your question…",
268
+ "examples": "Quick examples",
269
+ "no_ctx": "No confident passage found. Try another question or lower the threshold."
270
+ }
271
+ }
272
+
273
+ # =========================
274
+ # Gradio App
275
+ # =========================
276
+ def run(q, top_k, max_len, thresh, lang):
277
+ if not q:
278
+ return "—", "", "", "—", "—"
279
+ cands = retrieve(q, int(top_k), float(thresh))
280
  if not cands:
281
+ msg = TXT[lang]["no_ctx"]
282
+ return msg, "", "", "—", "—"
283
+ return answer_with_cleaning(q, cands, int(max_len))
284
+
285
+ with gr.Blocks(title="Salatus QA — Question-Only", css=CSS, theme=gr.themes.Soft()) as demo:
286
+ # Header
287
+ gr.HTML(
288
+ f"""
 
 
 
 
 
 
 
289
  <div id="hero">
290
+ <div style="font-size:1.8rem">🏰</div>
291
  <div>
292
+ <h1>{TXT['ms']['title']}</h1>
293
+ <p>{TXT['ms']['subtitle']}</p>
294
+ </div>
295
+ <div style="margin-left:auto; background: rgba(255,255,255,.15); border: 1px solid rgba(255,255,255,.25); padding: 4px 10px; border-radius: 999px;">
296
+ RAG Demo
297
  </div>
 
298
  </div>
299
+ """
300
+ )
301
+
302
+ with gr.Tabs():
303
+ with gr.Tab(TXT["ms"]["tab_demo"]):
304
+ with gr.Row():
305
+ lang = gr.Radio(choices=[("Bahasa Melayu","ms"), ("English","en")],
306
+ value=("Bahasa Melayu","ms"),
307
+ label="Bahasa / Language", interactive=True)
308
+
309
+ with gr.Row():
310
+ q = gr.Textbox(label=TXT["ms"]["q"], lines=2, placeholder=TXT["ms"]["placeholder_q"])
311
+
312
+ with gr.Row():
313
+ top_k = gr.Slider(1, 10, value=5, step=1, label=TXT["ms"]["topk"])
314
+ max_len = gr.Slider(6, 24, value=12, step=1, label=TXT["ms"]["maxlen"])
315
+ thresh = gr.Slider(0.0, 0.4, value=0.12, step=0.01, label=TXT["ms"]["thresh"])
316
+
317
+ ask = gr.Button(TXT["ms"]["ask"], size="lg")
318
+
319
+ with gr.Row():
320
+ ans = gr.Textbox(label=TXT["ms"]["best"], interactive=False)
321
+ with gr.Row():
322
+ sent = gr.Textbox(label=TXT["ms"]["sent"], interactive=False)
323
+
324
+ html = gr.HTML(label="Sorotan / Highlight", elem_classes=["card"])
325
+ tops = gr.Textbox(label=TXT["ms"]["cands"], interactive=False)
326
+ rets = gr.Textbox(label=TXT["ms"]["retr"], interactive=False)
327
+
328
+ # Quick examples (click to auto-fill then click Ask)
329
+ examples = [
330
+ ["Siapakah yang dikahwinkan dengan Tuan Puteri Mahtabu’l-Bahri?", 5, 12, 0.12],
331
+ ["Siapakah yang memerintahkan pihak Sultan Alau'd-Din?", 5, 12, 0.12],
332
+ ]
333
+ eg = gr.Examples(
334
+ examples=[[e[0], e[1], e[2], e[3]] for e in examples],
335
+ inputs=[q, top_k, max_len, thresh],
336
+ label=TXT["ms"]["examples"],
337
+ cache_examples=False
338
+ )
339
+
340
+ def ask(q, top_k_, max_len_, thresh_, lang_tuple):
341
+ # lang comes in as (label,value)
342
+ lang_code = lang_tuple[1] if isinstance(lang_tuple, (tuple, list)) else lang_tuple
343
+ return run(q_, top_k_, max_len_, thresh_, lang_code)
344
+
345
+ ask.click(_ask, inputs=[q, top_k, max_len, thresh, lang], outputs=[ans, sent, html, tops, rets])
346
+
347
+ # Live update labels when language toggles
348
+ def switch_labels(lang_tuple):
349
+ code = lang_tuple[1] if isinstance(lang_tuple, (tuple, list)) else lang_tuple
350
+ T = TXT[code]
351
+ return (
352
+ gr.update(label=T["q"], placeholder=T["placeholder_q"]),
353
+ gr.update(label=T["topk"]),
354
+ gr.update(label=T["maxlen"]),
355
+ gr.update(label=T["thresh"]),
356
+ gr.update(value=T["ask"]),
357
+ gr.update(label=T["best"]),
358
+ gr.update(label=T["sent"]),
359
+ gr.update(label=T["cands"]),
360
+ gr.update(label=T["retr"]),
361
+ )
362
+
363
+ lang.change(
364
+ switch_labels, inputs=[lang],
365
+ outputs=[q, top_k, max_len, thresh, ask, ans, sent, tops, rets]
366
+ )
367
 
368
  demo.queue().launch()