Update app.py

app.py

@@ -8,7 +8,7 @@
 # - NEW: Lightweight instrumentation (JSONL logs per RAG turn)
 # - UPDATED THEME: Dark-blue tabs + Evaluate tab + k-slider styling
 # - PATCH: Per-question/aggregate File + JSON outputs now dark-themed via elem_id hooks
-# - OPTIONAL JS: Adds .eval-active class when Evaluate tab is selected
+# - NEW: APA-like inline citations via CITATION_MAP + format_citation()
 # ================================================================
 
 # ---------------------- Runtime flags (HF-safe) ----------------------
@@ -50,7 +50,37 @@ except Exception:
     OpenAI = None
 
 # LLM availability flag — used internally; UI remains hidden
-LLM_AVAILABLE = (OPENAI_API_KEY is not None and OPENAI_API_KEY.strip() != "" and OpenAI is not None)
+LLM_AVAILABLE = (
+    OPENAI_API_KEY is not None
+    and OPENAI_API_KEY.strip() != ""
+    and OpenAI is not None
+)
+
+# ========================= APA-style citation helpers =========================
+
+# Map PDF basenames → short APA-style in-text citations.
+# IMPORTANT: edit these entries to match your actual filenames + desired citations.
+CITATION_MAP: Dict[str, str] = {
+    # "filename.pdf": "Author et al., YEAR",
+    # Examples (you can edit/remove these):
+    "S92-Research-on-the-self-sensing-and-mechanical-properties-of_2021_Cement-and-Co.pdf": "Omar et al., 2021",
+    "S60-Study-on-self-sensing-capabilities-of-smart-cements-filled-with-graphene-oxide-under-dynamic-cyclic-loading.pdf": "Zhang et al., 2019",
+    "S59-Modifying self-sensing cement-based composites through multiscale composition.pdf": "Li et al., 2020",
+}
+
+def format_citation(doc_name: str, page: Any) -> str:
+    """
+    Convert a doc filename + page into an APA-like inline citation.
+    Fallback: use filename if doc not in CITATION_MAP.
+    """
+    base = Path(doc_name).name
+    short = CITATION_MAP.get(base, base)
+    if page is None:
+        return short
+    page_str = str(page).strip()
+    if page_str == "" or page_str == "?":
+        return short
+    return f"{short}, p. {page_str}"
 
 # ========================= Predictor (kept) =========================
 CF_COL     = "Conductive Filler Conc. (wt%)"
@@ -140,7 +170,9 @@ def _try_load_model():
                 traceback.print_exc()
                 MODEL = None
     if MODEL is None:
-        MODEL_STATUS = "🔴 Model not found (place stress_gf_xgb.joblib at repo root or models/, or set MODEL_PATH)"
+        MODEL_STATUS = (
+            "🔴 Model not found (place stress_gf_xgb.joblib at repo root or models/, or set MODEL_PATH)"
+        )
         print("[ModelLoad]", MODEL_STATUS)
 
 _try_load_model()  # load at import time
@@ -331,7 +363,8 @@ def chunk_by_sentence_windows(text: str, win_size=8, overlap=2) -> List[str]:
     chunks, step = [], max(1, win_size - overlap)
     for i in range(0, len(sents), step):
         window = sents[i:i+win_size]
-        if not window: break
+        if not window:
+            break
         chunks.append(" ".join(window))
     return chunks
 
@@ -348,10 +381,12 @@ def _safe_init_st_model(name: str):
 
 def build_or_load_hybrid(pdf_dir: Path):
     # Build or load the hybrid retriever cache
-    have_cache = (TFIDF_VECT_PATH.exists() and TFIDF_MAT_PATH.exists()
-                  and RAG_META_PATH.exists()
-                  and (BM25_TOK_PATH.exists() or BM25Okapi is None)
-                  and (EMB_NPY_PATH.exists() or not USE_DENSE))
+    have_cache = (
+        TFIDF_VECT_PATH.exists() and TFIDF_MAT_PATH.exists()
+        and RAG_META_PATH.exists()
+        and (BM25_TOK_PATH.exists() or BM25Okapi is None)
+        and (EMB_NPY_PATH.exists() or not USE_DENSE)
+    )
     if have_cache:
         vectorizer = joblib.load(TFIDF_VECT_PATH)
         X_tfidf    = joblib.load(TFIDF_MAT_PATH)
@@ -378,7 +413,7 @@ def build_or_load_hybrid(pdf_dir: Path):
     meta = pd.DataFrame(rows)
     from sklearn.feature_extraction.text import TfidfVectorizer
     vectorizer = TfidfVectorizer(
-        ngram_range=(1,2),
+        ngram_range=(1, 2),
         min_df=1, max_df=0.95,
         sublinear_tf=True, smooth_idf=True,
         lowercase=True,
@@ -389,10 +424,17 @@ def build_or_load_hybrid(pdf_dir: Path):
     emb = None
     if USE_DENSE:
         try:
-            st_model = _safe_init_st_model(os.getenv("EMB_MODEL_NAME", "sentence-transformers/all-MiniLM-L6-v2"))
+            st_model = _safe_init_st_model(
+                os.getenv("EMB_MODEL_NAME", "sentence-transformers/all-MiniLM-L6-v2")
+            )
             if st_model is not None:
                 from sklearn.preprocessing import normalize as sk_normalize
-                em = st_model.encode(meta["text"].tolist(), batch_size=64, show_progress_bar=False, convert_to_numpy=True)
+                em = st_model.encode(
+                    meta["text"].tolist(),
+                    batch_size=64,
+                    show_progress_bar=False,
+                    convert_to_numpy=True
+                )
                 emb = sk_normalize(em)
                 np.save(EMB_NPY_PATH, emb)
         except Exception as e:
@@ -415,7 +457,10 @@ def _extract_page(text_chunk: str) -> str:
     m = list(re.finditer(r"\[\[PAGE=(\d+)\]\]", text_chunk or ""))
     return (m[-1].group(1) if m else "?")
 
-def hybrid_search(query: str, k=8, w_tfidf=W_TFIDF_DEFAULT, w_bm25=W_BM25_DEFAULT, w_emb=W_EMB_DEFAULT):
+def hybrid_search(query: str, k=8,
+                  w_tfidf=W_TFIDF_DEFAULT,
+                  w_bm25=W_BM25_DEFAULT,
+                  w_emb=W_EMB_DEFAULT):
     if rag_meta is None or rag_meta.empty:
         return pd.DataFrame()
 
@@ -472,7 +517,8 @@ def split_sentences(text: str) -> List[str]:
     sents = sent_split(text)
     return [s for s in sents if 6 <= len(s.split()) <= 60]
 
-def mmr_select_sentences(question: str, hits: pd.DataFrame, top_n=4, pool_per_chunk=6, lambda_div=0.7):
+def mmr_select_sentences(question: str, hits: pd.DataFrame,
+                         top_n=4, pool_per_chunk=6, lambda_div=0.7):
     """
     Robust MMR sentence picker:
     - Handles empty pools
@@ -548,9 +594,17 @@ def mmr_select_sentences(question: str, hits: pd.DataFrame, top_n=4, pool_per_ch
     return selected
 
 def compose_extractive(selected: List[Dict[str, Any]]) -> str:
+    """
+    Build an extractive answer with APA-style inline citations.
+    Each sentence ends with (Author et al., YEAR, p. X) when mapped in CITATION_MAP.
+    """
     if not selected:
         return ""
-    return " ".join(f"{s['sent']} ({s['doc']}, p.{s['page']})" for s in selected)
+    parts = []
+    for s in selected:
+        cite = format_citation(s["doc"], s["page"])
+        parts.append(f"{s['sent']} ({cite})")
+    return " ".join(parts)
 
 # ========================= NEW: Instrumentation helpers =========================
 LOG_PATH = ARTIFACT_DIR / "rag_logs.jsonl"
@@ -570,7 +624,8 @@ def _calc_cost_usd(prompt_toks, completion_toks):
     return (prompt_toks / 1000.0) * OPENAI_IN_COST_PER_1K + (completion_toks / 1000.0) * OPENAI_OUT_COST_PER_1K
 
 # ----------------- Modified to return (text, usage_dict) -----------------
-def synthesize_with_llm(question: str, sentence_lines: List[str], model: str = None, temperature: float = 0.2):
+def synthesize_with_llm(question: str, sentence_lines: List[str],
+                        model: str = None, temperature: float = 0.2):
     if not LLM_AVAILABLE:
         return None, None
     client = OpenAI(api_key=OPENAI_API_KEY)
@@ -579,7 +634,7 @@ def synthesize_with_llm(question: str, sentence_lines: List[str], model: str = N
         "You are a scientific assistant for self-sensing cementitious materials.\n"
         "Answer STRICTLY using the provided sentences.\n"
         "Do not invent facts. Keep it concise (3–6 sentences).\n"
-        "Retain inline citations like (Doc.pdf, p.X) exactly as given."
+        "Retain inline citations exactly as given (e.g., Omar et al., 2021, p. X).\n"
     )
     user_prompt = (
         f"Question: {question}\n\n"
@@ -627,7 +682,10 @@ def rag_reply(
     t0_retr  = time.time()
 
     # --- Retrieval ---
-    hits = hybrid_search(question, k=k, w_tfidf=w_tfidf, w_bm25=w_bm25, w_emb=w_emb)
+    hits = hybrid_search(
+        question, k=k,
+        w_tfidf=w_tfidf, w_bm25=w_bm25, w_emb=w_emb
+    )
     t1_retr = time.time()
     latency_ms_retriever = int((t1_retr - t0_retr) * 1000)
 
@@ -650,10 +708,26 @@ def rag_reply(
         return final
 
     # Select sentences
-    selected = mmr_select_sentences(question, hits, top_n=int(n_sentences), pool_per_chunk=6, lambda_div=0.7)
-    header_cites = "; ".join(f"{Path(r['doc_path']).name} (p.{_extract_page(r['text'])})" for _, r in hits.head(6).iterrows())
-    srcs = {Path(r['doc_path']).name for _, r in hits.iterrows()}
-    coverage_note = "" if len(srcs) >= 3 else f"\n\n> Note: Only {len(srcs)} unique source(s) contributed. Add more PDFs or increase Top-K."
+    selected = mmr_select_sentences(
+        question, hits,
+        top_n=int(n_sentences),
+        pool_per_chunk=6,
+        lambda_div=0.7
+    )
+
+    # APA-style header citations: e.g., "Omar et al., 2021, p. 5; Zhang et al., 2019, p. 12"
+    header_cites = "; ".join(
+        format_citation(
+            Path(r["doc_path"]).name,
+            _extract_page(r["text"])
+        )
+        for _, r in hits.head(6).iterrows()
+    )
+    srcs = {Path(r["doc_path"]).name for _, r in hits.iterrows()}
+    coverage_note = "" if len(srcs) >= 3 else (
+        f"\n\n> Note: Only {len(srcs)} unique source(s) contributed. "
+        f"Add more PDFs or increase Top-K."
+    )
 
     # Prepare retrieval list for logging
     retr_list = []
@@ -670,12 +744,24 @@ def rag_reply(
     # Strict quotes only (no LLM)
     if strict_quotes_only:
         if not selected:
-            final = f"**Quoted Passages:**\n\n---\n" + "\n\n".join(hits['text'].tolist()[:2]) + f"\n\n**Citations:** {header_cites}{coverage_note}"
+            final = (
+                f"**Quoted Passages:**\n\n---\n"
+                + "\n\n".join(hits['text'].tolist()[:2])
+                + f"\n\n**Citations:** {header_cites}{coverage_note}"
+            )
         else:
-            final = "**Quoted Passages:**\n- " + "\n- ".join(f"{s['sent']} ({s['doc']}, p.{s['page']})" for s in selected)
+            final = (
+                "**Quoted Passages:**\n- "
+                + "\n- ".join(
+                    f"{s['sent']} ({format_citation(s['doc'], s['page'])})"
+                    for s in selected
+                )
+            )
             final += f"\n\n**Citations:** {header_cites}{coverage_note}"
             if include_passages:
-                final += "\n\n---\n" + "\n\n".join(hits['text'].tolist()[:2])
+                final += (
+                    "\n\n---\n" + "\n\n".join(hits['text'].tolist()[:2])
+                )
 
         record = {
             "run_id": run_id,
@@ -688,7 +774,10 @@ def rag_reply(
             "retrieval": {"hits": retr_list, "latency_ms_retriever": latency_ms_retriever},
             "output": {
                 "final_answer": final,
-                "used_sentences": [{"sent": s["sent"], "doc": s["doc"], "page": s["page"]} for s in selected]
+                "used_sentences": [
+                    {"sent": s["sent"], "doc": s["doc"], "page": s["page"]}
+                    for s in selected
+                ]
             },
             "latency_ms_total": int((time.time()-t0_total)*1000),
             "openai": None
@@ -701,30 +790,60 @@ def rag_reply(
     llm_usage = None
     llm_latency_ms = None
     if use_llm and selected:
-        lines = [f"{s['sent']} ({s['doc']}, p.{s['page']})" for s in selected]
+        # Pass APA-style citations into the LLM so it preserves them
+        lines = [
+            f"{s['sent']} ({format_citation(s['doc'], s['page'])})"
+            for s in selected
+        ]
         t0_llm = time.time()
-        llm_text, llm_usage = synthesize_with_llm(question, lines, model=model, temperature=temperature)
+        llm_text, llm_usage = synthesize_with_llm(
+            question, lines,
+            model=model, temperature=temperature
+        )
         t1_llm = time.time()
         llm_latency_ms = int((t1_llm - t0_llm) * 1000)
 
         if llm_text:
-            final = f"**Answer (LLM synthesis):** {llm_text}\n\n**Citations:** {header_cites}{coverage_note}"
+            final = (
+                f"**Answer (LLM synthesis):** {llm_text}\n\n"
+                f"**Citations:** {header_cites}{coverage_note}"
+            )
             if include_passages:
-                final += "\n\n---\n" + "\n\n".join(hits['text'].tolist()[:2])
+                final += (
+                    "\n\n---\n" + "\n\n".join(hits['text'].tolist()[:2])
+                )
         else:
             if not extractive:
-                final = f"**Answer:** Here are relevant passages.\n\n**Citations:** {header_cites}{coverage_note}\n\n---\n" + "\n\n".join(hits['text'].tolist()[:2])
+                final = (
+                    f"**Answer:** Here are relevant passages.\n\n"
+                    f"**Citations:** {header_cites}{coverage_note}\n\n---\n"
+                    + "\n\n".join(hits['text'].tolist()[:2])
+                )
             else:
-                final = f"**Answer:** {extractive}\n\n**Citations:** {header_cites}{coverage_note}"
+                final = (
+                    f"**Answer:** {extractive}\n\n"
+                    f"**Citations:** {header_cites}{coverage_note}"
+                )
                 if include_passages:
-                    final += "\n\n---\n" + "\n\n".join(hits['text'].tolist()[:2])
+                    final += (
+                        "\n\n---\n" + "\n\n".join(hits['text'].tolist()[:2])
+                    )
     else:
         if not extractive:
-            final = f"**Answer:** Here are relevant passages.\n\n**Citations:** {header_cites}{coverage_note}\n\n---\n" + "\n\n".join(hits['text'].tolist()[:2])
+            final = (
+                f"**Answer:** Here are relevant passages.\n\n"
+                f"**Citations:** {header_cites}{coverage_note}\n\n---\n"
+                + "\n\n".join(hits['text'].tolist()[:2])
+            )
         else:
-            final = f"**Answer:** {extractive}\n\n**Citations:** {header_cites}{coverage_note}"
+            final = (
+                f"**Answer:** {extractive}\n\n"
+                f"**Citations:** {header_cites}{coverage_note}"
+            )
             if include_passages:
-                final += "\n\n---\n" + "\n\n".join(hits['text'].tolist()[:2])
+                final += (
+                    "\n\n---\n" + "\n\n".join(hits['text'].tolist()[:2])
+                )
 
     # --------- Log full run ---------
     prompt_toks = llm_usage.get("prompt_tokens") if llm_usage else None
@@ -743,7 +862,10 @@ def rag_reply(
         "retrieval": {"hits": retr_list, "latency_ms_retriever": latency_ms_retriever},
         "output": {
             "final_answer": final,
-            "used_sentences": [{"sent": s['sent'], "doc": s['doc'], "page": s['page']} for s in selected]
+            "used_sentences": [
+                {"sent": s['sent'], "doc": s['doc'], "page": s['page']}
+                for s in selected
+            ]
         },
         "latency_ms_total": total_ms,
         "latency_ms_llm": llm_latency_ms,
@@ -1034,7 +1156,8 @@ with gr.Blocks(css=CSS, theme=theme, fill_height=True) as demo:
         "<h1 style='margin:0'>Self-Sensing Concrete Assistant</h1>"
         "<p style='opacity:.9'>"
         "Left: ML prediction for Stress Gauge Factor (original scale, MPa<sup>-1</sup>). "
-        "Right: Literature Q&A via Hybrid RAG (BM25 + TF-IDF + optional dense) with MMR sentence selection."
+        "Right: Literature Q&A via Hybrid RAG (BM25 + TF-IDF + optional dense) with MMR sentence selection. "
+        "Answers use APA-style inline citations (e.g., Omar et al., 2021, p. X)."
         "</p>"
     )
 
@@ -1078,7 +1201,12 @@ with gr.Blocks(css=CSS, theme=theme, fill_height=True) as demo:
 
                 with gr.Column(scale=5):
                     with gr.Group(elem_classes=["card"]):
-                        out_pred = gr.Number(label="Predicted Stress GF (MPa-1)", value=0.0, precision=6, elem_id="pred-out")
+                        out_pred = gr.Number(
+                            label="Predicted Stress GF (MPa-1)",
+                            value=0.0,
+                            precision=6,
+                            elem_id="pred-out"
+                        )
                         gr.Markdown(f"<small>{MODEL_STATUS}</small>")
                         with gr.Row():
                             btn_pred  = gr.Button("Predict", variant="primary")
@@ -1088,7 +1216,8 @@ with gr.Blocks(css=CSS, theme=theme, fill_height=True) as demo:
                     with gr.Accordion("About this model", open=False, elem_classes=["card"]):
                         gr.Markdown(
                             "- Pipeline: ColumnTransformer → (RobustScaler + OneHot) → XGBoost\n"
-                            "- Target: Stress GF (MPa<sup>-1</sup>) on original scale (model may train on log1p; saved flag used at inference).\n"
+                            "- Target: Stress GF (MPa<sup>-1</sup>) on original scale "
+                            "(model may train on log1p; saved flag used at inference).\n"
                             "- Missing values are safely imputed per-feature.\n"
                             "- Trained columns:\n"
                             f"  `{', '.join(MAIN_VARIABLES)}`",
@@ -1108,26 +1237,51 @@ with gr.Blocks(css=CSS, theme=theme, fill_height=True) as demo:
                 data = {k: v for k, v in zip(MAIN_VARIABLES, vals)}
                 return predict_fn(**data)
 
-            btn_pred.click(_predict_wrapper, inputs=inputs_in_order, outputs=out_pred)
-            btn_clear.click(lambda: _clear_all(), inputs=None, outputs=inputs_in_order).then(lambda: 0.0, outputs=out_pred)
-            btn_demo.click(lambda: _fill_example(), inputs=None, outputs=inputs_in_order)
+            btn_pred.click(
+                _predict_wrapper,
+                inputs=inputs_in_order,
+                outputs=out_pred
+            )
+            btn_clear.click(
+                lambda: _clear_all(),
+                inputs=None,
+                outputs=inputs_in_order
+            ).then(
+                lambda: 0.0,
+                outputs=out_pred
+            )
+            btn_demo.click(
+                lambda: _fill_example(),
+                inputs=None,
+                outputs=inputs_in_order
+            )
 
         # ------------------------- Literature Tab -------------------------
         with gr.Tab("📚 Ask the Literature (Hybrid RAG + MMR)", elem_id="rag-tab"):
             pdf_count = len(list(LOCAL_PDF_DIR.glob("**/*.pdf")))
             gr.Markdown(
                 f"Using local folder <code>papers/</code> — **{pdf_count} PDF(s)** indexed. "
-                "Upload more PDFs and reload the Space to expand coverage. Answers cite (Doc.pdf, p.X)."
+                "Upload more PDFs and reload the Space to expand coverage. "
+                "Answers use APA-like inline citations (e.g., Omar et al., 2021, p. X)."
             )
             with gr.Row():
                 top_k = gr.Slider(5, 12, value=8, step=1, label="Top-K chunks")
                 n_sentences = gr.Slider(2, 6, value=4, step=1, label="Answer length (sentences)")
-                include_passages = gr.Checkbox(value=False, label="Include supporting passages", interactive=True)
+                include_passages = gr.Checkbox(
+                    value=False,
+                    label="Include supporting passages",
+                    interactive=True
+                )
 
             with gr.Accordion("Retriever weights (advanced)", open=False):
                 w_tfidf = gr.Slider(0.0, 1.0, value=W_TFIDF_DEFAULT, step=0.05, label="TF-IDF weight")
                 w_bm25  = gr.Slider(0.0, 1.0, value=W_BM25_DEFAULT,  step=0.05, label="BM25 weight")
-                w_emb   = gr.Slider(0.0, 1.0, value=(0.0 if not USE_DENSE else 0.40), step=0.05, label="Dense weight (set 0 if disabled)")
+                w_emb   = gr.Slider(
+                    0.0, 1.0,
+                    value=(0.0 if not USE_DENSE else 0.40),
+                    step=0.05,
+                    label="Dense weight (set 0 if disabled)"
+                )
 
             # Hidden states (unchanged)
             state_use_llm     = gr.State(LLM_AVAILABLE)
@@ -1143,7 +1297,7 @@ with gr.Blocks(css=CSS, theme=theme, fill_height=True) as demo:
                     w_tfidf, w_bm25, w_emb
                 ],
                 title="Literature Q&A",
-                description="Hybrid retrieval with diversity. Answers carry inline (Doc, p.X) citations."
+                description="Hybrid retrieval with diversity. Answers carry APA-style inline citations."
             )
 
         # ====== Evaluate (Gold vs Logs) — darker, higher-contrast ======
@@ -1151,7 +1305,8 @@ with gr.Blocks(css=CSS, theme=theme, fill_height=True) as demo:
             gr.Markdown("Upload your **gold.csv** and compute metrics against the app logs.")
             with gr.Row():
                 gold_file = gr.File(label="gold.csv", file_types=[".csv"], interactive=True)
-                k_slider  = gr.Slider(3, 12, value=8, step=1, label="k for Hit/Recall/nDCG", elem_id="k-slider")
+                k_slider  = gr.Slider(3, 12, value=8, step=1,
+                                      label="k for Hit/Recall/nDCG", elem_id="k-slider")
             with gr.Row():
                 btn_eval = gr.Button("Compute Metrics", variant="primary")
             with gr.Row():
@@ -1172,7 +1327,12 @@ with gr.Blocks(css=CSS, theme=theme, fill_height=True) as demo:
                     "--out_dir", out_dir
                 ]
                 try:
-                    p = subprocess.run(cmd, capture_output=True, text=True, check=False)
+                    p = subprocess.run(
+                        cmd,
+                        capture_output=True,
+                        text=True,
+                        check=False
+                    )
                     stdout = p.stdout or ""
                     stderr = p.stderr or ""
                     perq = ARTIFACT_DIR / "metrics_per_question.csv"
@@ -1180,18 +1340,25 @@ with gr.Blocks(css=CSS, theme=theme, fill_height=True) as demo:
                     agg_json = {}
                     if agg.exists():
                         agg_json = _json.loads(agg.read_text(encoding="utf-8"))
-                    report = "```\n" + (stdout.strip() or "(no stdout)") + ("\n" + stderr.strip() if stderr else "") + "\n```"
-                    return (str(perq) if perq.exists() else None,
-                            str(agg)  if agg.exists()  else None,
-                            agg_json,
-                            report)
+                    report = (
+                        "```\n"
+                        + (stdout.strip() or "(no stdout)")
+                        + ("\n" + stderr.strip() if stderr else "")
+                        + "\n```"
+                    )
+                    return (
+                        str(perq) if perq.exists() else None,
+                        str(agg)  if agg.exists()  else None,
+                        agg_json,
+                        report
+                    )
                 except Exception as e:
                     return (None, None, {}, f"**Eval error:** {e}")
 
             def _eval_wrapper(gf, k):
-                from pathlib import Path
+                from pathlib import Path as _Path
                 if gf is None:
-                    default_gold = Path("gold.csv")
+                    default_gold = _Path("gold.csv")
                     if not default_gold.exists():
                         return None, None, {}, "**No gold.csv provided or found in repo root.**"
                     gold_path = str(default_gold)
@@ -1199,23 +1366,26 @@ with gr.Blocks(css=CSS, theme=theme, fill_height=True) as demo:
                     gold_path = gf.name
                 return _run_eval_inproc(gold_path, int(k))
 
-            btn_eval.click(_eval_wrapper, inputs=[gold_file, k_slider],
-                           outputs=[out_perq, out_agg, out_json, out_log])
+            btn_eval.click(
+                _eval_wrapper,
+                inputs=[gold_file, k_slider],
+                outputs=[out_perq, out_agg, out_json, out_log]
+            )
 
 # ------------- Launch -------------
 if __name__ == "__main__":
     demo.queue().launch()
-    import os
-import pandas as pd
 
-# Folder where your RAG files are stored
-folder = "papers"  # change if needed
+    import os
+    import pandas as pd
 
-# List all files in the folder
-files = sorted(os.listdir(folder))
+    # Folder where your RAG files are stored
+    folder = "papers"  # change if needed
 
-# Save them to a CSV file
-pd.DataFrame({"doc": files}).to_csv("paper_list.csv", index=False)
+    # List all files in the folder
+    files = sorted(os.listdir(folder))
 
-print("✅ Saved paper_list.csv with", len(files), "papers")
+    # Save them to a CSV file
+    pd.DataFrame({"doc": files}).to_csv("paper_list.csv", index=False)
 
+    print("✅ Saved paper_list.csv with", len(files), "papers")