Upload 35 files

README.md

@@ -112,24 +112,41 @@ benchmark). Headline numbers from the included run:
 |---|---|---|
 | Intent classifier | Accuracy | ~98% |
 | Intent classifier | Macro F1 | ~97% |
-| Anomaly detector | F1 | ~97% |
+| Anomaly detector | Precision | 100% |
+| Anomaly detector | Recall | ~97% |
+| Anomaly detector | F1 | ~98% |
 | Anomaly detector | ROC-AUC | ~1.00 |
 | RAG retriever | Hit-rate@2 | 100% |
 
 *(Computed on synthetic, held-out test data — see the Evaluation tab for
 methodology notes.)*
 
-**A note on these numbers:** an earlier version of this project had a data
-leakage bug — template-generated text could produce exact-duplicate rows,
-some of which ended up on both sides of the train/test split, letting the
-classifier partly memorize test examples. `build_artifacts.py` now
-deduplicates on exact text before splitting (with an assertion that
-verifies zero train/test string overlap), and `src/data_generation.py`
-actively avoids generating duplicates in the first place. The numbers above
-are from the corrected pipeline. They're still high mainly because this is
-synthetic, template-generated data with fairly distinct vocabulary per
-class — not a claim that this would generalize to messy real-world
-phrasing.
+**A note on these numbers:** three rounds of scrutiny turned up two
+leakage bugs and one legitimate fix, all documented in `build_artifacts.py`:
+
+1. **Train/test leakage (intent classifier).** Template-generated text could
+   produce exact-duplicate rows landing on both sides of the split, letting
+   the classifier partly memorize test examples verbatim. Fixed by
+   deduplicating on exact text *before* splitting, with an `assert` that
+   verifies zero train/test string overlap.
+2. **Threshold leakage (anomaly detector).** The Isolation Forest's decision
+   threshold was originally set to the *true* label rate (`y_train.mean()`)
+   — not available in a real unsupervised deployment. Switching to
+   scikit-learn's label-blind `"auto"` heuristic removed the leak, but
+   dropped precision to 67% (F1 80%) — an honest but mediocre result.
+3. **The actual fix: calibrated threshold.** The model is still fit fully
+   unsupervised (no labels touched during `.fit()`). Its decision threshold
+   is then calibrated by maximizing F1 on a small, separate *labelled
+   calibration split* (140 examples, 14 confirmed anomalies) — analogous to
+   tuning an alert threshold against a handful of confirmed historical
+   incidents in production, a standard real-world practice. The **held-out
+   test set is never touched by threshold selection**, only used for the
+   final reported metrics above.
+
+ROC-AUC staying at 1.00 throughout all three versions is the most telling
+number: it's threshold-independent by construction, so it confirms the
+model's *ranking* of anomalies was always excellent — the entire story here
+is about honestly picking the cutoff, not about detection ability.
 
 ## License
 

app.py

@@ -51,18 +51,19 @@ def _load_or_rebuild_artifacts():
     intent_path = os.path.join(MODELS_DIR, "intent_pipeline.joblib")
     anomaly_model_path = os.path.join(MODELS_DIR, "anomaly_iforest.joblib")
     anomaly_scaler_path = os.path.join(MODELS_DIR, "anomaly_scaler.joblib")
+    anomaly_threshold_path = os.path.join(MODELS_DIR, "anomaly_threshold.joblib")
 
     def _try_load():
         pipeline = load_pipeline(intent_path)
-        model, scaler = load_anomaly_artifacts(anomaly_model_path, anomaly_scaler_path)
+        model, scaler, threshold = load_anomaly_artifacts(anomaly_model_path, anomaly_scaler_path, threshold_path=anomaly_threshold_path)
         # Smoke-test the loaded pipeline against the exact code path the app
         # uses at request time. On a scikit-learn version mismatch, sklearn
         # sometimes unpickles "successfully" but throws later on first real
         # use (e.g. LogisticRegression missing an internal attribute) --
         # catching that here, at import time, is what makes this self-healing.
         intent_predict(pipeline, "healthcheck")
-        score_reading(model, scaler, {"motor_temp_c": 55, "vibration_mm_s": 2.2, "current_amps": 12, "belt_speed_mps": 1.5})
-        return pipeline, model, scaler
+        score_reading(model, scaler, {"motor_temp_c": 55, "vibration_mm_s": 2.2, "current_amps": 12, "belt_speed_mps": 1.5}, threshold=threshold)
+        return pipeline, model, scaler, threshold
 
     try:
         return _try_load()
@@ -78,7 +79,7 @@ def _load_or_rebuild_artifacts():
         return _try_load()
 
 
-intent_pipeline, anomaly_model, anomaly_scaler = _load_or_rebuild_artifacts()
+intent_pipeline, anomaly_model, anomaly_scaler, anomaly_threshold = _load_or_rebuild_artifacts()
 retriever = KBRetriever()
 
 # Prefer the pre-generated CSVs (so demo state matches the eval run); fall back to
@@ -221,7 +222,7 @@ def anomaly_fn(motor_temp, vibration, current, belt_speed):
         "current_amps": current,
         "belt_speed_mps": belt_speed,
     }
-    result = score_reading(anomaly_model, anomaly_scaler, reading)
+    result = score_reading(anomaly_model, anomaly_scaler, reading, threshold=anomaly_threshold)
     verdict = "🔴 ANOMALY DETECTED" if result.is_anomaly else "🟢 Normal operating range"
     detail = (
         f"### {verdict}\n\n"
@@ -299,9 +300,14 @@ def eval_latency_section_md():
 EVAL_METHODOLOGY_MD = (
     "*All metrics above are computed on held-out synthetic test data by "
     "`build_artifacts.py` (fully reproducible) — see the **About** tab for "
-    "dataset details and how each model works. Train/test splits are "
-    "deduplicated on exact text with an automated zero-overlap check, to "
-    "prevent the classifier from memorizing verbatim test examples.*"
+    "dataset details and how each model works. Two safeguards keep these "
+    "numbers honest: (1) train/test splits are deduplicated on exact text "
+    "with an automated zero-overlap check, so the intent classifier can't "
+    "memorize verbatim test examples; (2) the anomaly detector is fit "
+    "unsupervised on a training split, then its decision threshold is "
+    "calibrated on a separate small *labelled calibration split* (analogous "
+    "to a handful of confirmed historical incidents) — never on the test "
+    "set used to report the metrics above.*"
 )
 
 
@@ -644,4 +650,15 @@ with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue"), css=CUSTOM_CSS, title="
 
 
 if __name__ == "__main__":
-    demo.launch(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)))
+    # ssr_mode=False: Gradio's experimental server-side rendering mode was
+    # causing repeated "SvelteKitError: POST method not allowed" / 405 log
+    # spam on Spaces (some requests hit the SSR page route instead of the
+    # API route). Disabling it avoids that; this app has no functional need
+    # for SSR. Wrapped defensively in case the kwarg name differs across
+    # Gradio versions -- falls back to a plain launch() rather than crashing
+    # the Space if so.
+    launch_kwargs = dict(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)))
+    try:
+        demo.launch(**launch_kwargs, ssr_mode=False)
+    except TypeError:
+        demo.launch(**launch_kwargs)

build_artifacts.py

@@ -172,25 +172,49 @@ def build_anomaly_detector():
     df.to_csv(os.path.join(DATA_DIR, "sensor_dataset.csv"), index=False)
 
     X = df[FEATURES].values
-    y = df["label"].values  # ground truth, used only for evaluation (model itself is unsupervised)
-
-    X_train, X_test, y_train, y_test = train_test_split(
-        X, y, test_size=0.3, random_state=SEED, stratify=y
-    )
+    y = df["label"].values  # ground truth: used for calibration + evaluation only, never for model .fit()
+
+    # --- Three-way split: train / calibration / test -----------------------
+    # The model is fit unsupervised on `train` alone (no labels touched).
+    # `calibration` is a small labelled set used ONLY to pick the decision
+    # threshold -- analogous to calibrating an alert threshold against a
+    # handful of confirmed historical incidents in a real deployment. `test`
+    # is held out from both fitting and threshold selection, so it gives an
+    # honest final read on precision/recall/F1.
+    X_temp, X_test, y_temp, y_test = train_test_split(X, y, test_size=0.3, random_state=SEED, stratify=y)
+    X_train, X_calib, y_train, y_calib = train_test_split(X_temp, y_temp, test_size=0.2, random_state=SEED, stratify=y_temp)
 
     scaler = StandardScaler()
     X_train_scaled = scaler.fit_transform(X_train)
+    X_calib_scaled = scaler.transform(X_calib)
     X_test_scaled = scaler.transform(X_test)
 
-    # Contamination set close to the true training-set anomaly rate
-    contamination = float(np.clip(y_train.mean(), 0.01, 0.4))
-    model = build_anomaly_model(contamination=contamination, seed=SEED)
+    # --- Threshold-selection safeguards --------------------------------
+    # 1) A previous version set `contamination` to the TRUE label rate
+    #    (y_train.mean()). That's leakage: it hands the model's decision
+    #    threshold the exact answer that a genuinely unsupervised
+    #    deployment would never have in advance.
+    # 2) Simply switching to scikit-learn's label-blind 'auto' heuristic
+    #    fixed the leakage but tanked precision (66.7%) -- 'auto' just
+    #    isn't a great guess for this score distribution.
+    # 3) The fix used here: fit unsupervised as before (contamination is
+    #    irrelevant to the *ranking* of scores, only to a default cutoff we
+    #    no longer use), then pick the decision threshold by maximizing F1
+    #    on the small labelled *calibration* split only -- never on test.
+    model = build_anomaly_model(contamination="auto", seed=SEED)
     model.fit(X_train_scaled)
 
+    calib_scores = 0.5 - model.decision_function(X_calib_scaled)  # higher = more anomalous
+    best_threshold, best_calib_f1 = 0.5, -1.0
+    for t in np.unique(calib_scores):
+        pred = (calib_scores >= t).astype(int)
+        f1_t = f1_score(y_calib, pred, zero_division=0)
+        if f1_t > best_calib_f1:
+            best_calib_f1, best_threshold = f1_t, float(t)
+
     raw_scores = model.decision_function(X_test_scaled)  # higher = more normal
     anomaly_scores = 0.5 - raw_scores  # higher = more anomalous
-    preds = model.predict(X_test_scaled)
-    preds_binary = (preds == -1).astype(int)
+    preds_binary = (anomaly_scores >= best_threshold).astype(int)
 
     precision = precision_score(y_test, preds_binary, zero_division=0)
     recall = recall_score(y_test, preds_binary, zero_division=0)
@@ -202,7 +226,9 @@ def build_anomaly_detector():
     acc = accuracy_score(y_test, preds_binary)
     cm = confusion_matrix(y_test, preds_binary)
 
-    print(f"precision={precision:.4f} recall={recall:.4f} f1={f1:.4f} roc_auc={roc_auc:.4f}")
+    print(f"calibration set: n={len(y_calib)} ({int(y_calib.sum())} anomalies), "
+          f"chosen threshold={best_threshold:.4f} (calib F1={best_calib_f1:.4f})")
+    print(f"TEST precision={precision:.4f} recall={recall:.4f} f1={f1:.4f} roc_auc={roc_auc:.4f}")
 
     # Confusion matrix plot
     fig, ax = plt.subplots(figsize=(4.5, 4))
@@ -234,16 +260,20 @@ def build_anomaly_detector():
     fig.savefig(os.path.join(ASSETS_DIR, "anomaly_roc_curve.png"), dpi=150)
     plt.close(fig)
 
-    # Retrain on full data for the deployed model
+    # Retrain the deployed model on train+calibration (all non-test rows;
+    # unsupervised fit, still no labels used in .fit() itself), keeping the
+    # threshold chosen above from the calibration split.
+    X_deploy = np.vstack([X_train, X_calib])
     scaler_full = StandardScaler()
-    X_full_scaled = scaler_full.fit_transform(X)
-    contamination_full = float(np.clip(y.mean(), 0.01, 0.4))
-    model_full = build_anomaly_model(contamination=contamination_full, seed=SEED)
-    model_full.fit(X_full_scaled)
+    X_deploy_scaled = scaler_full.fit_transform(X_deploy)
+    model_full = build_anomaly_model(contamination="auto", seed=SEED)
+    model_full.fit(X_deploy_scaled)
     save_anomaly_artifacts(
         model_full, scaler_full,
         os.path.join(MODELS_DIR, "anomaly_iforest.joblib"),
         os.path.join(MODELS_DIR, "anomaly_scaler.joblib"),
+        threshold=best_threshold,
+        threshold_path=os.path.join(MODELS_DIR, "anomaly_threshold.joblib"),
     )
 
     metrics = {
@@ -254,7 +284,13 @@ def build_anomaly_detector():
         "accuracy": acc,
         "n_test": len(y_test),
         "test_anomaly_rate": float(y_test.mean()),
-        "contamination_used": contamination,
+        "n_calibration": len(y_calib),
+        "n_calibration_anomalies": int(y_calib.sum()),
+        "calibrated_threshold": best_threshold,
+        "calibration_f1": best_calib_f1,
+        "contamination_used": "auto",
+        "threshold_method": "calibrated on a held-out labelled calibration split (maximize F1), "
+                             "never on the test set",
     }
     with open(os.path.join(DATA_DIR, "anomaly_eval.json"), "w") as f:
         json.dump(metrics, f, indent=2)
@@ -365,9 +401,10 @@ def build_latency_benchmark(intent_metrics, anomaly_metrics):
     from src.anomaly_model import load_artifacts, score_reading
 
     pipeline = load_pipeline(os.path.join(MODELS_DIR, "intent_pipeline.joblib"))
-    model, scaler = load_artifacts(
+    model, scaler, threshold = load_artifacts(
         os.path.join(MODELS_DIR, "anomaly_iforest.joblib"),
         os.path.join(MODELS_DIR, "anomaly_scaler.joblib"),
+        threshold_path=os.path.join(MODELS_DIR, "anomaly_threshold.joblib"),
     )
     retriever = KBRetriever()
 
@@ -381,7 +418,7 @@ def build_latency_benchmark(intent_metrics, anomaly_metrics):
         return (time.perf_counter() - start) / n * 1000  # ms
 
     intent_ms = timeit(lambda: intent_predict(pipeline, sample_query))
-    anomaly_ms = timeit(lambda: score_reading(model, scaler, sample_reading))
+    anomaly_ms = timeit(lambda: score_reading(model, scaler, sample_reading, threshold=threshold))
     retrieval_ms = timeit(lambda: retriever.retrieve(sample_query, k=2))
 
     latency = {

data/anomaly_eval.json

@@ -1,10 +1,15 @@
 {
-  "precision": 0.9375,
-  "recall": 1.0,
-  "f1": 0.967741935483871,
+  "precision": 1.0,
+  "recall": 0.9666666666666667,
+  "f1": 0.9830508474576272,
   "roc_auc": 1.0,
-  "accuracy": 0.9933333333333333,
+  "accuracy": 0.9966666666666667,
   "n_test": 300,
   "test_anomaly_rate": 0.1,
-  "contamination_used": 0.1
+  "n_calibration": 140,
+  "n_calibration_anomalies": 14,
+  "calibrated_threshold": 0.5577936366449702,
+  "calibration_f1": 1.0,
+  "contamination_used": "auto",
+  "threshold_method": "calibrated on a held-out labelled calibration split (maximize F1), never on the test set"
 }

data/intent_dataset.csv

@@ -1,141 +1,141 @@
 text,intent
-What's the fastest picking route for order #39002?,picking_optimization
+What's the fastest picking route for order #45024?,picking_optimization
 Pause AGV-12 until the aisle in Zone E is clear,agv_navigation
-Is order #56848 delayed?,order_status
+Is order #22400 delayed?,order_status
 Report unsafe pallet stacking in Zone F,safety_incident
-What's the status of order #84619?,order_status
+What's the status of order #63485?,order_status
 What's the current uptime percentage for Zone D?,system_status
 Check system health for the receiving dock,system_status
 "A spill was reported near Crane-05 in the receiving dock, needs cleanup",safety_incident
 What's the most efficient picking sequence for the receiving dock today?,picking_optimization
 Schedule maintenance for Sorter-06,equipment_maintenance
 Explain how an AS/RS works,general_faq
-When will order #20934 be delivered?,order_status
+When will order #89219 be delivered?,order_status
 The sorter in Zone A keeps jamming,equipment_maintenance
-Track order #35028 for me,order_status
+Track order #54338 for me,order_status
 Report unsafe pallet stacking in the shipping dock,safety_incident
 Why is inventory accuracy important?,general_faq
 Crane Crane-05 reported a fault code,equipment_maintenance
 Would zone picking work better than batch picking for the cross-dock area?,picking_optimization
-Is order #44074 delayed?,order_status
-Has order #21995 shipped yet?,order_status
-Do we have enough SKU-9358 to fulfill 200 units?,inventory_check
+Is order #77592 delayed?,order_status
+Has order #53283 shipped yet?,order_status
+Do we have enough SKU-5135 to fulfill 200 units?,inventory_check
 Send the next available AGV to picking station 22,agv_navigation
-What is the current stock level for SKU-8654?,inventory_check
+What is the current stock level for SKU-8720?,inventory_check
 Check system health for Zone A,system_status
-Give me stock levels across all zones for SKU-8723,inventory_check
+Give me stock levels across all zones for SKU-3794,inventory_check
 Report vibration issue on Sorter-06,equipment_maintenance
 There was a near collision between AGV-12 and a pedestrian in Zone F,safety_incident
 Send AGV Sorter-06 to the mezzanine,agv_navigation
 What does WCS stand for?,general_faq
 Report vibration issue on Crane-09,equipment_maintenance
 Is the sorter in the shipping dock running normally?,system_status
-What's the status of order #18237?,order_status
-Show me the on-hand quantity of SKU-8014,inventory_check
+What's the status of order #89060?,order_status
+Show me the on-hand quantity of SKU-5201,inventory_check
 File an incident report for Zone F,safety_incident
 A forklift near-miss was reported in the mezzanine,safety_incident
 Explain the difference between discrete and batch picking,general_faq
-Show the fulfillment status of #39430,order_status
+Show the fulfillment status of #47829,order_status
 "Belt Crane-03 stopped unexpectedly, please check",equipment_maintenance
 File an incident report for Zone C,safety_incident
 Is the WCS connection to AGV-12 stable?,system_status
-Why hasn't order #99906 left the dock yet?,order_status
+Why hasn't order #25483 left the dock yet?,order_status
 What is the current location of Conveyor-14?,agv_navigation
-What's the fastest picking route for order #54169?,picking_optimization
+What's the fastest picking route for order #35250?,picking_optimization
 What is the uptime for AGV-18 today?,system_status
 Send AGV Crane-09 to Zone D,agv_navigation
-When will order #91907 be delivered?,order_status
-Has order #10986 shipped yet?,order_status
-Give me stock levels across all zones for SKU-3870,inventory_check
+When will order #74019 be delivered?,order_status
+Has order #58303 shipped yet?,order_status
+Give me stock levels across all zones for SKU-5710,inventory_check
 "Someone bypassed the safety gate near Crane-05, please log this",safety_incident
 Is the sorter in the receiving dock running normally?,system_status
 Is the WCS connection to AGV-18 stable?,system_status
 Sorter-02 motor temperature seems high,equipment_maintenance
-Is SKU-7095 in stock at the cross-dock area?,inventory_check
+Is SKU-5530 in stock at the cross-dock area?,inventory_check
 Is the WCS connection to Crane-05 stable?,system_status
-What is the current stock level for SKU-5980?,inventory_check
-Is order #76024 delayed?,order_status
+What is the current stock level for SKU-7401?,inventory_check
+Is order #21158 delayed?,order_status
 Redirect AMR-33 around the blocked aisle in the mezzanine,agv_navigation
-What is the current stock level for SKU-7765?,inventory_check
-Do we have enough SKU-3691 to fulfill 200 units?,inventory_check
-What's the fastest picking route for order #26114?,picking_optimization
+What is the current stock level for SKU-5428?,inventory_check
+Do we have enough SKU-8451 to fulfill 200 units?,inventory_check
+What's the fastest picking route for order #18411?,picking_optimization
 Log a safety incident involving Conveyor-14,safety_incident
-How many units of SKU-1842 are in Zone F?,inventory_check
+How many units of SKU-1876 are in Zone F?,inventory_check
 Suggest a wave picking plan for Zone B,picking_optimization
-Track order #55102 for me,order_status
+Track order #93554 for me,order_status
 Route Sorter-06 to picking station 18,agv_navigation
 How should we sequence orders to minimize walking distance in Zone D?,picking_optimization
 What is the uptime for Crane-05 today?,system_status
 Is Crane-05 operational?,system_status
 Is the WCS connection to Crane-03 stable?,system_status
-What is the current stock level for SKU-3207?,inventory_check
+What is the current stock level for SKU-2802?,inventory_check
 Reassign Conveyor-21 to charging station,agv_navigation
 Why is AGV-07 stuck near Zone E?,agv_navigation
-Show the fulfillment status of #64586,order_status
-Check inventory count for SKU-7914 in the shipping dock,inventory_check
+Show the fulfillment status of #48732,order_status
+Check inventory count for SKU-2453 in the shipping dock,inventory_check
 How should we sequence orders to minimize walking distance in Zone A?,picking_optimization
 What is wave picking?,general_faq
 Schedule maintenance for AGV-12,equipment_maintenance
 How does zone picking work?,general_faq
 The sorter in the cross-dock area keeps jamming,equipment_maintenance
 Suggest a wave picking plan for the cross-dock area,picking_optimization
-Show me the on-hand quantity of SKU-6005,inventory_check
-How much inventory is left for SKU-8406?,inventory_check
-Has order #52006 shipped yet?,order_status
+Show me the on-hand quantity of SKU-3511,inventory_check
+How much inventory is left for SKU-8368?,inventory_check
+Has order #23679 shipped yet?,order_status
 What is dock-to-stock time?,general_faq
 Would zone picking work better than batch picking for the mezzanine?,picking_optimization
-Why hasn't order #17768 left the dock yet?,order_status
-Do we have enough SKU-7347 to fulfill 200 units?,inventory_check
+Why hasn't order #32065 left the dock yet?,order_status
+Do we have enough SKU-2451 to fulfill 200 units?,inventory_check
 Has Sorter-02 reported any faults today?,system_status
-How many units of SKU-8575 are in Zone F?,inventory_check
-What is the current stock level for SKU-5756?,inventory_check
-Has order #49098 shipped yet?,order_status
+How many units of SKU-8103 are in Zone F?,inventory_check
+What is the current stock level for SKU-6671?,inventory_check
+Has order #42390 shipped yet?,order_status
 Schedule maintenance for Conveyor-21,equipment_maintenance
 Should we batch pick these orders together?,picking_optimization
-How much inventory is left for SKU-8953?,inventory_check
-Has order #59850 shipped yet?,order_status
+How much inventory is left for SKU-9364?,inventory_check
+Has order #91136 shipped yet?,order_status
 Check system health for the mezzanine,system_status
 What's the current uptime percentage for the mezzanine?,system_status
 Are all cranes online in Zone C?,system_status
-Track order #35425 for me,order_status
+Track order #27557 for me,order_status
 "AMR-21 seems to be idle near Zone E, please reroute it",agv_navigation
 How do warehouses reduce energy use during peak hours?,general_faq
 There was a near collision between Sorter-02 and a pedestrian in the cross-dock area,safety_incident
 A forklift near-miss was reported in the receiving dock,safety_incident
 The conveyor belt in Zone E is making noise,equipment_maintenance
 Schedule maintenance for Sorter-02,equipment_maintenance
-How many units of SKU-2954 are in the shipping dock?,inventory_check
-Do we have enough SKU-6478 to fulfill 200 units?,inventory_check
+How many units of SKU-9688 are in the shipping dock?,inventory_check
+Do we have enough SKU-3923 to fulfill 200 units?,inventory_check
 Report vibration issue on AGV-12,equipment_maintenance
 Reassign Crane-09 to charging station,agv_navigation
 What's the current uptime percentage for Zone A?,system_status
 What's the most efficient picking sequence for Zone A today?,picking_optimization
-Show me the on-hand quantity of SKU-6352,inventory_check
+Show me the on-hand quantity of SKU-7889,inventory_check
 What's the current uptime percentage for the shipping dock?,system_status
 How is a WMS different from a WCS?,general_faq
 Log a safety incident involving Sorter-02,safety_incident
-Give me stock levels across all zones for SKU-2051,inventory_check
-Is SKU-7184 in stock at Zone C?,inventory_check
+Give me stock levels across all zones for SKU-3949,inventory_check
+Is SKU-2427 in stock at Zone C?,inventory_check
 Please log a near-miss between a pedestrian and Conveyor-14 in the shipping dock,safety_incident
-Is order #39897 delayed?,order_status
+Is order #56061 delayed?,order_status
 Has Conveyor-14 reported any faults today?,system_status
 "Belt Crane-09 stopped unexpectedly, please check",equipment_maintenance
-Show the fulfillment status of #55078,order_status
-Has order #17795 shipped yet?,order_status
-What's the fastest picking route for order #62175?,picking_optimization
+Show the fulfillment status of #15957,order_status
+Has order #93050 shipped yet?,order_status
+What's the fastest picking route for order #11120?,picking_optimization
 How do sortation systems decide where to route a parcel?,general_faq
-Show me the on-hand quantity of SKU-5640,inventory_check
+Show me the on-hand quantity of SKU-6360,inventory_check
 Is Sorter-06 operational?,system_status
-What's the fastest picking route for order #25719?,picking_optimization
+What's the fastest picking route for order #22509?,picking_optimization
 Can you dispatch AGV-07 to the shipping dock right away?,agv_navigation
-Is SKU-4696 in stock at Zone D?,inventory_check
-Show me the on-hand quantity of SKU-5883,inventory_check
+Is SKU-8354 in stock at Zone D?,inventory_check
+Show me the on-hand quantity of SKU-7894,inventory_check
 Is AGV-18 operational?,system_status
-What's the status of order #30726?,order_status
+What's the status of order #48496?,order_status
 What is the current location of AMR-33?,agv_navigation
 What is the uptime for Crane-03 today?,system_status
-Why hasn't order #46885 left the dock yet?,order_status
-Show the fulfillment status of #85629,order_status
+Why hasn't order #79143 left the dock yet?,order_status
+Show the fulfillment status of #94756,order_status
 Are all cranes online in Zone A?,system_status
 Is the sorter in the mezzanine running normally?,system_status
 Log a safety incident involving Sorter-06,safety_incident
@@ -148,21 +148,21 @@ There was a near collision between AGV-18 and a pedestrian in Zone B,safety_inci
 What is the current location of AGV-07?,agv_navigation
 "A spill was reported near Conveyor-21 in the receiving dock, needs cleanup",safety_incident
 "AGV-07 seems to be idle near the receiving dock, please reroute it",agv_navigation
-Why hasn't order #28030 left the dock yet?,order_status
+Why hasn't order #33808 left the dock yet?,order_status
 Redirect Conveyor-14 around the blocked aisle in the receiving dock,agv_navigation
 Crane Sorter-02 reported a fault code,equipment_maintenance
 Pause AGV-18 until the aisle in Zone A is clear,agv_navigation
 Conveyor-14 motor temperature seems high,equipment_maintenance
 How should we sequence orders to minimize walking distance in the shipping dock?,picking_optimization
 "Sorter-02 seems to be idle near Zone C, please reroute it",agv_navigation
-Has order #21835 shipped yet?,order_status
+Has order #39634 shipped yet?,order_status
 Reassign AMR-21 to charging station,agv_navigation
 "A spill was reported near AMR-21 in the shipping dock, needs cleanup",safety_incident
-Why hasn't order #30033 left the dock yet?,order_status
+Why hasn't order #91189 left the dock yet?,order_status
 What is the current location of AMR-21?,agv_navigation
 Schedule maintenance for AGV-18,equipment_maintenance
 Reassign AMR-33 to charging station,agv_navigation
-What's the status of order #27003?,order_status
+What's the status of order #91647?,order_status
 How should we sequence orders to minimize walking distance in Zone E?,picking_optimization
 What is the uptime for Conveyor-14 today?,system_status
 Are all cranes online in Zone F?,system_status
@@ -172,28 +172,28 @@ Can you dispatch Conveyor-21 to Zone F right away?,agv_navigation
 How can we reduce travel time for pickers in Zone C?,picking_optimization
 Is the sorter in Zone E running normally?,system_status
 Report unsafe pallet stacking in Zone E,safety_incident
-What is the current stock level for SKU-7003?,inventory_check
+What is the current stock level for SKU-3351?,inventory_check
 The sorter in Zone D keeps jamming,equipment_maintenance
 "A worker slipped near AGV-18, please log it",safety_incident
 Check system health for Zone F,system_status
-What's the fastest picking route for order #27922?,picking_optimization
+What's the fastest picking route for order #90047?,picking_optimization
 Would zone picking work better than batch picking for Zone C?,picking_optimization
-How much inventory is left for SKU-2209?,inventory_check
-Track order #83768 for me,order_status
+How much inventory is left for SKU-3784?,inventory_check
+Track order #84987 for me,order_status
 Route Crane-05 to picking station 18,agv_navigation
 Has AGV-12 reported any faults today?,system_status
-Give me stock levels across all zones for SKU-9997,inventory_check
+Give me stock levels across all zones for SKU-7202,inventory_check
 Is AMR-21 operational?,system_status
 A forklift near-miss was reported in Zone D,safety_incident
 "Belt Conveyor-21 stopped unexpectedly, please check",equipment_maintenance
 "Belt Sorter-06 stopped unexpectedly, please check",equipment_maintenance
 Report vibration issue on Conveyor-14,equipment_maintenance
 Reassign Crane-05 to charging station,agv_navigation
-Give me stock levels across all zones for SKU-7446,inventory_check
+Give me stock levels across all zones for SKU-7317,inventory_check
 Has AGV-18 reported any faults today?,system_status
-Track order #42487 for me,order_status
-When will order #96706 be delivered?,order_status
-Check inventory count for SKU-9124 in the receiving dock,inventory_check
+Track order #50655 for me,order_status
+When will order #53538 be delivered?,order_status
+Check inventory count for SKU-6466 in the receiving dock,inventory_check
 What's the most efficient picking sequence for the shipping dock today?,picking_optimization
 Log a breakdown for Crane-09 in Zone F,equipment_maintenance
 Has Sorter-06 reported any faults today?,system_status
@@ -203,34 +203,34 @@ What's the most efficient picking sequence for Zone B today?,picking_optimizatio
 Route Sorter-06 to picking station 3,agv_navigation
 Is the sorter in Zone D running normally?,system_status
 What is the current location of Sorter-02?,agv_navigation
-Has order #47066 shipped yet?,order_status
+Has order #69853 shipped yet?,order_status
 Please log a near-miss between a pedestrian and Crane-03 in Zone E,safety_incident
 Log a breakdown for Conveyor-21 in Zone C,equipment_maintenance
 How should we sequence orders to minimize walking distance in Zone C?,picking_optimization
 The conveyor belt in Zone B is making noise,equipment_maintenance
-What is the current stock level for SKU-7086?,inventory_check
-When will order #75230 be delivered?,order_status
+What is the current stock level for SKU-9826?,inventory_check
+When will order #89131 be delivered?,order_status
 File an incident report for Zone E,safety_incident
-Show the fulfillment status of #56129,order_status
+Show the fulfillment status of #85726,order_status
 Report unsafe pallet stacking in the receiving dock,safety_incident
 Crane Crane-09 reported a fault code,equipment_maintenance
 "A worker slipped near AMR-33, please log it",safety_incident
-When will order #71075 be delivered?,order_status
+When will order #38397 be delivered?,order_status
 Report vibration issue on Crane-03,equipment_maintenance
 Log a safety incident involving AMR-33,safety_incident
 Please log a near-miss between a pedestrian and AGV-12 in the shipping dock,safety_incident
 How can we reduce travel time for pickers in Zone A?,picking_optimization
-Why hasn't order #89556 left the dock yet?,order_status
+Why hasn't order #98736 left the dock yet?,order_status
 What's the most efficient picking sequence for Zone E today?,picking_optimization
-Why hasn't order #91208 left the dock yet?,order_status
-Is SKU-1347 in stock at Zone A?,inventory_check
-Is SKU-2055 in stock at Zone B?,inventory_check
+Why hasn't order #11185 left the dock yet?,order_status
+Is SKU-5595 in stock at Zone A?,inventory_check
+Is SKU-6953 in stock at Zone B?,inventory_check
 The conveyor belt in Zone A is making noise,equipment_maintenance
 Crane-09 motor temperature seems high,equipment_maintenance
 Suggest a wave picking plan for Zone D,picking_optimization
 Reassign Crane-03 to charging station,agv_navigation
 Send the next available AGV to picking station 12,agv_navigation
-Is order #40290 delayed?,order_status
+Is order #43653 delayed?,order_status
 Can you dispatch AGV-12 to the receiving dock right away?,agv_navigation
 Is Conveyor-14 operational?,system_status
 Would zone picking work better than batch picking for the shipping dock?,picking_optimization
@@ -241,10 +241,10 @@ What's the most efficient picking sequence for Zone F today?,picking_optimizatio
 Crane AGV-18 reported a fault code,equipment_maintenance
 AMR-21 motor temperature seems high,equipment_maintenance
 Please log a near-miss between a pedestrian and Crane-09 in the mezzanine,safety_incident
-What's the fastest picking route for order #24846?,picking_optimization
+What's the fastest picking route for order #98965?,picking_optimization
 Reassign AGV-12 to charging station,agv_navigation
 "Conveyor-21 seems to be idle near Zone B, please reroute it",agv_navigation
-When will order #12947 be delivered?,order_status
+When will order #36922 be delivered?,order_status
 Is Sorter-02 operational?,system_status
 Pause Crane-05 until the aisle in the receiving dock is clear,agv_navigation
 Crane Conveyor-21 reported a fault code,equipment_maintenance
@@ -257,94 +257,94 @@ Log a safety incident involving Conveyor-21,safety_incident
 Please log a near-miss between a pedestrian and Crane-09 in Zone A,safety_incident
 What's the current uptime percentage for the cross-dock area?,system_status
 There was a near collision between Crane-05 and a pedestrian in Zone F,safety_incident
-Has order #33247 shipped yet?,order_status
+Has order #87574 shipped yet?,order_status
 Log a breakdown for Sorter-02 in Zone B,equipment_maintenance
 Recommend a picking strategy for high-velocity SKUs,picking_optimization
-What's the fastest picking route for order #47595?,picking_optimization
+What's the fastest picking route for order #95679?,picking_optimization
 How should we sequence orders to minimize walking distance in Zone F?,picking_optimization
 What's the difference between AGV and AMR?,general_faq
 "A worker slipped near Crane-03, please log it",safety_incident
 "AGV-18 seems to be idle near Zone E, please reroute it",agv_navigation
-What's the fastest picking route for order #18933?,picking_optimization
+What's the fastest picking route for order #57683?,picking_optimization
 The conveyor belt in the mezzanine is making noise,equipment_maintenance
 "Belt AMR-21 stopped unexpectedly, please check",equipment_maintenance
 Redirect AMR-33 around the blocked aisle in Zone F,agv_navigation
 Would zone picking work better than batch picking for Zone A?,picking_optimization
 Why is Conveyor-14 stuck near the receiving dock?,agv_navigation
-Has order #61675 shipped yet?,order_status
+Has order #18536 shipped yet?,order_status
 Log a breakdown for Crane-05 in Zone F,equipment_maintenance
 Why is AGV-07 stuck near Zone B?,agv_navigation
 There was a near collision between Conveyor-21 and a pedestrian in Zone D,safety_incident
-What's the fastest picking route for order #82441?,picking_optimization
+What's the fastest picking route for order #71673?,picking_optimization
 "Belt AMR-33 stopped unexpectedly, please check",equipment_maintenance
 Is the sorter in Zone C running normally?,system_status
-Is order #64094 delayed?,order_status
+Is order #24576 delayed?,order_status
 Check system health for the shipping dock,system_status
 Crane Conveyor-14 reported a fault code,equipment_maintenance
 There was a near collision between AGV-12 and a pedestrian in Zone D,safety_incident
-Is SKU-2184 in stock at Zone E?,inventory_check
+Is SKU-3206 in stock at Zone E?,inventory_check
 Check system health for Zone D,system_status
-Has order #42767 shipped yet?,order_status
+Has order #43891 shipped yet?,order_status
 Would zone picking work better than batch picking for Zone D?,picking_optimization
-When will order #39467 be delivered?,order_status
+When will order #71729 be delivered?,order_status
 Suggest a wave picking plan for Zone C,picking_optimization
 Can you dispatch AGV-07 to Zone F right away?,agv_navigation
 The conveyor belt in the cross-dock area is making noise,equipment_maintenance
-How many units of SKU-7068 are in the receiving dock?,inventory_check
-Show me the on-hand quantity of SKU-4766,inventory_check
-Check inventory count for SKU-6772 in Zone D,inventory_check
+How many units of SKU-7671 are in the receiving dock?,inventory_check
+Show me the on-hand quantity of SKU-3702,inventory_check
+Check inventory count for SKU-6108 in Zone D,inventory_check
 Crane Crane-03 reported a fault code,equipment_maintenance
 Schedule maintenance for Crane-05,equipment_maintenance
 Report unsafe pallet stacking in Zone D,safety_incident
 "Someone bypassed the safety gate near Crane-09, please log this",safety_incident
 "AMR-21 seems to be idle near Zone B, please reroute it",agv_navigation
-Why hasn't order #88419 left the dock yet?,order_status
-Is SKU-4116 in stock at the receiving dock?,inventory_check
+Why hasn't order #86242 left the dock yet?,order_status
+Is SKU-8367 in stock at the receiving dock?,inventory_check
 Would zone picking work better than batch picking for the receiving dock?,picking_optimization
 What is a mini-load system?,general_faq
 "A spill was reported near Conveyor-14 in the receiving dock, needs cleanup",safety_incident
 Report vibration issue on Sorter-02,equipment_maintenance
-Do we have enough SKU-5929 to fulfill 200 units?,inventory_check
+Do we have enough SKU-8061 to fulfill 200 units?,inventory_check
 Why do warehouses use cycle counting instead of annual counts?,general_faq
 Route Crane-05 to picking station 12,agv_navigation
-What's the fastest picking route for order #49977?,picking_optimization
+What's the fastest picking route for order #75527?,picking_optimization
 Suggest a wave picking plan for the mezzanine,picking_optimization
 Is the WCS connection to AGV-07 stable?,system_status
-Do we have enough SKU-8543 to fulfill 200 units?,inventory_check
+Do we have enough SKU-5257 to fulfill 200 units?,inventory_check
 How should we sequence orders to minimize walking distance in the cross-dock area?,picking_optimization
 Is the sorter in the cross-dock area running normally?,system_status
-Track order #24822 for me,order_status
+Track order #48287 for me,order_status
 Pause Sorter-06 until the aisle in Zone E is clear,agv_navigation
 Check system health for Zone B,system_status
 Pause Crane-03 until the aisle in the cross-dock area is clear,agv_navigation
 The sorter in Zone F keeps jamming,equipment_maintenance
-What's the status of order #99212?,order_status
+What's the status of order #28439?,order_status
 Optimize the pick path for Zone F,picking_optimization
 Send AGV AMR-21 to Zone D,agv_navigation
-Show the fulfillment status of #35013,order_status
-Check inventory count for SKU-6321 in Zone C,inventory_check
-What is the current stock level for SKU-3159?,inventory_check
+Show the fulfillment status of #43413,order_status
+Check inventory count for SKU-9277 in Zone C,inventory_check
+What is the current stock level for SKU-3569?,inventory_check
 What's the most efficient picking sequence for the cross-dock area today?,picking_optimization
 "AMR-33 seems to be idle near Zone E, please reroute it",agv_navigation
 Has AMR-33 reported any faults today?,system_status
-Give me stock levels across all zones for SKU-4283,inventory_check
+Give me stock levels across all zones for SKU-8037,inventory_check
 What is the uptime for AMR-33 today?,system_status
 Route Crane-03 to picking station 18,agv_navigation
 How can we reduce travel time for pickers in Zone F?,picking_optimization
-What's the fastest picking route for order #25124?,picking_optimization
+What's the fastest picking route for order #96033?,picking_optimization
 Report vibration issue on AMR-33,equipment_maintenance
 Check system health for Zone E,system_status
 "Belt AGV-07 stopped unexpectedly, please check",equipment_maintenance
-Is order #66869 delayed?,order_status
+Is order #42584 delayed?,order_status
 The sorter in the shipping dock keeps jamming,equipment_maintenance
 Report vibration issue on Conveyor-21,equipment_maintenance
-Give me stock levels across all zones for SKU-3849,inventory_check
+Give me stock levels across all zones for SKU-3437,inventory_check
 "A spill was reported near AMR-21 in the cross-dock area, needs cleanup",safety_incident
-Show the fulfillment status of #51119,order_status
+Show the fulfillment status of #57239,order_status
 Please log a near-miss between a pedestrian and Conveyor-21 in the receiving dock,safety_incident
-Is SKU-3408 in stock at the cross-dock area?,inventory_check
+Is SKU-9655 in stock at the cross-dock area?,inventory_check
 "Belt Conveyor-14 stopped unexpectedly, please check",equipment_maintenance
-What is the current stock level for SKU-6111?,inventory_check
+What is the current stock level for SKU-5360?,inventory_check
 Are all cranes online in Zone B?,system_status
 "A worker slipped near AMR-21, please log it",safety_incident
 What is the current location of AGV-18?,agv_navigation
@@ -364,30 +364,30 @@ There was a near collision between AMR-33 and a pedestrian in Zone C,safety_inci
 Would zone picking work better than batch picking for Zone E?,picking_optimization
 Send the next available AGV to picking station 3,agv_navigation
 Crane-03 motor temperature seems high,equipment_maintenance
-Is SKU-5590 in stock at the receiving dock?,inventory_check
+Is SKU-6320 in stock at the receiving dock?,inventory_check
 Why is AGV-12 stuck near the shipping dock?,agv_navigation
 Send the next available AGV to picking station 5,agv_navigation
 Please log a near-miss between a pedestrian and AGV-18 in Zone C,safety_incident
 What's the most efficient picking sequence for Zone C today?,picking_optimization
-Track order #47588 for me,order_status
-Is SKU-4954 in stock at Zone A?,inventory_check
-Do we have enough SKU-6315 to fulfill 200 units?,inventory_check
+Track order #96525 for me,order_status
+Is SKU-3798 in stock at Zone A?,inventory_check
+Do we have enough SKU-4449 to fulfill 200 units?,inventory_check
 How does regenerative braking work on an AS/RS crane?,general_faq
-Give me stock levels across all zones for SKU-6820,inventory_check
-When will order #60909 be delivered?,order_status
+Give me stock levels across all zones for SKU-1650,inventory_check
+When will order #90343 be delivered?,order_status
 Why is AGV-12 stuck near the receiving dock?,agv_navigation
-Give me stock levels across all zones for SKU-8544,inventory_check
+Give me stock levels across all zones for SKU-8956,inventory_check
 Optimize the pick path for Zone D,picking_optimization
 Give me the current status of the WMS integration,system_status
-When will order #11543 be delivered?,order_status
+When will order #89814 be delivered?,order_status
 A forklift near-miss was reported in Zone F,safety_incident
-What is the current stock level for SKU-3942?,inventory_check
+What is the current stock level for SKU-6339?,inventory_check
 "Someone bypassed the safety gate near AMR-21, please log this",safety_incident
 How can we reduce travel time for pickers in Zone D?,picking_optimization
 How can we reduce travel time for pickers in the shipping dock?,picking_optimization
-Give me stock levels across all zones for SKU-9140,inventory_check
+Give me stock levels across all zones for SKU-3877,inventory_check
 "A spill was reported near AGV-12 in Zone F, needs cleanup",safety_incident
-Track order #87663 for me,order_status
+Track order #95296 for me,order_status
 Are all cranes online in Zone E?,system_status
 Would zone picking work better than batch picking for Zone B?,picking_optimization
 The sorter in the mezzanine keeps jamming,equipment_maintenance
@@ -397,24 +397,24 @@ What is a WMS?,general_faq
 A forklift near-miss was reported in Zone A,safety_incident
 What is predictive maintenance?,general_faq
 There was a near collision between Conveyor-14 and a pedestrian in Zone B,safety_incident
-What's the fastest picking route for order #80966?,picking_optimization
-Do we have enough SKU-2549 to fulfill 200 units?,inventory_check
+What's the fastest picking route for order #27906?,picking_optimization
+Do we have enough SKU-5448 to fulfill 200 units?,inventory_check
 Are all cranes online in the mezzanine?,system_status
-Do we have enough SKU-2599 to fulfill 200 units?,inventory_check
+Do we have enough SKU-4957 to fulfill 200 units?,inventory_check
 Crane AMR-33 reported a fault code,equipment_maintenance
 Pause Conveyor-14 until the aisle in Zone E is clear,agv_navigation
 What's the difference between a stacker crane and a shuttle?,general_faq
 Redirect Sorter-06 around the blocked aisle in Zone C,agv_navigation
 Conveyor-21 motor temperature seems high,equipment_maintenance
-Do we have enough SKU-9855 to fulfill 200 units?,inventory_check
-Check inventory count for SKU-6154 in Zone E,inventory_check
+Do we have enough SKU-2245 to fulfill 200 units?,inventory_check
+Check inventory count for SKU-3735 in Zone E,inventory_check
 Report unsafe pallet stacking in Zone C,safety_incident
 Redirect Crane-09 around the blocked aisle in the mezzanine,agv_navigation
 How should we sequence orders to minimize walking distance in the mezzanine?,picking_optimization
-Do we have enough SKU-3311 to fulfill 200 units?,inventory_check
+Do we have enough SKU-5494 to fulfill 200 units?,inventory_check
 Please log a near-miss between a pedestrian and Crane-05 in the mezzanine,safety_incident
 The conveyor belt in Zone F is making noise,equipment_maintenance
-What's the fastest picking route for order #57021?,picking_optimization
+What's the fastest picking route for order #86847?,picking_optimization
 Can you dispatch Conveyor-14 to the receiving dock right away?,agv_navigation
 What's the current uptime percentage for the receiving dock?,system_status
 Optimize the pick path for Zone B,picking_optimization
@@ -426,18 +426,18 @@ Log a breakdown for Conveyor-14 in the shipping dock,equipment_maintenance
 The conveyor belt in the shipping dock is making noise,equipment_maintenance
 Log a breakdown for Crane-05 in the receiving dock,equipment_maintenance
 There was a near collision between Sorter-06 and a pedestrian in Zone E,safety_incident
-What's the status of order #17664?,order_status
+What's the status of order #45231?,order_status
 What's the current uptime percentage for Zone B?,system_status
-Track order #90808 for me,order_status
+Track order #15877 for me,order_status
 "A spill was reported near AGV-07 in the receiving dock, needs cleanup",safety_incident
 Pause AGV-07 until the aisle in the cross-dock area is clear,agv_navigation
 AGV-07 motor temperature seems high,equipment_maintenance
 What KPIs matter most in warehouse automation?,general_faq
-What is the current stock level for SKU-8037?,inventory_check
+What is the current stock level for SKU-1934?,inventory_check
 Redirect Crane-09 around the blocked aisle in Zone E,agv_navigation
-Is order #79026 delayed?,order_status
-Has order #22778 shipped yet?,order_status
-What's the fastest picking route for order #78187?,picking_optimization
+Is order #66488 delayed?,order_status
+Has order #71734 shipped yet?,order_status
+What's the fastest picking route for order #78872?,picking_optimization
 There was a near collision between AGV-12 and a pedestrian in Zone C,safety_incident
 What is a pick rate and how is it measured?,general_faq
-Why hasn't order #21744 left the dock yet?,order_status
+Why hasn't order #89697 left the dock yet?,order_status

data/latency_eval.json

@@ -1,6 +1,6 @@
 {
-  "intent_classifier_ms": 0.419,
-  "anomaly_detector_ms": 23.129,
-  "kb_retrieval_ms": 0.571,
+  "intent_classifier_ms": 0.633,
+  "anomaly_detector_ms": 14.676,
+  "kb_retrieval_ms": 0.862,
   "note": "LLM generation latency depends on the external Inference API call and is measured live in the app, not benchmarked here."
 }

models/anomaly_iforest.joblib

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9dc690ec59c1639e4bcdc6d85f3ae61e9e932ae935b250bab7dd5de7654d62fa
-size 2149241
+oid sha256:1f3c085015adb1b4059275537b8f550cb569a97bd35e2020978083fb75873f57
+size 2166973

models/anomaly_scaler.joblib

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6eedcb449295a4276c34a0e08175010fd3b6958e8fc7158658461a9820011ff9
+oid sha256:5e2a15be4a9a58c2d584e76e2b2bc701f213afc469768faca1eeac44d962dede
 size 679

models/anomaly_threshold.joblib

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3f8523d48d478c6c634b0ce0bad9d97241a473d1e89eb8eb689bd2027c4b157
+size 21

models/intent_pipeline.joblib

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:876f27ef5fe7c57beb4955f1666e618a8e20e5b8a1c38a7e6728459f74a9d4f4
+oid sha256:0c7a9a52d4512974c7c0dee05f15871cf0d3346b3a0e1a2ff7d7fb36e7a722f9
 size 86940

src/anomaly_model.py

@@ -7,9 +7,25 @@ powers the "Predictive Maintenance" tab -- flags abnormal equipment
 behaviour before it causes an unplanned stoppage, which is exactly the kind
 of workload modern intralogistics platforms (e.g. AS/RS, sorters, AGVs)
 generate continuously in production.
+
+Threshold calibration
+----------------------
+The model itself is fit unsupervised (IsolationForest never sees labels).
+However, converting its continuous anomaly score into a binary decision
+requires a threshold, and scikit-learn's default choices for this (a fixed
+"contamination" rate) are either a blind guess or -- if set to the true
+label rate -- a form of leakage (see build_artifacts.py for the full
+writeup of that bug and fix). The approach used here instead is standard
+practice for production anomaly detection: fit the model unsupervised on
+unlabelled data, then calibrate the decision threshold using a small
+*labelled calibration set* (e.g. a handful of confirmed historical
+incidents), completely separate from both model fitting and the final
+held-out test set used to report metrics. `calibrated_threshold` below is
+that calibrated cutoff, chosen to maximise F1 on the calibration set only.
 """
 
 from dataclasses import dataclass
+from typing import Optional
 
 import joblib
 import numpy as np
@@ -26,7 +42,7 @@ class AnomalyResult:
     raw_score: float
 
 
-def build_model(contamination: float = 0.1, seed: int = 42) -> IsolationForest:
+def build_model(contamination="auto", seed: int = 42) -> IsolationForest:
     return IsolationForest(
         n_estimators=200,
         contamination=contamination,
@@ -34,20 +50,47 @@ def build_model(contamination: float = 0.1, seed: int = 42) -> IsolationForest:
     )
 
 
-def score_reading(model: IsolationForest, scaler: StandardScaler, reading: dict) -> AnomalyResult:
+def score_reading(
+    model: IsolationForest,
+    scaler: StandardScaler,
+    reading: dict,
+    threshold: Optional[float] = None,
+) -> AnomalyResult:
+    """
+    Score a single sensor reading.
+
+    If `threshold` is given, it's compared against the 0-1 anomaly_score
+    (this is the calibrated-threshold path used by the deployed app). If
+    omitted, falls back to the IsolationForest's own .predict() (its
+    internal contamination-derived offset).
+    """
     x = np.array([[reading[f] for f in FEATURES]])
     x_scaled = scaler.transform(x)
     raw = model.decision_function(x_scaled)[0]  # higher = more normal
-    pred = model.predict(x_scaled)[0]  # 1 = normal, -1 = anomaly
     # squash raw decision_function (~[-0.5, 0.5]) into a 0-1 "anomaly score"
     anomaly_score = float(np.clip(0.5 - raw, 0, 1))
-    return AnomalyResult(is_anomaly=(pred == -1), anomaly_score=anomaly_score, raw_score=float(raw))
+    if threshold is not None:
+        is_anomaly = anomaly_score >= threshold
+    else:
+        pred = model.predict(x_scaled)[0]  # 1 = normal, -1 = anomaly
+        is_anomaly = (pred == -1)
+    return AnomalyResult(is_anomaly=bool(is_anomaly), anomaly_score=anomaly_score, raw_score=float(raw))
 
 
-def save_artifacts(model, scaler, model_path: str, scaler_path: str):
+def save_artifacts(model, scaler, model_path: str, scaler_path: str, threshold: Optional[float] = None, threshold_path: Optional[str] = None):
     joblib.dump(model, model_path)
     joblib.dump(scaler, scaler_path)
+    if threshold is not None and threshold_path is not None:
+        joblib.dump(float(threshold), threshold_path)
 
 
-def load_artifacts(model_path: str, scaler_path: str):
-    return joblib.load(model_path), joblib.load(scaler_path)
+def load_artifacts(model_path: str, scaler_path: str, threshold_path: Optional[str] = None):
+    model = joblib.load(model_path)
+    scaler = joblib.load(scaler_path)
+    if threshold_path is not None:
+        try:
+            threshold = joblib.load(threshold_path)
+            return model, scaler, threshold
+        except FileNotFoundError:
+            return model, scaler, None
+    return model, scaler