ChogwuCY commited on
Commit
345dddc
·
verified ·
1 Parent(s): 1a0d25f

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +236 -251
app.py CHANGED
@@ -1,288 +1,273 @@
1
- # app.py — Salatus QA (Question-only RAG), robust version for HF Spaces
2
- import os, json, re, warnings, traceback
3
  from pathlib import Path
 
4
  import numpy as np
 
 
 
 
5
  import gradio as gr
6
- from transformers import pipeline
7
-
8
- warnings.filterwarnings("ignore", category=UserWarning)
9
- warnings.filterwarnings("ignore", category=FutureWarning)
10
-
11
- # --------------------------- Paths & config ---------------------------
12
- try:
13
- HERE = Path(_file_).parent
14
- except NameError:
15
- HERE = Path(".").resolve()
16
-
17
- RDIR = HERE / "retriever"
18
- CTX_PATH = RDIR / "contexts.json"
19
- TFIDF_PATH = RDIR / "tfidf_pipe.joblib"
20
- X_NPZ_PATH = RDIR / "X.npz"
21
-
22
- MODEL_ID = os.getenv("MODEL_ID", "ChogwuCY/salatus-qa-xlmrb-base")
23
-
24
- DEFAULT_TOPK = 5
25
- DEFAULT_MAXLEN = 12
26
- DEFAULT_THRESH = 0.05
27
-
28
- # --------------------------- Load retriever ---------------------------
29
- PIPE = None
30
- X = None
31
- CONTEXTS = []
32
-
33
- def _load_retriever():
34
- """Loads TF-IDF pipeline + matrix and contexts."""
35
- global PIPE, X, CONTEXTS
36
- if not TFIDF_PATH.exists() or not CTX_PATH.exists():
37
- raise RuntimeError(
38
- "Missing retriever assets in ./retriever/. "
39
- "Upload tfidf_pipe.joblib and contexts.json (and X.npz if available)."
40
- )
41
- import joblib
42
- CONTEXTS = json.loads(CTX_PATH.read_text(encoding="utf-8"))
43
- PIPE = joblib.load(TFIDF_PATH)
44
- if X_NPZ_PATH.exists():
45
- from scipy.sparse import load_npz
46
- X = load_npz(str(X_NPZ_PATH))
47
- else:
48
- X = PIPE.transform(CONTEXTS)
49
-
50
- _load_retriever()
51
-
52
- # --------------------------- QA pipeline ------------------------------
53
- # Keep on CPU in Spaces
54
- qa = pipeline("question-answering", model=MODEL_ID, tokenizer=MODEL_ID)
55
-
56
- # --------------------------- Helpers ----------------------------------
57
- APOS = {"’": "'", "‘": "'", "“": '"', "”": '"'}
58
- def norm(s: str) -> str:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
59
  if not s: return ""
60
- s = s.replace("\u200b","").replace("\u200c","")
61
  for k,v in APOS.items(): s = s.replace(k,v)
62
- return re.sub(r"\s+", " ", s).strip()
63
-
64
- def cosine_topk(query: str, k: int):
65
- qv = PIPE.transform([query])
66
- sims = (qv @ X.T).A.ravel()
67
- if k <= 0: k = 1
68
- idxs = np.argpartition(-sims, min(k, len(sims)-1))[:k]
69
- return sorted([(i, float(sims[i])) for i in idxs], key=lambda t: t[1], reverse=True)
70
-
71
- def highlight_span(text: str, start: int, end: int, color="#fff59d"):
72
- if start is None or end is None or start < 0 or end <= start or end > len(text):
73
- return text
74
- return text[:start] + f'<mark style="background:{color}">' + text[start:end] + "</mark>" + text[end:]
75
-
76
- def snippet_around(text: str, start: int, end: int, win=240):
77
- if start is None or end is None or start < 0 or end <= start or end > len(text):
78
- return text[:win]
79
- mid = (start + end)//2
80
- a = max(0, mid - win)
81
- b = min(len(text), mid + win)
82
- local_s = start - a
83
- local_e = end - a
84
- return highlight_span(text[a:b], local_s, local_e)
85
-
86
- # Siapa-aware reranker
87
- _TITLES = r"(Raja|Sultan|Bendahara|Tuan|Tun|Puteri|Raden|Megat|Hang|Syah|Shah)"
88
- NAME_RE = re.compile(rf"(?:{_TITLES}\s+[A-Z][\w’'\-]+(?:\s+[A-Z][\w’'\-]+)*)|(?:[A-Z][\w’'\-]+(?:\s+[A-Z][\w’'\-]+)+)")
89
- BAD_WHO = re.compile(r"\b(tahun|bulan|hari|antara|pada|ke|satu|dua|tiga|empat|\d+)\b", re.I)
90
-
91
- def rerank_by_type(cands, question: str):
92
- is_who = "siapa" in question.lower()
93
- rescored = []
94
- for c in cands:
95
- ans = c.get("answer","").strip()
96
- score = float(c.get("score", 0.0))
97
- bonus = 0.0
98
- if is_who:
99
- if NAME_RE.search(ans): bonus += 0.50
100
- if ans.lower() in {"raja suran","suran"}: bonus += 0.25
101
- if BAD_WHO.search(ans): bonus -= 0.40
102
- if len(ans) <= 3 and len(ans.split())==1: bonus -= 0.05
103
- rescored.append((score+bonus, c))
104
- rescored.sort(key=lambda x:x[0], reverse=True)
105
- return [c for _,c in rescored]
106
-
107
- def soft_trim_words(s: str, max_words: int):
108
- w = s.split()
109
- if len(w) <= max_words: return s
110
- return " ".join(w[:max_words])
111
-
112
- # --------------------------- Core inference ---------------------------
113
- __MARRY__PAT = re.compile(
114
- r"(dikahwinkan(?:\s+dengan)?|dikahwini|diperisterikan|beristeri|married(?:\s+to)?)",
115
- re.I,
116
- )
117
-
118
- def _fallback_name(ctx: str, q: str):
119
- """Heuristic: if QA returns nothing, try to pick a plausible person name,
120
- prioritizing names near 'dikahwinkan/ married' keywords."""
121
- m = __MARRY__PAT.search(ctx)
122
  if m:
123
- # look forward ~120 chars for a proper name
124
- seg = ctx[m.end(): m.end() + 120]
125
  n = NAME_RE.search(seg)
126
  if n:
127
- ans = n.group(0).strip()
128
- start = m.end() + n.start()
129
- end = m.end() + n.end()
130
- return {"answer": ans, "start": start, "end": end, "score": 0.0}
131
- # else: any proper name in the whole ctx
132
- n2 = NAME_RE.search(ctx)
133
- if n2:
134
- ans = n2.group(0).strip()
135
- return {"answer": ans, "start": n2.start(), "end": n2.end(), "score": 0.0}
136
- return None
137
-
138
- def run_question_only(q: str, top_k: int, max_len: int, thresh: float, lang: str):
 
 
 
 
 
 
 
 
 
 
 
 
 
139
  q = norm(q)
140
  if not q:
141
- return "", ("Sila masukkan soalan." if lang=="ms" else "Please enter a question."), "", ""
 
 
142
 
143
  try:
144
- # retrieve a bit more, then filter by threshold; always keep at least 1
145
- pairs = cosine_topk(q, k=max(1, top_k*3))
146
- filt = [(i,s) for (i,s) in pairs if s >= thresh]
 
 
 
 
147
  if not filt:
148
- filt = pairs[:max(1, top_k)]
149
- pairs = filt[:max(1, top_k)]
150
 
151
  answers = []
152
- selected = []
153
- for idx, sim in pairs:
154
  ctx = CONTEXTS[idx]
155
- # extremely compatible QA call (old/new transformers)
156
- try:
157
- out = qa(question=q, context=ctx) # single best
158
- except Exception as e:
159
- # last-resort: treat as no answer from model
160
- out = None
161
-
162
  if isinstance(out, dict) and out.get("answer"):
163
- a = out
164
- a["answer"] = soft_trim_words(a["answer"].strip(), int(max_len))
165
- a["ctx_idx"] = idx
166
- a["sim"] = float(sim)
167
- answers.append(a)
 
 
 
168
  else:
169
- # fallback heuristic: extract a name from context
170
- fb = _fallback_name(ctx, q)
171
- if fb:
172
- fb["answer"] = soft_trim_words(fb["answer"], int(max_len))
173
- fb["ctx_idx"] = idx
174
- fb["sim"] = float(sim)
175
- answers.append(fb)
176
 
177
  preview = ctx[:160].replace("\n"," ") + ("…" if len(ctx)>160 else "")
178
- selected.append(f"{idx} (sim={sim:.3f}) {preview}")
179
 
180
  if not answers:
181
- msg = "Tiada jawapan ditemui." if lang=="ms" else "No answer found."
182
- return "", msg, "", "\n".join(selected)
 
183
 
184
- answers = rerank_by_type(answers, q)
 
185
  best = answers[0]
186
- ans_text = best.get("answer","").strip()
187
-
188
- idx = best.get("ctx_idx")
189
- ctx = CONTEXTS[idx]
190
- snip = snippet_around(ctx, best.get("start",-1), best.get("end",-1))
191
 
192
- tops = "\n".join([
193
- f"- {c.get('answer','').strip()} (confidence≈{float(c.get('score',0.0)):.3f})"
194
- for c in answers[:3]
195
- ])
196
-
197
- return ans_text, snip, tops, "\n".join(selected)
198
 
199
  except Exception as e:
200
- print("ERROR in run_question_only:", repr(e))
201
  traceback.print_exc()
202
- msg = "Ralat dalaman semasa memproses." if lang=="ms" else "Internal error during processing."
203
  return "", msg, "", ""
204
- def run_ui_callback(q, top_k, max_len, thresh, lang):
205
- # Ensure proper numeric types from sliders/text
206
- try:
207
- return run_question_only(q, int(top_k), int(max_len), float(thresh), lang)
208
- except Exception as e:
209
- print("ERROR in run_ui_callback:", repr(e))
210
- traceback.print_exc()
211
- return "", "Internal error.", "", ""
212
-
213
- # --------------------------- UI ---------------------------------------
214
- BRAND_PRIMARY = "#0f172a"
215
- BRAND_ACCENT = "#f59e0b"
216
- BRAND_BG = "#f8fafc"
217
- HILITE_BG = "#fff59d"
218
-
219
- CSS = f"""
220
- :root {{
221
- --brand-primary: {BRAND_PRIMARY};
222
- --brand-accent: {BRAND_ACCENT};
223
- --brand-bg: {BRAND_BG};
224
- --hilite-bg: {HILITE_BG};
225
- }}
226
- .gradio-container {{ background: var(--brand-bg); }}
227
- #hero {{
228
- background: linear-gradient(135deg, #1e293b, #0f172a 70%);
229
- color: white; border-radius: 14px; padding: 16px 20px; display:flex; align-items:center; gap:12px;
230
- }}
231
- #hero .title {{ font-weight:800; font-size:1.15rem; }}
232
- button, .gr-button {{
233
- background: var(--brand-accent) !important; color:#111 !important; border:0 !important; border-radius:12px !important; font-weight:700 !important;
234
- }}
235
- mark {{ background: var(--hilite-bg); }}
236
- """
237
 
 
 
 
238
  def build_ui():
239
- with gr.Blocks(title="Salatus QA — Question-Only",
240
- theme=gr.themes.Soft(primary_hue="slate", secondary_hue="amber"),
241
- css=CSS) as demo:
242
- gr.HTML("""<div id="hero">
243
  <div style="font-size:1.6rem">🏰</div>
244
- <div><div class="title">Salatus QA — Question-Only</div>
245
- <div style="opacity:.85">Masukkan soalan sahaja. Aplikasi memilih konteks terbaik, kemudian mengekstrak jawapan.</div></div>
246
- <div style="margin-left:auto;opacity:.75">RAG Demo</div></div>""")
247
-
248
- lang = gr.Radio(["ms","en"], value="ms", label="Bahasa / Language (ms/en)")
249
- q = gr.Textbox(lines=2, label="Soalan / Question",
250
- value="Siapakah yang dikahwinkan dengan Tuan Puteri Mahtabu’l-Bahri?")
251
-
252
- with gr.Row():
253
- top_k = gr.Slider(1, 12, value=DEFAULT_TOPK, step=1,
254
- label="Bilangan konteks dicadang (top_k) / Number of candidate passages")
255
- maxlen = gr.Slider(6, 24, value=DEFAULT_MAXLEN, step=1,
256
- label="Maks panjang jawapan (token) / Max answer length (tokens)")
257
- thresh = gr.Slider(0.00, 0.50, value=DEFAULT_THRESH, step=0.01,
258
- label="Ambang persamaan (retrieval) / Similarity threshold")
259
-
260
- ask_btn = gr.Button("Tanya / Ask", size="lg")
261
-
262
- best = gr.Textbox(label="Jawapan terbaik / Best answer", interactive=False)
263
- sent = gr.HTML(label="Jawapan (ayat) / Answer (sentence)") # HTML to keep highlight
264
- tops = gr.Textbox(label="Calon jawapan (Top-3) / Top-3 candidates", interactive=False)
265
- rets = gr.Textbox(label="Konteks terpilih / Selected passages", interactive=False)
266
-
267
- ask_btn.click(run_ui_callback, inputs=[q, top_k, maxlen, thresh, lang],
268
- outputs=[best, sent, tops, rets])
269
-
270
- gr.Markdown("*Contoh pantas / Quick examples*")
271
- examples = [
272
- ["Siapakah yang dikahwinkan dengan Tuan Puteri Mahtabu’l-Bahri?", "ms", 5, 12, 0.12],
273
- ["Siapakah yang memerintahkan pihak Sultan Alau'd-Din?", "ms", 5, 12, 0.12],
274
- ["Who married Tuan Puteri Mahtabu’l-Bahri?", "en", 5, 12, 0.12],
275
- ]
276
- gr.Examples(
277
- examples=examples,
278
- inputs=[q, lang, top_k, maxlen, thresh],
279
- outputs=[best, sent, tops, rets],
280
- fn=lambda q_, lang_, k_, ml_, th_: run_ui_callback(q_, int(k_), int(ml_), float(th_), lang_),
281
- cache_examples=False
282
- )
 
283
  return demo
284
 
285
  demo = build_ui()
286
 
287
- if __name__ == "__main__":
288
  demo.queue().launch()
 
1
+ # app.py — Salatus QA (Question-only, RAG + robust fallbacks)
2
+ import os, re, json, traceback
3
  from pathlib import Path
4
+
5
  import numpy as np
6
+ import joblib
7
+ from scipy.sparse import csr_matrix
8
+ from transformers import pipeline, AutoTokenizer, AutoModelForQuestionAnswering
9
+
10
  import gradio as gr
11
+
12
+ # -------------------------------
13
+ # Config
14
+ # -------------------------------
15
+ HERE = Path(_file_).parent
16
+ REPO = os.environ.get("MODEL_REPO", "ChogwuCY/salatus-qa-xlmrb-base") # <— change if needed
17
+
18
+ RETR_DIR = HERE / "retriever"
19
+ CTX_JSON = RETR_DIR / "contexts.json"
20
+ TFIDF_PIP = RETR_DIR / "tfidf_pipe.joblib"
21
+
22
+ DEFAULT_TOPK = 5
23
+ DEFAULT_MAXLEN = 12
24
+ DEFAULT_THRESH = 0.05 # be permissive so we always return something
25
+
26
+ # Colors
27
+ BRAND_PRIMARY = "#0f172a" # blue-ish slate
28
+ BRAND_ACCENT = "#f59e0b" # amber
29
+ BRAND_BG = "#f8fafc"
30
+ HILITE_BG = "#fff59d"
31
+
32
+ CSS = f"""
33
+ :root {{
34
+ --brand-primary:{BRAND_PRIMARY};
35
+ --brand-accent:{BRAND_ACCENT};
36
+ --brand-bg:{BRAND_BG};
37
+ --hilite-bg:{HILITE_BG};
38
+ }}
39
+ .gradio-container {{
40
+ background: var(--brand-bg);
41
+ }}
42
+ #hero {{
43
+ background: linear-gradient(135deg, #1d2a52, #0f172a 70%);
44
+ color: #fff;
45
+ border-radius: 14px;
46
+ padding: 18px 20px;
47
+ margin-bottom: 10px;
48
+ display:flex; align-items:center; gap:12px;
49
+ }}
50
+ #hero .title {{ font-weight:800; font-size:1.25rem; }}
51
+ #hero .sub {{ opacity:.9; font-size:.95rem; }}
52
+ button, .gr-button {{
53
+ background: var(--brand-accent) !important; color:#111 !important;
54
+ border-radius: 12px !important; font-weight:700 !important; border:0 !important;
55
+ }}
56
+ .markdown-body mark, mark {{
57
+ background: var(--hilite-bg); padding:0 .15em; border-radius:.2em;
58
+ }}
59
+ """
60
+
61
+ # -------------------------------
62
+ # Load retriever
63
+ # -------------------------------
64
+ def load_retriever():
65
+ if not CTX_JSON.exists() or not TFIDF_PIP.exists():
66
+ return None, None
67
+ contexts = json.loads(CTX_JSON.read_text(encoding="utf-8"))
68
+ pipe = joblib.load(TFIDF_PIP) # (TfidfVectorizer -> Normalizer) pipeline
69
+ return contexts, pipe
70
+
71
+ CONTEXTS, TFIDF = load_retriever()
72
+
73
+ # -------------------------------
74
+ # Model / pipeline
75
+ # -------------------------------
76
+ device = 0 if os.environ.get("CUDA_VISIBLE_DEVICES") else -1 # Space will set CPU → -1
77
+ tokenizer = AutoTokenizer.from_pretrained(REPO)
78
+ model = AutoModelForQuestionAnswering.from_pretrained(REPO)
79
+ qa = pipeline("question-answering", model=model, tokenizer=tokenizer, device=device)
80
+
81
+ # -------------------------------
82
+ # Utilities
83
+ # -------------------------------
84
+ APOS = {"’": "'", "‘": "'", "“": '"', "”": '"', "\u200b":"", "\u200c":""}
85
+ def norm(s:str)->str:
86
  if not s: return ""
 
87
  for k,v in APOS.items(): s = s.replace(k,v)
88
+ return re.sub(r"\s+"," ", s).strip()
89
+
90
+ def cosine_topk(query:str, k:int=5):
91
+ """Return list[(idx, sim)] by TF-IDF cosine. Keep k even if threshold used later."""
92
+ if TFIDF is None or CONTEXTS is None or not CONTEXTS:
93
+ return []
94
+ qv = TFIDF.transform([query])
95
+ # Project into the Normalized space (last step in pipeline already Normalizer)
96
+ # Equivalent of dot-product cosine.
97
+ # We can get raw matrix by re-fitting; instead, just score each doc via vectorizer vocabulary overlap.
98
+ # Faster: build a TFIDF matrix once? For simplicity use transform+pairwise with contexts? Not feasible.
99
+ # Minimal hack: select candidates by simple token overlap and then re-score with vectorizer:
100
+
101
+ # Lightweight filter: term overlap
102
+ toks = set(norm(query).lower().split())
103
+ prelim = []
104
+ # sample a subset cheaply for speed; or score all but not feasible here
105
+ # We'll just scan a windowed subset using simple overlap rank and take top 200 to vectorize properly.
106
+ # But we don't have vectors for all docs; we only can transform texts. Transforming all at startup is heavy.
107
+ # So: do a cheap overlap pass, pick top 200 indices, then compute real tfidf cosine on those.
108
+ for i, c in enumerate(CONTEXTS):
109
+ s = c.lower()
110
+ overlap = sum(1 for t in toks if t in s)
111
+ if overlap:
112
+ prelim.append((i, overlap))
113
+ prelim.sort(key=lambda x: x[1], reverse=True)
114
+ cand_idx = [i for i,_ in prelim[:200]] if prelim else list(range(min(200, len(CONTEXTS))))
115
+ cand_texts = [CONTEXTS[i] for i in cand_idx]
116
+ Cv = TFIDF.transform(cand_texts)
117
+ sims = (qv @ Cv.T).A[0] # dense array
118
+ pairs = sorted(zip(cand_idx, sims), key=lambda x: x[1], reverse=True)
119
+ return pairs[:max(1,k)]
120
+
121
+ # Proper-name (Malay / English) for last-resort
122
+ NAME_RE = re.compile(r"\b([A-Z][a-z]+(?:\s+[A-Z][a-z]+){0,3})\b")
123
+ MARRY_RE = re.compile(r"(dikahwinkan(?:\s+dengan)?|dikahwini|diperisterikan|beristeri|married(?:\s+to)?)", re.I)
124
+
125
+ def fallback_person(ctx:str):
126
+ m = MARRY_RE.search(ctx)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
127
  if m:
128
+ seg = ctx[m.end():m.end()+120]
 
129
  n = NAME_RE.search(seg)
130
  if n:
131
+ return n.group(0), m.end()+n.start(), m.end()+n.end()
132
+ n = NAME_RE.search(ctx)
133
+ if n:
134
+ return n.group(0), n.start(), n.end()
135
+ return "", -1, -1
136
+
137
+ def soft_trim_words(s: str, max_tokens:int):
138
+ if not s: return s
139
+ toks = s.split()
140
+ if len(toks) <= max_tokens: return s
141
+ return " ".join(toks[:max_tokens])
142
+
143
+ def snippet_around(ctx:str, s:int, e:int):
144
+ if isinstance(s,int) and isinstance(e,int) and 0 <= s < e <= len(ctx):
145
+ pre = max(0, s-140)
146
+ post = min(len(ctx), e+140)
147
+ return ctx[pre:s] + "<mark>" + ctx[s:e] + "</mark>" + ctx[e:post]
148
+ # fallback: just show a shortened ctx
149
+ cut = ctx[:280].replace("\n"," ")
150
+ return cut + ("…" if len(ctx)>280 else "")
151
+
152
+ # -------------------------------
153
+ # Core runner
154
+ # -------------------------------
155
+ def run(q: str, top_k: int, max_len: int, thresh: float, lang: str):
156
  q = norm(q)
157
  if not q:
158
+ return ("",
159
+ "Sila masukkan soalan." if lang=="ms" else "Please enter a question.",
160
+ "", "")
161
 
162
  try:
163
+ pairs = cosine_topk(q, k=max(1, int(top_k)))
164
+ if not pairs:
165
+ msg = "Aset retriever tiada atau kosong." if lang=="ms" else "Retriever assets missing or empty."
166
+ return "", msg, "", ""
167
+
168
+ # filter by threshold but always keep at least one
169
+ filt = [(i,s) for (i,s) in pairs if s >= float(thresh)]
170
  if not filt:
171
+ filt = pairs[:1]
172
+ sel = filt[:max(1, int(top_k))]
173
 
174
  answers = []
175
+ chosen = []
176
+ for idx, sim in sel:
177
  ctx = CONTEXTS[idx]
178
+ # robust QA call (works across versions)
179
+ out = qa({"question": q, "context": ctx})
180
+ best = None
 
 
 
 
181
  if isinstance(out, dict) and out.get("answer"):
182
+ best = out
183
+ elif isinstance(out, list) and out:
184
+ best = out[0]
185
+
186
+ if best and str(best.get("answer","")).strip():
187
+ ans = soft_trim_words(best["answer"].strip(), int(max_len))
188
+ s = int(best.get("start", -1)); e = int(best.get("end", -1))
189
+ answers.append({"answer": ans, "start": s, "end": e, "score": float(best.get("score",0.0)), "sim": float(sim), "idx": idx})
190
  else:
191
+ # fallback person name
192
+ name, s, e = fallback_person(ctx)
193
+ if name:
194
+ answers.append({"answer": soft_trim_words(name, int(max_len)),
195
+ "start": s, "end": e, "score": 0.0, "sim": float(sim), "idx": idx})
 
 
196
 
197
  preview = ctx[:160].replace("\n"," ") + ("…" if len(ctx)>160 else "")
198
+ chosen.append(f"{idx} (sim={sim:.3f}) {preview}")
199
 
200
  if not answers:
201
+ msg = "Tiada jawapan ditemui. Cuba turunkan ambang persamaan." if lang=="ms" else \
202
+ "No answer found. Try lowering the similarity threshold."
203
+ return "", msg, "", "\n".join(chosen)
204
 
205
+ # pick by (score then sim)
206
+ answers.sort(key=lambda x: (x["score"], x["sim"]), reverse=True)
207
  best = answers[0]
208
+ ctx = CONTEXTS[best["idx"]]
209
+ sent = snippet_around(ctx, best["start"], best["end"])
210
+ tops = "\n".join([f"- {a['answer']} (confidence≈{a['score']:.3f})" for a in answers[:3]])
 
 
211
 
212
+ return best["answer"], sent, tops, "\n".join(chosen)
 
 
 
 
 
213
 
214
  except Exception as e:
215
+ print("ERROR:", repr(e))
216
  traceback.print_exc()
217
+ msg = "Ralat dalaman." if lang=="ms" else "Internal error."
218
  return "", msg, "", ""
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
219
 
220
+ # -------------------------------
221
+ # UI
222
+ # -------------------------------
223
  def build_ui():
224
+ with gr.Blocks(title="Salatus QA — Question-Only", theme=gr.themes.Soft(primary_hue="slate", secondary_hue="amber"), css=CSS) as demo:
225
+ gr.HTML(f"""
226
+ <div id="hero">
 
227
  <div style="font-size:1.6rem">🏰</div>
228
+ <div>
229
+ <div class="title">Salatus QA — Question-Only</div>
230
+ <div class="sub">Masukkan soalan sahaja. Aplikasi memilih konteks terbaik, kemudian mengekstrak jawapan.</div>
231
+ </div>
232
+ <div style="margin-left:auto;opacity:.9">RAG Demo</div>
233
+ </div>
234
+ """)
235
+
236
+ with gr.Tab("Demo"):
237
+ with gr.Row():
238
+ lang = gr.Radio(choices=["ms","en"], value="ms", label="Bahasa / Language (ms/en)")
239
+ q = gr.Textbox(label="Soalan / Question", placeholder="Tulis soalan di sini…")
240
+
241
+ with gr.Row():
242
+ top_k = gr.Slider(1, 12, value=DEFAULT_TOPK, step=1, label="Bilangan konteks dicadang (top_k) / Number of candidate passages")
243
+ maxlen = gr.Slider(6, 32, value=DEFAULT_MAXLEN, step=1, label="Maks panjang jawapan (token) / Max answer length (tokens)")
244
+ thresh = gr.Slider(0.00, 0.50, value=DEFAULT_THRESH, step=0.01, label="Ambang persamaan (retrieval) / Similarity threshold")
245
+
246
+ ask_btn = gr.Button("Tanya / Ask", size="lg")
247
+
248
+ best = gr.Textbox(label="Jawapan terbaik / Best answer")
249
+ sent = gr.HTML(label="Jawapan (ayat) / Answer (sentence)")
250
+ tops = gr.Textbox(label="Calon jawapan (Top-3) / Top-3 candidates")
251
+ rets = gr.Textbox(label="Konteks terpilih / Selected passages")
252
+
253
+ ask_btn.click(fn=run, inputs=[q, top_k, maxlen, thresh, lang], outputs=[best, sent, tops, rets])
254
+
255
+ gr.Examples(
256
+ examples=[
257
+ ["Siapakah yang dikahwinkan dengan Tuan Puteri Mahtabu’l-Bahri?", DEFAULT_TOPK, DEFAULT_MAXLEN, DEFAULT_THRESH, "ms"],
258
+ ["Siapakah yang memerintahkan pihak Sultan Alau'd-Din?", DEFAULT_TOPK, DEFAULT_MAXLEN, DEFAULT_THRESH, "ms"],
259
+ ["Who married Tuan Puteri Mahtabu’l-Bahri?", DEFAULT_TOPK, DEFAULT_MAXLEN, DEFAULT_THRESH, "en"],
260
+ ],
261
+ inputs=[q, top_k, maxlen, thresh, lang],
262
+ fn=run,
263
+ outputs=[best, sent, tops, rets],
264
+ label="Contoh pantas / Quick examples",
265
+ cache_examples=False
266
+ )
267
+
268
  return demo
269
 
270
  demo = build_ui()
271
 
272
+ if __name__name_ == "__main__":
273
  demo.queue().launch()