Update app.py

app.py

@@ -23,12 +23,7 @@ import docx2txt
 # Models
 # =========================================================
 EMBED_MODEL_NAME = os.getenv("EMBED_MODEL_NAME", "BAAI/bge-base-en-v1.5")
-
-# If CPU Space is slow, set Space Variable:
-#   RERANK_MODEL_NAME=BAAI/bge-reranker-base
 RERANK_MODEL_NAME = os.getenv("RERANK_MODEL_NAME", "BAAI/bge-reranker-large")
-
-# Default LLM judge (override via Space Variables)
 LLM_MODEL = os.getenv("LLM_MODEL", "meta-llama/Meta-Llama-3.1-8B-Instruct")
 
 
@@ -41,7 +36,6 @@ CHUNK_OVERLAP_CHARS = 180
 TOP_CHUNKS_PER_CV = 10
 EVIDENCE_CHUNKS_PER_CV = 4
 
-# Smaller batch reduces truncation and "only 1 candidate" outputs
 LLM_BATCH_SIZE = int(os.getenv("LLM_BATCH_SIZE", "4"))
 LLM_MAX_TOKENS = int(os.getenv("LLM_MAX_TOKENS", "3500"))
 LLM_TEMPERATURE = float(os.getenv("LLM_TEMPERATURE", "0.15"))
@@ -49,6 +43,8 @@ LLM_TEMPERATURE = float(os.getenv("LLM_TEMPERATURE", "0.15"))
 MAX_CV_CHARS = 120_000
 MAX_JD_CHARS = 60_000
 
+MAX_CV_UPLOADS = 20  # ✅ requested max
+
 # Global singletons
 _embedder: Optional[SentenceTransformer] = None
 _reranker: Optional[CrossEncoder] = None
@@ -258,7 +254,6 @@ def compute_local_score(retr_sims: List[float], rerank_logits: List[float]) -> f
 # LLM Prompt (compact to avoid truncation)
 # =========================================================
 def build_llm_prompt(jd_text: str, must_haves: str, candidates: List[Dict[str, Any]]) -> str:
-    # IMPORTANT: no example filename like "example.pdf"
     schema_example = {
         "ranked": [
             {
@@ -374,11 +369,6 @@ def fallback_candidate(filename: str, score: float) -> CandidateLLMResult:
 
 
 def llm_judge_rank_batch(jd_text: str, must_haves: str, batch: List[Dict[str, Any]]) -> LLMRankingOutput:
-    """
-    Guarantees: returns a result for EVERY candidate in `batch`.
-    If LLM returns incomplete list, re-judge missing CVs individually.
-    If still missing, fallback to local_score.
-    """
     client = get_hf_client()
 
     prompt = build_llm_prompt(
@@ -401,7 +391,6 @@ def llm_judge_rank_batch(jd_text: str, must_haves: str, batch: List[Dict[str, An
 
     out: Optional[LLMRankingOutput] = None
 
-    # Attempt 1
     text = _call(LLM_TEMPERATURE, LLM_MAX_TOKENS, prompt)
     try:
         out = LLMRankingOutput.model_validate(json.loads(text))
@@ -410,7 +399,6 @@ def llm_judge_rank_batch(jd_text: str, must_haves: str, batch: List[Dict[str, An
         if obj:
             out = LLMRankingOutput.model_validate(json.loads(obj))
 
-    # Retry once if parsing failed
     if out is None:
         text2 = _call(0.0, max(LLM_MAX_TOKENS, 4500), prompt)
         try:
@@ -420,7 +408,6 @@ def llm_judge_rank_batch(jd_text: str, must_haves: str, batch: List[Dict[str, An
             if obj2:
                 out = LLMRankingOutput.model_validate(json.loads(obj2))
 
-    # If still failing: fallback all
     if out is None:
         ranked = [fallback_candidate(b["filename"], b.get("local_score", 50.0)) for b in batch]
         return LLMRankingOutput(ranked=ranked, overall_notes="LLM parsing failed; used local scoring fallback.")
@@ -428,7 +415,6 @@ def llm_judge_rank_batch(jd_text: str, must_haves: str, batch: List[Dict[str, An
     returned = {c.filename: c for c in out.ranked}
     missing = [b for b in batch if b["filename"] not in returned]
 
-    # Re-judge missing individually (more reliable)
     for b in missing:
         single_prompt = build_llm_prompt(
             jd_text,
@@ -590,22 +576,19 @@ def render_top10_html(ranked: List[CandidateLLMResult], total_count: int) -> str
 
 
 # =========================================================
-# Shortlist export (FIXED for pandas DataFrame)
+# Shortlist export (DataFrame-safe)
 # =========================================================
 def export_shortlist(shortlist_table: pd.DataFrame) -> Tuple[str, str, str]:
     if shortlist_table is None or shortlist_table.empty:
         raise gr.Error("No shortlist data yet. Run ranking first.")
 
-    # First column is Shortlisted (bool)
     shortlisted_df = shortlist_table[shortlist_table.iloc[:, 0] == True]
     if shortlisted_df.empty:
         raise gr.Error("No candidates marked as shortlisted.")
 
-    # Export CSV
     tmp = tempfile.NamedTemporaryFile(delete=False, suffix=".csv")
     shortlisted_df.to_csv(tmp.name, index=False)
 
-    # Email list from Email column (index 6 in our table)
     emails = (
         shortlisted_df.iloc[:, 6]
         .dropna()
@@ -622,7 +605,7 @@ def export_shortlist(shortlist_table: pd.DataFrame) -> Tuple[str, str, str]:
 
 
 # =========================================================
-# Main pipeline
+# Main pipeline (with progress bar)
 # =========================================================
 def rank_app(
     jd_file_obj,
@@ -630,13 +613,15 @@ def rank_app(
     must_haves: str,
     mask_pii_toggle: bool,
     show_contacts_toggle: bool,
+    progress=gr.Progress(track_tqdm=False),  # ✅ progress bar
 ):
     t0 = time.time()
     ensure_models()
     embedder = _embedder
     reranker = _reranker
 
-    # ---- Load JD
+    progress(0.02, desc="Loading Job Description...")
+
     jd_path = gr_file_to_path(jd_file_obj)
     if not jd_path:
         raise gr.Error("Please upload a Job Description file (PDF/DOCX/TXT).")
@@ -645,10 +630,13 @@ def rank_app(
     if not jd_text:
         raise gr.Error("Could not extract text from the Job Description file.")
 
-    # ---- CV paths
     if not cv_file_objs:
         raise gr.Error("Please upload at least 1 CV.")
 
+    # ✅ enforce max 20
+    if len(cv_file_objs) > MAX_CV_UPLOADS:
+        raise gr.Error(f"Maximum allowed CV uploads is {MAX_CV_UPLOADS}. You uploaded {len(cv_file_objs)}.")
+
     cv_paths = []
     for f in cv_file_objs:
         p = gr_file_to_path(f)
@@ -657,7 +645,8 @@ def rank_app(
     if not cv_paths:
         raise gr.Error("Could not read uploaded CV files (no valid paths).")
 
-    # ---- Duplicate detection
+    progress(0.06, desc="Checking duplicates...")
+
     seen = {}
     duplicates = []
     unique_paths = []
@@ -674,14 +663,19 @@ def rank_app(
         seen[h] = fname
         unique_paths.append(p)
 
-    # ---- Embed JD once
+    progress(0.10, desc="Embedding Job Description...")
+
     jd_vec = np.array(embedder.encode([jd_text], normalize_embeddings=True), dtype=np.float32)
 
-    # ---- Process ALL CVs (retrieval + rerank + local_score + contacts)
     local_pool = []
     contacts_map: Dict[str, Dict[str, str]] = {}
 
-    for p in unique_paths:
+    total = len(unique_paths)
+    for idx, p in enumerate(unique_paths, start=1):
+        # progress 10% -> 70% while processing CVs
+        prog = 0.10 + 0.60 * (idx / max(1, total))
+        progress(prog, desc=f"Processing CVs ({idx}/{total}) — {os.path.basename(p)}")
+
         raw = clean_text(read_file_to_text(p))[:MAX_CV_CHARS]
         if not raw:
             continue
@@ -721,42 +715,50 @@ def rank_app(
     if not local_pool:
         raise gr.Error("Could not extract usable text from the uploaded CVs.")
 
-    # ---- Sort by local_score, then LLM judge ALL (batched)
+    progress(0.72, desc="Preparing LLM ranking...")
+
     local_pool = sorted(local_pool, key=lambda x: float(x["local_score"]), reverse=True)
 
     batch_outputs: List[LLMRankingOutput] = []
-    for i in range(0, len(local_pool), LLM_BATCH_SIZE):
-        batch = local_pool[i:i + LLM_BATCH_SIZE]
+    batches = max(1, (len(local_pool) + LLM_BATCH_SIZE - 1) // LLM_BATCH_SIZE)
+
+    for b in range(batches):
+        start = b * LLM_BATCH_SIZE
+        end = start + LLM_BATCH_SIZE
+        batch = local_pool[start:end]
+
+        # progress 72% -> 92% while LLM runs
+        prog = 0.72 + 0.20 * ((b + 1) / batches)
+        progress(prog, desc=f"LLM judging batches ({b+1}/{batches})...")
 
         llm_batch = [
             {
                 "filename": c["filename"],
                 "evidence_chunks": c["evidence_chunks"],
-                "local_score": c["local_score"],  # fallback only
+                "local_score": c["local_score"],
             }
             for c in batch
         ]
-
         out = llm_judge_rank_batch(jd_text, must_haves or "", llm_batch)
         batch_outputs.append(out)
 
+    progress(0.94, desc="Finalizing report...")
+
     judged = merge_llm_batches(batch_outputs)
     ranked = judged.ranked
     if not ranked:
         raise gr.Error("LLM returned an empty ranking.")
 
-    # ---- Top 10 report
     report_html = render_top10_html(ranked, total_count=len(ranked))
 
-    # ---- Full ranking export (with contacts)
     tmp = tempfile.NamedTemporaryFile(delete=False, suffix=".csv")
     with open(tmp.name, "w", newline="", encoding="utf-8") as f:
         w = csv.writer(f)
         w.writerow(["Rank", "Filename", "FinalScore(0-100)", "FitLevel", "Name", "Email", "Phone", "Summary"])
-        for idx, c in enumerate(ranked, start=1):
+        for ridx, c in enumerate(ranked, start=1):
             ci = contacts_map.get(c.filename, {"name": "", "email": "", "phone": ""})
             w.writerow([
-                idx,
+                ridx,
                 c.filename,
                 round(float(c.final_score), 2),
                 c.fit_level,
@@ -766,13 +768,12 @@ def rank_app(
                 c.summary,
             ])
 
-    # ---- Shortlist table (pandas DataFrame so export works)
     shortlist_rows = []
-    for idx, c in enumerate(ranked, start=1):
+    for ridx, c in enumerate(ranked, start=1):
         ci = contacts_map.get(c.filename, {"name": "", "email": "", "phone": ""})
         shortlist_rows.append([
             False,
-            idx,
+            ridx,
             c.filename,
             round(float(c.final_score), 2),
             c.fit_level,
@@ -788,28 +789,27 @@ def rank_app(
 
     elapsed = time.time() - t0
     meta = (
-        f"**CVs uploaded:** {len(cv_paths)} → **Unique processed:** {len(unique_paths)}  \n"
+        f"**CVs uploaded:** {len(cv_paths)} → **Unique processed:** {len(unique_paths)} (Max allowed: {MAX_CV_UPLOADS})  \n"
         f"**Ranked (ALL):** {len(ranked)}  \n"
-        f"**LLM batches:** {len(batch_outputs)} (batch size={LLM_BATCH_SIZE})  \n"
+        f"**LLM batches:** {batches} (batch size={LLM_BATCH_SIZE})  \n"
         f"**Time:** {elapsed:.2f}s  \n"
         f"**Duplicates skipped:** {len(duplicates)}  \n\n"
         f"**LLM Notes:** {(judged.overall_notes or '').strip()}"
     )
 
+    progress(1.0, desc="Done ✅")
     return report_html, meta, tmp.name, shortlist_df, "", ""
 
 
 # =========================================================
-# SGS Theme / CSS (with WHITE FONT FIX)
+# SGS Theme / CSS (white text + MET green + nice touches)
 # =========================================================
 CUSTOM_CSS = """
 :root{
   --sgs-blue:#0B3D91;
   --sgs-green:#00A651;
   --text:#F3F7FF;
-  --muted:#D7E3FF;
   --line:rgba(255,255,255,.14);
-  --soft:rgba(255,255,255,.09);
 }
 
 .gradio-container{max-width:1180px !important;}
@@ -819,11 +819,9 @@ body, .gradio-container{
               linear-gradient(180deg, #060914, #060914) !important;
 }
 
-/* Improve default component text contrast */
-.gradio-container, .gradio-container *{
-  color: var(--text);
-}
+.gradio-container, .gradio-container *{ color: var(--text); }
 
+/* Hero */
 .hero{
   border:1px solid var(--line);
   background: linear-gradient(135deg, rgba(11,61,145,.40), rgba(0,166,81,.20));
@@ -835,10 +833,20 @@ body, .gradio-container{
   gap:16px;
   box-shadow: 0 18px 40px rgba(0,0,0,.38);
   margin: 12px 0 16px;
+  position: relative;
+  overflow: hidden;
 }
-.hero-title{color:var(--text);font-weight:900;font-size:22px;}
-.hero-sub{color:rgba(243,247,255,.88);margin-top:6px;font-size:13px;}
-.hero-right{display:flex;gap:10px;flex-wrap:wrap;justify-content:flex-end;}
+.hero:before{
+  content:"";
+  position:absolute;
+  inset:-40%;
+  background: radial-gradient(circle at 30% 30%, rgba(255,255,255,.10), transparent 45%);
+  transform: rotate(18deg);
+  pointer-events:none;
+}
+.hero-title{font-weight:900;font-size:22px;position:relative;}
+.hero-sub{color:rgba(243,247,255,.90);margin-top:6px;font-size:13px;position:relative;}
+.hero-right{display:flex;gap:10px;flex-wrap:wrap;justify-content:flex-end;position:relative;}
 
 .kpi{
   background: rgba(255,255,255,.08);
@@ -846,17 +854,25 @@ body, .gradio-container{
   border-radius: 16px;
   padding: 10px 12px;
   min-width: 140px;
+  backdrop-filter: blur(6px);
 }
-.kpi-label{color:rgba(243,247,255,.80);font-size:12px;font-weight:700;}
-.kpi-val{color:var(--text);font-size:18px;font-weight:900;margin-top:2px;}
+.kpi-label{color:rgba(243,247,255,.82);font-size:12px;font-weight:700;}
+.kpi-val{font-size:18px;font-weight:900;margin-top:2px;}
 
+/* Cards */
 .cards{display:grid;grid-template-columns: 1fr; gap: 12px;}
 .card{
   background: linear-gradient(180deg, rgba(16,26,44,.98), rgba(12,19,34,.88));
-  border:1px solid var(--line);
+  border:1px solid rgba(255,255,255,.14);
   border-radius: 18px;
   padding: 14px;
   box-shadow: 0 14px 28px rgba(0,0,0,.28);
+  transition: transform .18s ease, box-shadow .18s ease, border-color .18s ease;
+}
+.card:hover{
+  transform: translateY(-2px);
+  box-shadow: 0 20px 40px rgba(0,0,0,.38);
+  border-color: rgba(255,255,255,.20);
 }
 .card-top{display:flex;align-items:flex-start;justify-content:space-between;gap:10px;}
 .card-title{display:flex;gap:10px;align-items:baseline;flex-wrap:wrap;}
@@ -871,6 +887,7 @@ body, .gradio-container{
 .file{font-weight:900;font-size:16px;}
 .card-meta{display:flex;gap:8px;align-items:center;flex-wrap:wrap;justify-content:flex-end;}
 
+/* Badges */
 .badge{
   display:inline-flex;align-items:center;
   padding: 6px 10px;border-radius: 999px;font-size:12px;font-weight:900;
@@ -892,6 +909,7 @@ body, .gradio-container{
 .p-low{  background: rgba(245,158,11,.16); border-color: rgba(245,158,11,.28); }
 .p-bad{  background: rgba(239,68,68,.14); border-color: rgba(239,68,68,.28); }
 
+/* Score bar */
 .bar{
   width: 100%; height: 10px; border-radius: 999px;
   background: rgba(255,255,255,.10); overflow: hidden;
@@ -903,79 +921,60 @@ body, .gradio-container{
   background: linear-gradient(90deg, var(--sgs-green), #4fb2ff, var(--sgs-blue));
 }
 
-.summary{
-  color:rgba(243,247,255,.92);
-  font-size:13px;
-  line-height:1.55rem;
-  margin: 6px 0 10px;
-}
-.section-title{
-  color:rgba(224,234,255,.98);
-  font-size:13px;
-  font-weight:900;
-  margin:10px 0 6px;
-}
+.summary{font-size:13px;line-height:1.55rem;margin: 6px 0 10px;color:#fff;}
+.section-title{font-size:13px;font-weight:900;margin:10px 0 6px;color:#fff;}
 
 .grid{display:grid;grid-template-columns: 1fr 1fr; gap: 14px;}
 @media(max-width:860px){.grid{grid-template-columns:1fr;}}
 
-.list{margin:0;padding-left:18px;color:rgba(243,247,255,.92);}
-.list li{margin:6px 0;line-height:1.30rem;color:rgba(243,247,255,.92);}
+.list{margin:0;padding-left:18px;color:#fff;}
+.list li{margin:6px 0;line-height:1.30rem;color:#fff;}
 
+/* Quotes / Evidence */
 .quotes{display:grid;gap:10px;margin-top:6px;}
 .quote{
-  background: rgba(255,255,255,.09);
-  border:1px solid rgba(255,255,255,.14);
+  background: rgba(255,255,255,.10);
+  border:1px solid rgba(255,255,255,.16);
   border-radius: 14px;
   padding: 10px 12px;
-  color: rgba(243,247,255,.94);
+  color: #fff;
   font-size: 13px;
   line-height: 1.45rem;
 }
 
+/* Checklist */
 .checklist{display:grid;gap:8px;margin-top:6px;}
 .checkrow{
   display:grid; grid-template-columns: 1.1fr .4fr 1.5fr; gap:10px;
   padding:10px 12px; border-radius:14px;
-  border:1px solid rgba(255,255,255,.14);
-  background: rgba(255,255,255,.07);
+  border:1px solid rgba(255,255,255,.18);
+  background: rgba(255,255,255,.10);
   font-size:13px;
+  position: relative;
+  overflow: hidden;
+}
+.checkrow:before{
+  content:"";
+  position:absolute;
+  left:0; top:0; bottom:0;
+  width:4px;
+  background: rgba(255,255,255,.20);
 }
-.checkrow .req{font-weight:900;}
-.checkrow .st{font-weight:900;text-align:center;}
-.checkrow .ev{opacity:.95;}
+.checkrow .req{font-weight:900;color:#fff;}
+.checkrow .ev{color:rgba(255,255,255,0.95);}
+.checkrow .st{font-weight:1000;text-align:center;letter-spacing:.4px;}
 
-.checkrow.ok{border-color: rgba(0,166,81,.30); background: rgba(0,166,81,.12);}
-.checkrow.partial{border-color: rgba(245,158,11,.30); background: rgba(245,158,11,.12);}
-.checkrow.miss{border-color: rgba(239,68,68,.30); background: rgba(239,68,68,.11);}
+/* ✅ Status colors (MET green) */
+.checkrow.ok:before{ background: rgba(0,166,81,.95); }
+.checkrow.partial:before{ background: rgba(245,158,11,.95); }
+.checkrow.miss:before{ background: rgba(239,68,68,.95); }
 
-/* Dataframe border */
-table { border-color: rgba(255,255,255,.12) !important; }
+.checkrow.ok .st{ color:#22ffb6 !important; text-shadow: 0 0 10px rgba(34,255,182,.18); }
+.checkrow.partial .st{ color:#ffd27a !important; }
+.checkrow.miss .st{ color:#ff9a9a !important; }
 
-/* =========================================================
-   VISIBILITY FIX: Force WHITE text for checklist/evidence
-   ========================================================= */
-.checkrow, .checkrow *{
-  color: #FFFFFF !important;
-}
-.checkrow .ev{
-  color: rgba(255,255,255,0.95) !important;
-}
-.checkrow .st{
-  color: #FFFFFF !important;
-  opacity: 1 !important;
-}
-.quote, .quote *{
-  color: #FFFFFF !important;
-}
-.summary, .section-title, .list, .list li{
-  color: #FFFFFF !important;
-  opacity: 1 !important;
-}
-.checkrow{
-  background: rgba(255,255,255,0.10) !important;
-  border-color: rgba(255,255,255,0.18) !important;
-}
+/* Dataframe border */
+table { border-color: rgba(255,255,255,.14) !important; }
 """
 
 
@@ -991,15 +990,16 @@ theme = gr.themes.Soft(
 )
 
 with gr.Blocks(title="SGS ATS Candidate Matcher", theme=theme, css=CUSTOM_CSS) as demo:
-    gr.Markdown("""
+    gr.Markdown(f"""
 # SGS ATS Candidate Matcher
 Evidence-based CV ranking against a Job Description (Top 10 Report + Shortlisting).  
+**Max CV uploads:** {MAX_CV_UPLOADS}  
 **Important:** set `HF_TOKEN` in Space secrets.
 """)
 
     with gr.Row():
         jd_file = gr.File(label="Job Description file (PDF/DOCX/TXT)", file_types=[".pdf", ".docx", ".txt"])
-        cv_files = gr.File(label="Upload CVs (multiple)", file_count="multiple", file_types=[".pdf", ".docx", ".txt"])
+        cv_files = gr.File(label=f"Upload CVs (max {MAX_CV_UPLOADS})", file_count="multiple", file_types=[".pdf", ".docx", ".txt"])
 
     with gr.Accordion("Settings", open=False):
         must_haves = gr.Textbox(