Upload 34 files

README.md

@@ -27,16 +27,23 @@ metrics on held-out test data.
 
 1. **💬 AI Assistant** — ask free-text warehouse-ops questions; answers are
    grounded via TF-IDF retrieval over a small knowledge base and generated
-   by a hosted LLM (Hugging Face Inference API), with a transparent
-   retrieval-only fallback if no API key is configured.
+   by a hosted LLM (Hugging Face Inference API, with a multi-model fallback
+   chain), with a transparent retrieval-only fallback and a built-in
+   **"Test LLM connection" diagnostics button** if no API key is configured
+   or the call fails. Includes an "About the data" panel explaining the
+   knowledge base and training data.
 2. **📦 Inventory & Order Query** — natural-language queries over synthetic
-   inventory / order tables (SKU, zone, order-id extraction).
+   inventory / order tables (SKU, zone, order-id extraction), with an
+   "About the data" panel describing the synthetic tables.
 3. **⚠️ Predictive Maintenance** — Isolation Forest anomaly detector over
    conveyor/crane motor sensor readings (temperature, vibration, current,
-   belt speed).
+   belt speed), with an "About the data" panel describing the synthetic
+   sensor dataset and failure patterns.
 4. **📊 Model Evaluation** — accuracy, macro-F1, confusion matrices,
    ROC-AUC, retrieval hit-rate, and latency benchmarks, all computed on
-   held-out data by `build_artifacts.py`.
+   held-out data by `build_artifacts.py` and rendered as charts (bar
+   charts, confusion matrices, ROC curve, feature-distribution histograms)
+   alongside the underlying tables.
 5. **ℹ️ About** — project write-up, architecture diagram, tech stack.
 
 ## Quick start (local)

app.py

@@ -30,7 +30,7 @@ from src.anomaly_model import score_reading
 from src.data_generation import generate_inventory_db, generate_orders_db
 from src.intent_model import INTENT_DESCRIPTIONS, load_pipeline, predict as intent_predict
 from src.inventory_db import query_inventory, query_orders
-from src.llm_client import answer_query
+from src.llm_client import answer_query, test_connection
 from src.retriever import KBRetriever
 
 ROOT = os.path.dirname(os.path.abspath(__file__))
@@ -146,11 +146,33 @@ def chat_fn(message, history):
         f"**Generation:** {'LLM (' + response.model_id + ')' if response.used_llm else 'retrieval-only fallback'}"
         f" · {response.latency_s * 1000:.0f} ms"
     )
+    if not response.used_llm and response.debug_errors:
+        err_lines = "\n".join(f"  - `{e}`" for e in response.debug_errors)
+        meta_lines.append(f"**Why the LLM wasn't used:**\n{err_lines}")
 
     full_reply = response.answer + "\n\n---\n" + "\n".join(meta_lines)
     return full_reply
 
 
+def test_llm_fn():
+    response = test_connection(retriever)
+    if response.used_llm:
+        return (
+            f"✅ **LLM connection working.** Model: `{response.model_id}` · "
+            f"{response.latency_s * 1000:.0f} ms\n\nSample answer: {response.answer}"
+        )
+    err_lines = "\n".join(f"- `{e}`" for e in response.debug_errors) or "(no error detail captured)"
+    return (
+        "❌ **LLM connection failed** — running in retrieval-only fallback mode.\n\n"
+        f"**Errors from each candidate model tried:**\n{err_lines}\n\n"
+        "**Common causes:** missing/invalid `HF_TOKEN` secret, the token's account "
+        "lacking Inference API access, or the candidate models being temporarily "
+        "unavailable on HF's free serverless tier. See `src/llm_client.py` to add "
+        "or reorder candidate models, or set the `LLM_MODEL_ID` secret to force a "
+        "specific one."
+    )
+
+
 ASSISTANT_EXAMPLES = [
     "The conveyor belt in Zone C is making noise",
     "How many units of SKU-1042 are in Zone B?",
@@ -226,7 +248,7 @@ def load_preset(name):
 # TAB 4 -- Model Evaluation
 # ==========================================================================
 
-def build_evaluation_markdown():
+def eval_intent_section_md():
     cls_report = intent_eval.get("classification_report", {})
     per_class_rows = []
     for cls in intent_eval.get("classes", []):
@@ -236,19 +258,22 @@ def build_evaluation_markdown():
             f"{stats.get('f1-score', 0):.2f} | {int(stats.get('support', 0))} |"
         )
     per_class_table = "\n".join(per_class_rows)
+    return f"""
+## 1. Intent Classifier (TF-IDF + Logistic Regression)
 
-    retrieval_rows = "\n".join(
-        f"| {r['query']} | {r['expected']} | {r['retrieved_top1']} | "
-        f"{'✅' if r['hit@1'] else ('〰️' if r['hit@2'] else '❌')} | {r['top1_score']:.2f} |"
-        for r in retrieval_eval.get("rows", [])
-    )
+**What it does:** routes a free-text query (e.g. *"The conveyor belt in Zone C
+is making noise"*) into one of 8 operational categories, used by both the
+AI Assistant and Inventory & Order Query tabs to decide how to handle a
+request.
 
-    md = f"""
-## 1. Intent Classifier (TF-IDF + Logistic Regression)
+**Training data:** {intent_eval.get('n_train', '?') } synthetically generated
+example queries (see chart below), built from ~8 hand-written templates per
+category with randomised SKU codes, zone names, order IDs, and equipment IDs
+slotted in -- e.g. *"How many units of {{sku}} are in {{zone}}?"*. This keeps
+the language varied while being fully reproducible (`src/data_generation.py`).
 
-Trained on {intent_eval.get('n_train', '?')} examples, evaluated on a held-out
-stratified test split of {intent_eval.get('n_test', '?')} examples across
-{intent_eval.get('n_classes', '?')} intent classes.
+Evaluated on a **held-out stratified test split** of {intent_eval.get('n_test', '?')}
+examples the model never saw during training.
 
 | Metric | Score |
 |---|---|
@@ -260,16 +285,28 @@ stratified test split of {intent_eval.get('n_test', '?')} examples across
 | Intent | Precision | Recall | F1 | Support |
 |---|---|---|---|---|
 {per_class_table}
+"""
 
-![Confusion Matrix](assets/intent_confusion_matrix.png)
-
----
 
+def eval_anomaly_section_md():
+    return f"""
 ## 2. Predictive Maintenance Anomaly Detector (Isolation Forest)
 
-Trained unsupervised on scaled sensor features ({', '.join(SENSOR_FEATURES)}),
-evaluated against held-out ground-truth anomaly labels ({anomaly_eval.get('n_test', '?')} test readings,
-{anomaly_eval.get('test_anomaly_rate', 0):.1%} true anomaly rate).
+**What it does:** flags abnormal conveyor/crane motor sensor readings
+(temperature, vibration, current, belt speed) before they cause an
+unplanned stoppage -- powers the Predictive Maintenance tab.
+
+**Training data:** {900 + 100} synthetic sensor readings (900 "normal"
++ 100 "anomaly" patterns), each with 4 features. Normal readings are drawn
+from realistic operating ranges (e.g. ~55°C motor temp, ~2.2 mm/s vibration);
+anomalies simulate bearing wear / misalignment / overload (elevated temp,
+vibration, and current with reduced belt speed). See the distribution chart
+below for exactly how these two classes differ.
+
+The model itself is trained **unsupervised** (Isolation Forest never sees
+the anomaly label during fitting) -- labels are used only to *evaluate* it
+afterward, on a held-out test split of {anomaly_eval.get('n_test', '?')}
+readings ({anomaly_eval.get('test_anomaly_rate', 0):.1%} true anomaly rate).
 
 | Metric | Score |
 |---|---|
@@ -278,17 +315,27 @@ evaluated against held-out ground-truth anomaly labels ({anomaly_eval.get('n_tes
 | **F1 Score** | **{anomaly_eval.get('f1', 0):.2%}** |
 | **ROC-AUC** | **{anomaly_eval.get('roc_auc', 0):.3f}** |
 | Accuracy | {anomaly_eval.get('accuracy', 0):.2%} |
+"""
 
-![Anomaly Confusion Matrix](assets/anomaly_confusion_matrix.png)
-![ROC Curve](assets/anomaly_roc_curve.png)
-
----
 
+def eval_retrieval_section_md():
+    retrieval_rows = "\n".join(
+        f"| {r['query']} | {r['expected']} | {r['retrieved_top1']} | "
+        f"{'✅' if r['hit@1'] else ('〰️' if r['hit@2'] else '❌')} | {r['top1_score']:.2f} |"
+        for r in retrieval_eval.get("rows", [])
+    )
+    return f"""
 ## 3. Retrieval (RAG) Evaluation
 
-Hit-rate of the TF-IDF retriever against a hand-labelled query -> expected-document
-evaluation set (higher is better; hit@1 = correct doc ranked first, hit@2 = correct
-doc within top 2).
+**What it does:** before the LLM answers a question, this component finds
+the most relevant passages from a 10-article warehouse-operations knowledge
+base (AS/RS, AGV/AMR, WMS, sortation, picking strategy, safety, etc. -- see
+`src/knowledge_base.py`) using TF-IDF + cosine similarity, so the LLM answers
+from real context rather than guessing.
+
+**Evaluation data:** {retrieval_eval.get('n_queries', '?')} hand-labelled
+(query, expected-article) pairs -- a small ground-truth set built by hand to
+check the retriever finds the *right* article, not just *an* article.
 
 | Metric | Score |
 |---|---|
@@ -298,20 +345,25 @@ doc within top 2).
 | Query | Expected Doc | Retrieved (top-1) | Hit | Score |
 |---|---|---|---|---|
 {retrieval_rows}
+"""
 
----
 
+def eval_latency_section_md():
+    return f"""
 ## 4. Latency Benchmark (per-request, CPU)
 
+Average of 50 runs each, measured on the same CPU hardware the Space runs on.
+
 | Component | Avg. latency |
 |---|---|
 | Intent classification | {latency_eval.get('intent_classifier_ms', '?')} ms |
 | Anomaly scoring | {latency_eval.get('anomaly_detector_ms', '?')} ms |
 | KB retrieval (TF-IDF) | {latency_eval.get('kb_retrieval_ms', '?')} ms |
-| LLM generation | Depends on hosted Inference API (measured live per-request in the Assistant tab) |
+| LLM generation | Depends on the hosted Inference API (measured live per-request in the Assistant tab, not benchmarked here) |
+"""
 
----
 
+EVAL_METHODOLOGY_MD = """
 ### Evaluation methodology notes
 
 - All datasets are **synthetically generated** (see `src/data_generation.py`) using
@@ -326,7 +378,8 @@ doc within top 2).
 - In a production deployment, all three components would be continuously
   re-evaluated against real WMS/WCS/sensor logs rather than synthetic data.
 """
-    return md
+
+
 
 
 # ==========================================================================
@@ -452,6 +505,20 @@ with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue"), css=CUSTOM_CSS, title="
             "AGV routing, picking strategy, or general warehouse-automation concepts. "
             "Answers are grounded (RAG) in a small warehouse-operations knowledge base."
         )
+        with gr.Accordion("ℹ️ About the data behind this tab", open=False):
+            gr.Markdown(
+                "- **Knowledge base (RAG source):** 10 original, hand-written articles "
+                "covering AS/RS, AGV/AMR, WMS, conveyor/sortation, picking strategy, "
+                "predictive maintenance, safety protocol, inventory accuracy, KPIs, and "
+                "energy efficiency (`src/knowledge_base.py`). Every answer's *Retrieved "
+                "context* line shows exactly which article(s) it drew on.\n"
+                "- **Intent classifier:** trained on ~480 synthetically generated example "
+                "queries across 8 categories (see the Model Evaluation tab for accuracy).\n"
+                "- **LLM:** a hosted instruct model called via the Hugging Face Inference "
+                "API — not run locally. If no `HF_TOKEN` is configured, or the API call "
+                "fails, this tab automatically falls back to showing the retrieved "
+                "knowledge-base passages directly, so it never just breaks."
+            )
         gr.ChatInterface(
             fn=chat_fn,
             type="messages",
@@ -459,6 +526,15 @@ with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue"), css=CUSTOM_CSS, title="
             chatbot=gr.Chatbot(height=430, label="Warehouse Assistant", type="messages"),
             textbox=gr.Textbox(placeholder="e.g. The conveyor belt in Zone C is making noise"),
         )
+        with gr.Accordion("🔧 LLM connection diagnostics", open=False):
+            gr.Markdown(
+                "If the assistant keeps answering in retrieval-only fallback mode, "
+                "click below to test the LLM connection directly and see the exact "
+                "error from each candidate model."
+            )
+            test_llm_btn = gr.Button("Test LLM connection")
+            test_llm_output = gr.Markdown()
+            test_llm_btn.click(test_llm_fn, inputs=None, outputs=test_llm_output)
 
     with gr.Tab("📦 Inventory & Order Query"):
         gr.Markdown(
@@ -466,6 +542,22 @@ with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue"), css=CUSTOM_CSS, title="
             "decides whether to query the inventory table or the orders table, then "
             "extracts SKU / order-id / zone slots to filter the result."
         )
+        with gr.Accordion("ℹ️ About the data behind this tab", open=False):
+            gr.Markdown(
+                f"- **Inventory table:** {len(inventory_df)} synthetic SKUs across 5 "
+                "categories (Electronics, Apparel, Automotive Parts, Food & Beverage, "
+                "Household) and 4 warehouse zones, with randomised on-hand quantities, "
+                "reorder points, and unit costs.\n"
+                f"- **Orders table:** {len(orders_df)} synthetic orders with randomised "
+                "status (Received / Picking / Packed / Shipped / Delayed), line count, "
+                "priority, and zone.\n"
+                "- Both tables are generated by `src/data_generation.py` with a fixed "
+                "random seed, so they're reproducible but **not real operational data** "
+                "-- this is a stand-in for a live WMS/WCS query interface.\n"
+                "- Query parsing is regex-based slot extraction (SKU codes like `SKU-1042`, "
+                "order IDs like `#10007`, zone names) combined with the same intent "
+                "classifier used in the AI Assistant tab (`src/inventory_db.py`)."
+            )
         with gr.Row():
             inv_input = gr.Textbox(label="Query", placeholder="How many units of SKU-1042 are in Zone B?", scale=4)
             inv_btn = gr.Button("Search", variant="primary", scale=1)
@@ -495,6 +587,21 @@ with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue"), css=CUSTOM_CSS, title="
             "for anomalous behaviour using an Isolation Forest model trained on "
             "historical sensor patterns."
         )
+        with gr.Accordion("ℹ️ About the data behind this tab", open=False):
+            gr.Markdown(
+                "- **Training data:** 1,000 synthetic sensor readings (900 normal + 100 "
+                "anomalous) across 4 features -- motor temperature, vibration, motor "
+                "current, and belt speed -- generated by `src/data_generation.py`. "
+                "Anomalies simulate realistic failure signatures: elevated temperature, "
+                "vibration, and current combined with reduced/erratic belt speed (the "
+                "pattern of bearing wear, belt misalignment, or motor overload).\n"
+                "- **Model:** Isolation Forest (unsupervised) trained on scaled features "
+                "-- it never sees a 'this is an anomaly' label during training, only "
+                "learns what 'normal' looks like and flags deviations from it.\n"
+                "- See the **Model Evaluation** tab for the feature-distribution chart "
+                "showing exactly how normal vs. anomalous readings differ, plus "
+                "precision/recall/F1/ROC-AUC on held-out data."
+            )
         preset_dropdown = gr.Dropdown(
             choices=list(ANOMALY_PRESETS.keys()), label="Load a preset reading", value="Normal reading"
         )
@@ -518,10 +625,48 @@ with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue"), css=CUSTOM_CSS, title="
         )
 
     with gr.Tab("📊 Model Evaluation"):
-        gr.Markdown(build_evaluation_markdown())
+        gr.Markdown(
+            "Every metric on this tab is computed on **held-out test data** by "
+            "`build_artifacts.py` (not cherry-picked from a live demo run) -- "
+            "re-run that script any time to reproduce these numbers from scratch."
+        )
+
+        gr.Markdown(eval_intent_section_md())
+        gr.Markdown("**Dataset composition** (how many training examples per intent):")
+        gr.Image(os.path.join(ASSETS_DIR, "intent_dataset_composition.png"), show_label=False, container=False)
+        gr.Markdown("**Per-class precision / recall / F1:**")
+        gr.Image(os.path.join(ASSETS_DIR, "intent_per_class_bar.png"), show_label=False, container=False)
+        gr.Markdown("**Confusion matrix:**")
+        gr.Image(os.path.join(ASSETS_DIR, "intent_confusion_matrix.png"), show_label=False, container=False)
+
+        gr.Markdown("---")
+        gr.Markdown(eval_anomaly_section_md())
+        gr.Markdown("**Sensor feature distributions (normal vs. anomaly):**")
+        gr.Image(os.path.join(ASSETS_DIR, "sensor_distributions.png"), show_label=False, container=False)
+        with gr.Row():
+            with gr.Column():
+                gr.Markdown("**Evaluation metrics:**")
+                gr.Image(os.path.join(ASSETS_DIR, "anomaly_metrics_bar.png"), show_label=False, container=False)
+            with gr.Column():
+                gr.Markdown("**Confusion matrix:**")
+                gr.Image(os.path.join(ASSETS_DIR, "anomaly_confusion_matrix.png"), show_label=False, container=False)
+        gr.Markdown("**ROC curve:**")
+        gr.Image(os.path.join(ASSETS_DIR, "anomaly_roc_curve.png"), show_label=False, container=False)
+
+        gr.Markdown("---")
+        gr.Markdown(eval_retrieval_section_md())
+        gr.Image(os.path.join(ASSETS_DIR, "retrieval_hitrate_bar.png"), show_label=False, container=False)
+
+        gr.Markdown("---")
+        gr.Markdown(eval_latency_section_md())
+        gr.Image(os.path.join(ASSETS_DIR, "latency_bar.png"), show_label=False, container=False)
+
+        gr.Markdown("---")
+        gr.Markdown(EVAL_METHODOLOGY_MD)
 
     with gr.Tab("ℹ️ About"):
         gr.Markdown(ABOUT_MD)
 
+
 if __name__ == "__main__":
     demo.launch(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)))

build_artifacts.py

@@ -98,6 +98,27 @@ def build_intent_classifier():
     pipeline_full.fit(df["text"], df["intent"])
     save_pipeline(pipeline_full, os.path.join(MODELS_DIR, "intent_pipeline.joblib"))
 
+    # Per-class precision/recall/F1 bar chart (clearer at a glance than the table alone)
+    fig, ax = plt.subplots(figsize=(9, 5))
+    x = np.arange(len(labels))
+    width = 0.25
+    precisions = [report[l]["precision"] for l in labels]
+    recalls = [report[l]["recall"] for l in labels]
+    f1s = [report[l]["f1-score"] for l in labels]
+    ax.bar(x - width, precisions, width, label="Precision", color="#3b82f6")
+    ax.bar(x, recalls, width, label="Recall", color="#10b981")
+    ax.bar(x + width, f1s, width, label="F1", color="#f59e0b")
+    ax.set_xticks(x)
+    ax.set_xticklabels(labels, rotation=35, ha="right", fontsize=8)
+    ax.set_ylim(0, 1.15)
+    ax.set_ylabel("Score")
+    ax.set_title("Intent Classifier: Per-Class Precision / Recall / F1")
+    ax.legend(loc="lower right", ncol=3)
+    ax.grid(axis="y", alpha=0.3)
+    fig.tight_layout()
+    fig.savefig(os.path.join(ASSETS_DIR, "intent_per_class_bar.png"), dpi=150)
+    plt.close(fig)
+
     metrics = {
         "accuracy": acc,
         "macro_f1": macro_f1,
@@ -109,6 +130,18 @@ def build_intent_classifier():
     }
     with open(os.path.join(DATA_DIR, "intent_eval.json"), "w") as f:
         json.dump(metrics, f, indent=2)
+
+    # Dataset composition chart (helps a reader understand what the model was trained on)
+    counts = df["intent"].value_counts().reindex(labels)
+    fig, ax = plt.subplots(figsize=(8, 4.5))
+    ax.barh(labels, counts.values, color="#6366f1")
+    ax.set_xlabel("Number of examples")
+    ax.set_title(f"Intent Dataset Composition (n={len(df)}, synthetic, templated)")
+    ax.grid(axis="x", alpha=0.3)
+    fig.tight_layout()
+    fig.savefig(os.path.join(ASSETS_DIR, "intent_dataset_composition.png"), dpi=150)
+    plt.close(fig)
+
     return metrics
 
 
@@ -204,6 +237,45 @@ def build_anomaly_detector():
     }
     with open(os.path.join(DATA_DIR, "anomaly_eval.json"), "w") as f:
         json.dump(metrics, f, indent=2)
+
+    # Metrics bar chart
+    fig, ax = plt.subplots(figsize=(6.5, 4.5))
+    metric_names = ["Precision", "Recall", "F1", "ROC-AUC", "Accuracy"]
+    metric_vals = [precision, recall, f1, roc_auc, acc]
+    bars = ax.bar(metric_names, metric_vals, color=["#3b82f6", "#10b981", "#f59e0b", "#8b5cf6", "#ef4444"])
+    ax.set_ylim(0, 1.15)
+    ax.set_ylabel("Score")
+    ax.set_title("Anomaly Detector: Evaluation Metrics")
+    ax.grid(axis="y", alpha=0.3)
+    for bar, val in zip(bars, metric_vals):
+        ax.text(bar.get_x() + bar.get_width() / 2, val + 0.03, f"{val:.2f}", ha="center", fontsize=9)
+    fig.tight_layout()
+    fig.savefig(os.path.join(ASSETS_DIR, "anomaly_metrics_bar.png"), dpi=150)
+    plt.close(fig)
+
+    # Sensor feature distributions: normal vs anomaly (helps a reader see *why*
+    # the model flags what it flags -- directly supports the Predictive
+    # Maintenance tab's sliders)
+    fig, axes = plt.subplots(2, 2, figsize=(10, 7))
+    titles = {
+        "motor_temp_c": "Motor Temperature (°C)",
+        "vibration_mm_s": "Vibration (mm/s)",
+        "current_amps": "Motor Current (A)",
+        "belt_speed_mps": "Belt Speed (m/s)",
+    }
+    for ax, feat in zip(axes.flat, FEATURES):
+        normal_vals = df.loc[df["label"] == 0, feat]
+        anomaly_vals = df.loc[df["label"] == 1, feat]
+        ax.hist(normal_vals, bins=25, alpha=0.6, label="Normal", color="#10b981")
+        ax.hist(anomaly_vals, bins=25, alpha=0.6, label="Anomaly", color="#ef4444")
+        ax.set_title(titles[feat], fontsize=10)
+        ax.legend(fontsize=8)
+        ax.grid(alpha=0.3)
+    fig.suptitle("Sensor Feature Distributions: Normal vs. Anomaly (synthetic training data)", fontsize=11)
+    fig.tight_layout()
+    fig.savefig(os.path.join(ASSETS_DIR, "sensor_distributions.png"), dpi=150)
+    plt.close(fig)
+
     return metrics
 
 
@@ -238,6 +310,21 @@ def build_retrieval_eval():
     print(f"hit@1={metrics['hit_rate_at_1']:.2f}  hit@2={metrics['hit_rate_at_2']:.2f}")
     with open(os.path.join(DATA_DIR, "retrieval_eval.json"), "w") as f:
         json.dump(metrics, f, indent=2)
+
+    fig, ax = plt.subplots(figsize=(4.5, 4))
+    bars = ax.bar(["Hit Rate @ 1", "Hit Rate @ 2"],
+                   [metrics["hit_rate_at_1"], metrics["hit_rate_at_2"]],
+                   color=["#3b82f6", "#10b981"])
+    ax.set_ylim(0, 1.15)
+    ax.set_ylabel("Hit rate")
+    ax.set_title(f"RAG Retriever Hit Rate (n={n} labelled queries)")
+    ax.grid(axis="y", alpha=0.3)
+    for bar, val in zip(bars, [metrics["hit_rate_at_1"], metrics["hit_rate_at_2"]]):
+        ax.text(bar.get_x() + bar.get_width() / 2, val + 0.03, f"{val:.0%}", ha="center", fontsize=10)
+    fig.tight_layout()
+    fig.savefig(os.path.join(ASSETS_DIR, "retrieval_hitrate_bar.png"), dpi=150)
+    plt.close(fig)
+
     return metrics
 
 
@@ -287,6 +374,19 @@ def build_latency_benchmark(intent_metrics, anomaly_metrics):
         json.dump(latency, f, indent=2)
     print(latency)
 
+    fig, ax = plt.subplots(figsize=(6, 4))
+    components = ["Intent\nclassifier", "Anomaly\ndetector", "KB\nretrieval"]
+    values = [intent_ms, anomaly_ms, retrieval_ms]
+    bars = ax.bar(components, values, color=["#3b82f6", "#f59e0b", "#10b981"])
+    ax.set_ylabel("Latency (ms, avg of 50 runs)")
+    ax.set_title("Local Component Latency (CPU)")
+    ax.grid(axis="y", alpha=0.3)
+    for bar, val in zip(bars, values):
+        ax.text(bar.get_x() + bar.get_width() / 2, val, f"{val:.2f} ms", ha="center", va="bottom", fontsize=9)
+    fig.tight_layout()
+    fig.savefig(os.path.join(ASSETS_DIR, "latency_bar.png"), dpi=150)
+    plt.close(fig)
+
 
 if __name__ == "__main__":
     intent_metrics = build_intent_classifier()

data/intent_dataset.csv

@@ -1,5 +1,5 @@
 text,intent
-Is order #37898 delayed?,order_status
+Is order #69013 delayed?,order_status
 Is the sorter in Zone C running normally?,system_status
 What is the uptime for Sorter-02 today?,system_status
 Recommend a picking strategy for high-velocity SKUs,picking_optimization
@@ -7,70 +7,70 @@ Is the sorter in the receiving dock running normally?,system_status
 The sorter in the mezzanine keeps jamming,equipment_maintenance
 Optimize the pick path for Zone D,picking_optimization
 Is Crane-03 operational?,system_status
-Has order #26915 shipped yet?,order_status
+Has order #22208 shipped yet?,order_status
 What's the difference between AGV and AMR?,general_faq
-What is the current stock level for SKU-7663?,inventory_check
+What is the current stock level for SKU-1481?,inventory_check
 Optimize the pick path for the receiving dock,picking_optimization
 What's the difference between AGV and AMR?,general_faq
 How can we reduce travel time for pickers in Zone C?,picking_optimization
-What is the current stock level for SKU-1214?,inventory_check
-What's the status of order #50538?,order_status
-How much inventory is left for SKU-4959?,inventory_check
+What is the current stock level for SKU-6098?,inventory_check
+What's the status of order #34714?,order_status
+How much inventory is left for SKU-8044?,inventory_check
 What is a WMS?,general_faq
-Is SKU-9180 in stock at Zone A?,inventory_check
+Is SKU-4724 in stock at Zone A?,inventory_check
 Send AGV AMR-21 to Zone B,agv_navigation
 Redirect Sorter-02 around the blocked aisle in Zone B,agv_navigation
 Explain how an AS/RS works,general_faq
-When will order #66314 be delivered?,order_status
-What's the fastest picking route for order #52993?,picking_optimization
-Why hasn't order #81176 left the dock yet?,order_status
+When will order #38043 be delivered?,order_status
+What's the fastest picking route for order #11428?,picking_optimization
+Why hasn't order #82342 left the dock yet?,order_status
 What KPIs matter most in warehouse automation?,general_faq
 File an incident report for Zone D,safety_incident
 "A worker slipped near AMR-21, please log it",safety_incident
 "A worker slipped near Sorter-02, please log it",safety_incident
-What is the current stock level for SKU-1705?,inventory_check
+What is the current stock level for SKU-5101?,inventory_check
 Is Crane-03 operational?,system_status
-What is the current stock level for SKU-1151?,inventory_check
+What is the current stock level for SKU-7785?,inventory_check
 Are all cranes online in Zone D?,system_status
 What is predictive maintenance?,general_faq
 What is the uptime for AMR-21 today?,system_status
-Track order #76958 for me,order_status
+Track order #93776 for me,order_status
 Log a safety incident involving Crane-05,safety_incident
 "Belt AGV-12 stopped unexpectedly, please check",equipment_maintenance
 What KPIs matter most in warehouse automation?,general_faq
-Check inventory count for SKU-1240 in the mezzanine,inventory_check
+Check inventory count for SKU-2025 in the mezzanine,inventory_check
 A forklift near-miss was reported in the receiving dock,safety_incident
 Why is AGV-12 stuck near Zone D?,agv_navigation
-Has order #36763 shipped yet?,order_status
+Has order #58337 shipped yet?,order_status
 What is the current location of AGV-07?,agv_navigation
-Do we have enough SKU-5968 to fulfill 200 units?,inventory_check
+Do we have enough SKU-5631 to fulfill 200 units?,inventory_check
 Report unsafe pallet stacking in Zone D,safety_incident
 What's the difference between AGV and AMR?,general_faq
-Track order #45880 for me,order_status
+Track order #24961 for me,order_status
 "Belt AGV-07 stopped unexpectedly, please check",equipment_maintenance
-Is order #77822 delayed?,order_status
+Is order #13815 delayed?,order_status
 What is the uptime for Conveyor-14 today?,system_status
 There was a near collision between AGV-12 and a pedestrian in Zone B,safety_incident
 Is Sorter-02 operational?,system_status
 How does goods-to-person picking work?,general_faq
 Report unsafe pallet stacking in Zone A,safety_incident
 Suggest a wave picking plan for Zone C,picking_optimization
-Has order #69175 shipped yet?,order_status
-Show the fulfillment status of #72615,order_status
+Has order #86007 shipped yet?,order_status
+Show the fulfillment status of #15798,order_status
 What's the difference between AGV and AMR?,general_faq
-Why hasn't order #13114 left the dock yet?,order_status
+Why hasn't order #21644 left the dock yet?,order_status
 What's the difference between AGV and AMR?,general_faq
-What's the fastest picking route for order #22223?,picking_optimization
-Show the fulfillment status of #80304,order_status
+What's the fastest picking route for order #47398?,picking_optimization
+Show the fulfillment status of #65245,order_status
 Schedule maintenance for AMR-21,equipment_maintenance
 Redirect Crane-03 around the blocked aisle in Zone D,agv_navigation
 Report unsafe pallet stacking in the receiving dock,safety_incident
 Report unsafe pallet stacking in Zone D,safety_incident
-What's the status of order #84458?,order_status
-What's the status of order #26699?,order_status
+What's the status of order #26643?,order_status
+What's the status of order #31576?,order_status
 What is the current location of Sorter-02?,agv_navigation
 "A worker slipped near Crane-03, please log it",safety_incident
-Track order #37216 for me,order_status
+Track order #79604 for me,order_status
 Route Conveyor-14 to picking station 7,agv_navigation
 "Belt AGV-12 stopped unexpectedly, please check",equipment_maintenance
 AGV-12 motor temperature seems high,equipment_maintenance
@@ -78,27 +78,27 @@ Crane Crane-05 reported a fault code,equipment_maintenance
 Suggest a wave picking plan for Zone A,picking_optimization
 Explain how an AS/RS works,general_faq
 A forklift near-miss was reported in Zone D,safety_incident
-Is order #19388 delayed?,order_status
+Is order #69266 delayed?,order_status
 How do sortation systems decide where to route a parcel?,general_faq
-Give me stock levels across all zones for SKU-5790,inventory_check
+Give me stock levels across all zones for SKU-3600,inventory_check
 The sorter in Zone D keeps jamming,equipment_maintenance
 Explain how an AS/RS works,general_faq
-Check inventory count for SKU-9378 in Zone D,inventory_check
+Check inventory count for SKU-6428 in Zone D,inventory_check
 There was a near collision between Crane-05 and a pedestrian in Zone D,safety_incident
 A forklift near-miss was reported in the mezzanine,safety_incident
-How many units of SKU-1573 are in Zone D?,inventory_check
+How many units of SKU-1430 are in Zone D?,inventory_check
 Report unsafe pallet stacking in the receiving dock,safety_incident
 Reassign Sorter-02 to charging station,agv_navigation
 What KPIs matter most in warehouse automation?,general_faq
 Redirect AGV-12 around the blocked aisle in Zone A,agv_navigation
 Route AGV-12 to picking station 3,agv_navigation
 Check system health for the receiving dock,system_status
-Show me the on-hand quantity of SKU-7096,inventory_check
+Show me the on-hand quantity of SKU-2713,inventory_check
 Check system health for Zone A,system_status
 Crane AMR-21 reported a fault code,equipment_maintenance
 Redirect AMR-21 around the blocked aisle in the mezzanine,agv_navigation
 What KPIs matter most in warehouse automation?,general_faq
-Check inventory count for SKU-9515 in Zone C,inventory_check
+Check inventory count for SKU-5725 in Zone C,inventory_check
 Is the sorter in Zone A running normally?,system_status
 What is the uptime for Conveyor-14 today?,system_status
 Crane AGV-12 reported a fault code,equipment_maintenance
@@ -106,94 +106,94 @@ What is predictive maintenance?,general_faq
 Redirect AMR-21 around the blocked aisle in Zone B,agv_navigation
 Reassign Crane-05 to charging station,agv_navigation
 How do sortation systems decide where to route a parcel?,general_faq
-Give me stock levels across all zones for SKU-3527,inventory_check
+Give me stock levels across all zones for SKU-5593,inventory_check
 What KPIs matter most in warehouse automation?,general_faq
-Do we have enough SKU-7090 to fulfill 200 units?,inventory_check
+Do we have enough SKU-6494 to fulfill 200 units?,inventory_check
 Schedule maintenance for Conveyor-14,equipment_maintenance
 Is AGV-07 operational?,system_status
 Crane Conveyor-14 reported a fault code,equipment_maintenance
 What is predictive maintenance?,general_faq
 What's the difference between AGV and AMR?,general_faq
-What is the current stock level for SKU-3744?,inventory_check
+What is the current stock level for SKU-6132?,inventory_check
 Report vibration issue on Sorter-02,equipment_maintenance
 How can we reduce travel time for pickers in Zone B?,picking_optimization
 Is Crane-05 operational?,system_status
-Give me stock levels across all zones for SKU-5721,inventory_check
-Show me the on-hand quantity of SKU-5559,inventory_check
-Track order #19300 for me,order_status
+Give me stock levels across all zones for SKU-5332,inventory_check
+Show me the on-hand quantity of SKU-5115,inventory_check
+Track order #39306 for me,order_status
 Is Crane-05 operational?,system_status
-Show the fulfillment status of #13571,order_status
-Show me the on-hand quantity of SKU-5098,inventory_check
-Why hasn't order #10859 left the dock yet?,order_status
+Show the fulfillment status of #70351,order_status
+Show me the on-hand quantity of SKU-6361,inventory_check
+Why hasn't order #80826 left the dock yet?,order_status
 Report vibration issue on AMR-21,equipment_maintenance
 File an incident report for Zone B,safety_incident
 How does goods-to-person picking work?,general_faq
 What is predictive maintenance?,general_faq
-Show the fulfillment status of #96688,order_status
+Show the fulfillment status of #10862,order_status
 Report unsafe pallet stacking in the receiving dock,safety_incident
 Redirect AMR-21 around the blocked aisle in the mezzanine,agv_navigation
-Is SKU-4451 in stock at Zone B?,inventory_check
+Is SKU-1013 in stock at Zone B?,inventory_check
 Why is Conveyor-14 stuck near the receiving dock?,agv_navigation
 How can we reduce travel time for pickers in Zone A?,picking_optimization
 What is the uptime for Sorter-02 today?,system_status
-Do we have enough SKU-8632 to fulfill 200 units?,inventory_check
-What's the status of order #39040?,order_status
-How much inventory is left for SKU-9838?,inventory_check
-Check inventory count for SKU-4526 in Zone C,inventory_check
+Do we have enough SKU-1897 to fulfill 200 units?,inventory_check
+What's the status of order #56805?,order_status
+How much inventory is left for SKU-1659?,inventory_check
+Check inventory count for SKU-5362 in Zone C,inventory_check
 What KPIs matter most in warehouse automation?,general_faq
 Is Crane-05 operational?,system_status
-How many units of SKU-7701 are in Zone A?,inventory_check
+How many units of SKU-5627 are in Zone A?,inventory_check
 What KPIs matter most in warehouse automation?,general_faq
-Is order #29501 delayed?,order_status
-Track order #16617 for me,order_status
+Is order #59173 delayed?,order_status
+Track order #15605 for me,order_status
 Report unsafe pallet stacking in Zone B,safety_incident
 How can we reduce travel time for pickers in Zone D?,picking_optimization
-Check inventory count for SKU-3493 in the receiving dock,inventory_check
+Check inventory count for SKU-6500 in the receiving dock,inventory_check
 Report unsafe pallet stacking in Zone C,safety_incident
 What is the current location of Crane-05?,agv_navigation
-Give me stock levels across all zones for SKU-6612,inventory_check
-Is SKU-1080 in stock at Zone C?,inventory_check
+Give me stock levels across all zones for SKU-8781,inventory_check
+Is SKU-7378 in stock at Zone C?,inventory_check
 What is predictive maintenance?,general_faq
-Is order #42539 delayed?,order_status
-How many units of SKU-2112 are in the mezzanine?,inventory_check
+Is order #15022 delayed?,order_status
+How many units of SKU-7134 are in the mezzanine?,inventory_check
 Schedule maintenance for AGV-12,equipment_maintenance
 What KPIs matter most in warehouse automation?,general_faq
-Is order #71287 delayed?,order_status
-Show the fulfillment status of #96666,order_status
+Is order #49861 delayed?,order_status
+Show the fulfillment status of #76528,order_status
 Should we batch pick these orders together?,picking_optimization
 How do sortation systems decide where to route a parcel?,general_faq
 The sorter in the mezzanine keeps jamming,equipment_maintenance
 File an incident report for the mezzanine,safety_incident
 Give me the current status of the WMS integration,system_status
-What's the fastest picking route for order #67193?,picking_optimization
+What's the fastest picking route for order #40826?,picking_optimization
 How can we reduce travel time for pickers in Zone D?,picking_optimization
 Why is Sorter-02 stuck near the receiving dock?,agv_navigation
 What is predictive maintenance?,general_faq
 Report vibration issue on Conveyor-14,equipment_maintenance
 File an incident report for the receiving dock,safety_incident
-Has order #11872 shipped yet?,order_status
+Has order #60505 shipped yet?,order_status
 Redirect AMR-21 around the blocked aisle in Zone D,agv_navigation
 Schedule maintenance for AMR-21,equipment_maintenance
 What is the uptime for Sorter-02 today?,system_status
 Send AGV AMR-21 to Zone D,agv_navigation
 Route AGV-12 to picking station 7,agv_navigation
-Has order #23013 shipped yet?,order_status
+Has order #78446 shipped yet?,order_status
 Reassign Conveyor-14 to charging station,agv_navigation
 The sorter in Zone D keeps jamming,equipment_maintenance
 Should we batch pick these orders together?,picking_optimization
 "Belt AGV-07 stopped unexpectedly, please check",equipment_maintenance
 Give me the current status of the WMS integration,system_status
 Are all cranes online in Zone C?,system_status
-Track order #30107 for me,order_status
+Track order #31954 for me,order_status
 Check system health for the mezzanine,system_status
 Report vibration issue on AMR-21,equipment_maintenance
 "A worker slipped near Crane-03, please log it",safety_incident
 Route Conveyor-14 to picking station 7,agv_navigation
-When will order #20466 be delivered?,order_status
+When will order #73619 be delivered?,order_status
 Suggest a wave picking plan for Zone A,picking_optimization
 What is the uptime for Conveyor-14 today?,system_status
 Route AGV-12 to picking station 7,agv_navigation
-Is order #77099 delayed?,order_status
+Is order #50452 delayed?,order_status
 Suggest a wave picking plan for Zone A,picking_optimization
 Check system health for the receiving dock,system_status
 How do sortation systems decide where to route a parcel?,general_faq
@@ -203,18 +203,18 @@ A forklift near-miss was reported in Zone A,safety_incident
 What is the uptime for Sorter-02 today?,system_status
 Is Conveyor-14 operational?,system_status
 Are all cranes online in the mezzanine?,system_status
-Do we have enough SKU-8857 to fulfill 200 units?,inventory_check
+Do we have enough SKU-4848 to fulfill 200 units?,inventory_check
 Why is Conveyor-14 stuck near the receiving dock?,agv_navigation
 Log a breakdown for Conveyor-14 in Zone A,equipment_maintenance
 File an incident report for Zone D,safety_incident
 Why is AGV-12 stuck near Zone A?,agv_navigation
 Recommend a picking strategy for high-velocity SKUs,picking_optimization
 Should we batch pick these orders together?,picking_optimization
-Show me the on-hand quantity of SKU-5184,inventory_check
-Has order #84530 shipped yet?,order_status
+Show me the on-hand quantity of SKU-7831,inventory_check
+Has order #37050 shipped yet?,order_status
 Redirect Conveyor-14 around the blocked aisle in Zone C,agv_navigation
 Is AMR-21 operational?,system_status
-How many units of SKU-2141 are in Zone B?,inventory_check
+How many units of SKU-1033 are in Zone B?,inventory_check
 "A worker slipped near AGV-07, please log it",safety_incident
 A forklift near-miss was reported in Zone D,safety_incident
 "Belt Conveyor-14 stopped unexpectedly, please check",equipment_maintenance
@@ -222,12 +222,12 @@ Log a breakdown for AGV-07 in Zone D,equipment_maintenance
 Crane AMR-21 reported a fault code,equipment_maintenance
 Is Conveyor-14 operational?,system_status
 Why is Crane-05 stuck near Zone B?,agv_navigation
-What's the fastest picking route for order #52316?,picking_optimization
-Show me the on-hand quantity of SKU-8644,inventory_check
+What's the fastest picking route for order #12884?,picking_optimization
+Show me the on-hand quantity of SKU-2617,inventory_check
 Is the sorter in Zone D running normally?,system_status
-When will order #59204 be delivered?,order_status
-When will order #59602 be delivered?,order_status
-How many units of SKU-8300 are in Zone A?,inventory_check
+When will order #79189 be delivered?,order_status
+When will order #68111 be delivered?,order_status
+How many units of SKU-8165 are in Zone A?,inventory_check
 What is the uptime for Crane-03 today?,system_status
 The conveyor belt in the mezzanine is making noise,equipment_maintenance
 What is predictive maintenance?,general_faq
@@ -237,31 +237,31 @@ Optimize the pick path for Zone B,picking_optimization
 A forklift near-miss was reported in Zone C,safety_incident
 How does goods-to-person picking work?,general_faq
 Reassign AGV-07 to charging station,agv_navigation
-Track order #11091 for me,order_status
+Track order #79020 for me,order_status
 Crane-03 motor temperature seems high,equipment_maintenance
 The sorter in Zone D keeps jamming,equipment_maintenance
 Should we batch pick these orders together?,picking_optimization
 Crane AGV-12 reported a fault code,equipment_maintenance
-What is the current stock level for SKU-8634?,inventory_check
-What's the status of order #52708?,order_status
+What is the current stock level for SKU-2932?,inventory_check
+What's the status of order #50176?,order_status
 Check system health for the mezzanine,system_status
-Show the fulfillment status of #26180,order_status
+Show the fulfillment status of #92835,order_status
 What's the difference between AGV and AMR?,general_faq
 The sorter in the mezzanine keeps jamming,equipment_maintenance
 File an incident report for the mezzanine,safety_incident
-What's the status of order #77245?,order_status
+What's the status of order #97649?,order_status
 Log a breakdown for Crane-05 in Zone C,equipment_maintenance
 Send AGV Conveyor-14 to Zone D,agv_navigation
 Is AMR-21 operational?,system_status
 Are all cranes online in Zone A?,system_status
 Should we batch pick these orders together?,picking_optimization
-When will order #18447 be delivered?,order_status
+When will order #18048 be delivered?,order_status
 What is the uptime for Crane-05 today?,system_status
 Report vibration issue on Crane-03,equipment_maintenance
 Reassign Crane-05 to charging station,agv_navigation
 Redirect Crane-05 around the blocked aisle in the mezzanine,agv_navigation
 Suggest a wave picking plan for Zone A,picking_optimization
-What's the fastest picking route for order #11864?,picking_optimization
+What's the fastest picking route for order #67988?,picking_optimization
 How can we reduce travel time for pickers in the receiving dock?,picking_optimization
 "A worker slipped near Crane-05, please log it",safety_incident
 How can we reduce travel time for pickers in the receiving dock?,picking_optimization
@@ -282,7 +282,7 @@ Log a safety incident involving Crane-03,safety_incident
 What KPIs matter most in warehouse automation?,general_faq
 A forklift near-miss was reported in Zone C,safety_incident
 What KPIs matter most in warehouse automation?,general_faq
-Is order #74973 delayed?,order_status
+Is order #86830 delayed?,order_status
 Crane Crane-03 reported a fault code,equipment_maintenance
 Suggest a wave picking plan for Zone C,picking_optimization
 A forklift near-miss was reported in Zone B,safety_incident
@@ -296,7 +296,7 @@ Log a breakdown for AMR-21 in Zone A,equipment_maintenance
 Reassign Conveyor-14 to charging station,agv_navigation
 Suggest a wave picking plan for Zone D,picking_optimization
 What is the current location of Crane-05?,agv_navigation
-What's the status of order #87757?,order_status
+What's the status of order #48499?,order_status
 The conveyor belt in the receiving dock is making noise,equipment_maintenance
 A forklift near-miss was reported in Zone C,safety_incident
 Route Sorter-02 to picking station 12,agv_navigation
@@ -305,58 +305,58 @@ A forklift near-miss was reported in Zone A,safety_incident
 "Belt Crane-05 stopped unexpectedly, please check",equipment_maintenance
 Should we batch pick these orders together?,picking_optimization
 Are all cranes online in Zone A?,system_status
-Track order #36358 for me,order_status
+Track order #25112 for me,order_status
 A forklift near-miss was reported in the receiving dock,safety_incident
 Report unsafe pallet stacking in the receiving dock,safety_incident
 AGV-12 motor temperature seems high,equipment_maintenance
 Is Sorter-02 operational?,system_status
-Show me the on-hand quantity of SKU-2270,inventory_check
+Show me the on-hand quantity of SKU-9509,inventory_check
 What KPIs matter most in warehouse automation?,general_faq
-Is order #22755 delayed?,order_status
+Is order #81134 delayed?,order_status
 Suggest a wave picking plan for Zone B,picking_optimization
-When will order #46139 be delivered?,order_status
+When will order #14162 be delivered?,order_status
 Should we batch pick these orders together?,picking_optimization
 What is the current location of AMR-21?,agv_navigation
 Crane Sorter-02 reported a fault code,equipment_maintenance
-Is SKU-8193 in stock at the mezzanine?,inventory_check
-Check inventory count for SKU-4760 in the mezzanine,inventory_check
-What is the current stock level for SKU-7210?,inventory_check
+Is SKU-3755 in stock at the mezzanine?,inventory_check
+Check inventory count for SKU-1945 in the mezzanine,inventory_check
+What is the current stock level for SKU-4071?,inventory_check
 Report vibration issue on Conveyor-14,equipment_maintenance
 Log a breakdown for AMR-21 in Zone B,equipment_maintenance
 Is the sorter in Zone A running normally?,system_status
 What is a WMS?,general_faq
 Send AGV AGV-12 to the receiving dock,agv_navigation
-When will order #40339 be delivered?,order_status
-How much inventory is left for SKU-5187?,inventory_check
+When will order #90290 be delivered?,order_status
+How much inventory is left for SKU-5795?,inventory_check
 Should we batch pick these orders together?,picking_optimization
 Give me the current status of the WMS integration,system_status
 What is the uptime for AGV-07 today?,system_status
 The conveyor belt in Zone D is making noise,equipment_maintenance
-Check inventory count for SKU-9221 in Zone B,inventory_check
+Check inventory count for SKU-9882 in Zone B,inventory_check
 Give me the current status of the WMS integration,system_status
 Why is Crane-03 stuck near Zone B?,agv_navigation
-What's the fastest picking route for order #39330?,picking_optimization
+What's the fastest picking route for order #10016?,picking_optimization
 Recommend a picking strategy for high-velocity SKUs,picking_optimization
 Check system health for the mezzanine,system_status
-How many units of SKU-7921 are in the mezzanine?,inventory_check
-What's the fastest picking route for order #36192?,picking_optimization
+How many units of SKU-4779 are in the mezzanine?,inventory_check
+What's the fastest picking route for order #28928?,picking_optimization
 What KPIs matter most in warehouse automation?,general_faq
-Why hasn't order #72092 left the dock yet?,order_status
+Why hasn't order #99186 left the dock yet?,order_status
 Redirect AMR-21 around the blocked aisle in Zone C,agv_navigation
 Is the sorter in Zone A running normally?,system_status
 Route Sorter-02 to picking station 12,agv_navigation
 Schedule maintenance for AGV-07,equipment_maintenance
-When will order #49814 be delivered?,order_status
+When will order #75046 be delivered?,order_status
 Recommend a picking strategy for high-velocity SKUs,picking_optimization
 Send AGV Crane-05 to Zone B,agv_navigation
-Show the fulfillment status of #77475,order_status
-How many units of SKU-9050 are in the mezzanine?,inventory_check
-What is the current stock level for SKU-8851?,inventory_check
+Show the fulfillment status of #32378,order_status
+How many units of SKU-7206 are in the mezzanine?,inventory_check
+What is the current stock level for SKU-9688?,inventory_check
 Recommend a picking strategy for high-velocity SKUs,picking_optimization
 "A worker slipped near AMR-21, please log it",safety_incident
 Why is AGV-12 stuck near Zone C?,agv_navigation
 What KPIs matter most in warehouse automation?,general_faq
-Do we have enough SKU-3849 to fulfill 200 units?,inventory_check
+Do we have enough SKU-4869 to fulfill 200 units?,inventory_check
 Check system health for Zone C,system_status
 Redirect Crane-03 around the blocked aisle in Zone B,agv_navigation
 Optimize the pick path for the mezzanine,picking_optimization
@@ -364,18 +364,18 @@ Optimize the pick path for Zone B,picking_optimization
 Report vibration issue on Crane-05,equipment_maintenance
 What is the uptime for AGV-12 today?,system_status
 Crane AGV-12 reported a fault code,equipment_maintenance
-What's the status of order #14373?,order_status
+What's the status of order #42485?,order_status
 Log a breakdown for AMR-21 in Zone D,equipment_maintenance
 The sorter in Zone C keeps jamming,equipment_maintenance
 What is predictive maintenance?,general_faq
-Is SKU-3836 in stock at Zone A?,inventory_check
+Is SKU-9153 in stock at Zone A?,inventory_check
 File an incident report for Zone D,safety_incident
 What is cycle counting?,general_faq
-What's the status of order #34002?,order_status
+What's the status of order #25908?,order_status
 There was a near collision between Conveyor-14 and a pedestrian in Zone A,safety_incident
-What is the current stock level for SKU-5377?,inventory_check
+What is the current stock level for SKU-7712?,inventory_check
 Report vibration issue on AMR-21,equipment_maintenance
-What is the current stock level for SKU-7925?,inventory_check
+What is the current stock level for SKU-3014?,inventory_check
 Explain how an AS/RS works,general_faq
 Give me the current status of the WMS integration,system_status
 Route AGV-07 to picking station 12,agv_navigation
@@ -388,7 +388,7 @@ Redirect AGV-12 around the blocked aisle in the mezzanine,agv_navigation
 Send AGV Crane-03 to Zone A,agv_navigation
 Recommend a picking strategy for high-velocity SKUs,picking_optimization
 What KPIs matter most in warehouse automation?,general_faq
-What's the fastest picking route for order #52930?,picking_optimization
+What's the fastest picking route for order #60573?,picking_optimization
 What's the difference between AGV and AMR?,general_faq
 What's the difference between AGV and AMR?,general_faq
 Log a safety incident involving AGV-12,safety_incident
@@ -396,35 +396,35 @@ File an incident report for the receiving dock,safety_incident
 Recommend a picking strategy for high-velocity SKUs,picking_optimization
 Redirect Crane-03 around the blocked aisle in Zone B,agv_navigation
 "Belt AGV-12 stopped unexpectedly, please check",equipment_maintenance
-Do we have enough SKU-1497 to fulfill 200 units?,inventory_check
+Do we have enough SKU-9423 to fulfill 200 units?,inventory_check
 What is the current location of Crane-05?,agv_navigation
 What is the current location of Conveyor-14?,agv_navigation
 Log a safety incident involving Crane-05,safety_incident
 Suggest a wave picking plan for the mezzanine,picking_optimization
-Is order #15665 delayed?,order_status
-Give me stock levels across all zones for SKU-1705,inventory_check
+Is order #74036 delayed?,order_status
+Give me stock levels across all zones for SKU-6317,inventory_check
 Give me the current status of the WMS integration,system_status
-Do we have enough SKU-6979 to fulfill 200 units?,inventory_check
+Do we have enough SKU-4614 to fulfill 200 units?,inventory_check
 Explain how an AS/RS works,general_faq
 Explain how an AS/RS works,general_faq
-Do we have enough SKU-3890 to fulfill 200 units?,inventory_check
-When will order #23395 be delivered?,order_status
+Do we have enough SKU-9494 to fulfill 200 units?,inventory_check
+When will order #63042 be delivered?,order_status
 Route Crane-03 to picking station 5,agv_navigation
-Check inventory count for SKU-6599 in Zone D,inventory_check
+Check inventory count for SKU-1270 in Zone D,inventory_check
 Should we batch pick these orders together?,picking_optimization
 Is the sorter in Zone A running normally?,system_status
-When will order #29139 be delivered?,order_status
+When will order #86590 be delivered?,order_status
 There was a near collision between Crane-05 and a pedestrian in Zone B,safety_incident
 How does goods-to-person picking work?,general_faq
-How much inventory is left for SKU-1963?,inventory_check
+How much inventory is left for SKU-9371?,inventory_check
 "A worker slipped near AMR-21, please log it",safety_incident
 Optimize the pick path for Zone D,picking_optimization
 Recommend a picking strategy for high-velocity SKUs,picking_optimization
-How many units of SKU-1944 are in the receiving dock?,inventory_check
+How many units of SKU-5484 are in the receiving dock?,inventory_check
 "A worker slipped near AGV-07, please log it",safety_incident
-Show the fulfillment status of #66984,order_status
+Show the fulfillment status of #22905,order_status
 Check system health for Zone D,system_status
-What's the fastest picking route for order #21706?,picking_optimization
+What's the fastest picking route for order #36117?,picking_optimization
 The conveyor belt in Zone A is making noise,equipment_maintenance
 File an incident report for the receiving dock,safety_incident
 What is predictive maintenance?,general_faq
@@ -433,20 +433,20 @@ There was a near collision between AGV-07 and a pedestrian in Zone D,safety_inci
 How do sortation systems decide where to route a parcel?,general_faq
 A forklift near-miss was reported in the receiving dock,safety_incident
 Recommend a picking strategy for high-velocity SKUs,picking_optimization
-Give me stock levels across all zones for SKU-8767,inventory_check
+Give me stock levels across all zones for SKU-9571,inventory_check
 Is AGV-07 operational?,system_status
-What is the current stock level for SKU-6146?,inventory_check
+What is the current stock level for SKU-2340?,inventory_check
 AMR-21 motor temperature seems high,equipment_maintenance
 Why is Crane-03 stuck near Zone C?,agv_navigation
 What's the difference between AGV and AMR?,general_faq
 Why is Sorter-02 stuck near Zone A?,agv_navigation
 AMR-21 motor temperature seems high,equipment_maintenance
-How much inventory is left for SKU-6680?,inventory_check
-Do we have enough SKU-5318 to fulfill 200 units?,inventory_check
+How much inventory is left for SKU-5577?,inventory_check
+Do we have enough SKU-6653 to fulfill 200 units?,inventory_check
 There was a near collision between AGV-12 and a pedestrian in Zone A,safety_incident
 Reassign Conveyor-14 to charging station,agv_navigation
 Should we batch pick these orders together?,picking_optimization
-Do we have enough SKU-1063 to fulfill 200 units?,inventory_check
+Do we have enough SKU-7862 to fulfill 200 units?,inventory_check
 File an incident report for Zone A,safety_incident
 Log a breakdown for AGV-07 in Zone B,equipment_maintenance
 Recommend a picking strategy for high-velocity SKUs,picking_optimization
@@ -462,20 +462,20 @@ Crane Sorter-02 reported a fault code,equipment_maintenance
 Sorter-02 motor temperature seems high,equipment_maintenance
 "Belt Crane-05 stopped unexpectedly, please check",equipment_maintenance
 There was a near collision between Sorter-02 and a pedestrian in Zone D,safety_incident
-Has order #78709 shipped yet?,order_status
+Has order #55893 shipped yet?,order_status
 Is Conveyor-14 operational?,system_status
-Has order #14046 shipped yet?,order_status
+Has order #45816 shipped yet?,order_status
 What is a WMS?,general_faq
 Report unsafe pallet stacking in Zone C,safety_incident
 Route AMR-21 to picking station 3,agv_navigation
 How do sortation systems decide where to route a parcel?,general_faq
 Sorter-02 motor temperature seems high,equipment_maintenance
 What is a WMS?,general_faq
-Do we have enough SKU-3593 to fulfill 200 units?,inventory_check
+Do we have enough SKU-5882 to fulfill 200 units?,inventory_check
 Send AGV AGV-12 to Zone D,agv_navigation
-What's the status of order #38333?,order_status
-Why hasn't order #55768 left the dock yet?,order_status
+What's the status of order #25189?,order_status
+Why hasn't order #47151 left the dock yet?,order_status
 Suggest a wave picking plan for the receiving dock,picking_optimization
 "A worker slipped near Conveyor-14, please log it",safety_incident
 What's the difference between AGV and AMR?,general_faq
-Track order #93785 for me,order_status
+Track order #55871 for me,order_status

data/latency_eval.json

@@ -1,6 +1,6 @@
 {
-  "intent_classifier_ms": 0.309,
-  "anomaly_detector_ms": 17.703,
-  "kb_retrieval_ms": 0.446,
+  "intent_classifier_ms": 0.408,
+  "anomaly_detector_ms": 21.223,
+  "kb_retrieval_ms": 0.617,
   "note": "LLM generation latency depends on the external Inference API call and is measured live in the app, not benchmarked here."
 }

models/intent_pipeline.joblib

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9bbdf1c95a8a9dac3baa157e22462025c8e40dae5ea10b483a8d0e1506c106c1
-size 65196
+oid sha256:b8e49f35a4e88b2a61908f861e49beb4a912a34fa85484c1936878e6ce1a6620
+size 65356

src/llm_client.py

@@ -12,18 +12,38 @@ Design notes
   secrets). The public demo also works without a token: it falls back to a
   deterministic, still-useful extractive answer built from the retrieved
   knowledge-base passages, so the Space never shows a broken demo.
-* The model name is configurable via `LLM_MODEL_ID` (defaults to a small,
-  fast, freely-hostable instruct model).
+* Serverless model availability on the free HF Inference API changes over
+  time (models get gated, deprecated, or moved between providers), so
+  rather than hard-depending on a single model id, we try a short list of
+  candidates in order and use the first one that responds successfully.
+  `LLM_MODEL_ID` (env var) is tried first if set, ahead of the built-in list.
+* On failure, the *actual* exception message (not just its type) is
+  surfaced back to the UI, so a broken deployment is debuggable from the
+  Space itself instead of requiring log access.
 """
 
 import os
 import time
-from dataclasses import dataclass
-from typing import List
+from dataclasses import dataclass, field
+from typing import List, Optional
 
 from src.retriever import KBRetriever, RetrievedDoc
 
-DEFAULT_MODEL_ID = os.environ.get("LLM_MODEL_ID", "Qwen/Qwen2.5-7B-Instruct")
+# Small, widely-available instruct models known to work well on HF's free
+# serverless Inference API. Tried in order; first success wins. If
+# LLM_MODEL_ID is set as an env var, it is tried first, ahead of this list.
+MODEL_CANDIDATES = [
+    "Qwen/Qwen2.5-7B-Instruct",
+    "meta-llama/Llama-3.2-3B-Instruct",
+    "mistralai/Mistral-7B-Instruct-v0.3",
+    "HuggingFaceH4/zephyr-7b-beta",
+]
+
+_env_model = os.environ.get("LLM_MODEL_ID")
+if _env_model:
+    MODEL_CANDIDATES = [_env_model] + [m for m in MODEL_CANDIDATES if m != _env_model]
+
+DEFAULT_MODEL_ID = MODEL_CANDIDATES[0]
 
 SYSTEM_PROMPT = (
     "You are the Smart Warehouse AI Assistant, a helpful operations copilot "
@@ -42,6 +62,7 @@ class AssistantResponse:
     sources: List[RetrievedDoc]
     latency_s: float
     model_id: str
+    debug_errors: List[str] = field(default_factory=list)  # non-empty only when used_llm is False due to failures
 
 
 def _extractive_fallback(query: str, sources: List[RetrievedDoc]) -> str:
@@ -57,8 +78,8 @@ def _extractive_fallback(query: str, sources: List[RetrievedDoc]) -> str:
     lead = sources[0]
     bullets = "\n".join(f"- **{s.title}**: {s.text}" for s in sources)
     return (
-        f"(Offline / no LLM API key configured -- showing retrieved "
-        f"knowledge instead of a generated answer.)\n\n"
+        f"(Showing retrieved knowledge instead of an LLM-generated answer -- "
+        f"see the diagnostics below.)\n\n"
         f"Based on **{lead.title}**, here's the relevant information:\n\n{bullets}"
     )
 
@@ -67,7 +88,6 @@ def answer_query(
     query: str,
     retriever: KBRetriever,
     k: int = 2,
-    model_id: str = DEFAULT_MODEL_ID,
     max_tokens: int = 350,
 ) -> AssistantResponse:
     start = time.time()
@@ -84,35 +104,57 @@ def answer_query(
             sources=sources,
             latency_s=time.time() - start,
             model_id="extractive-fallback",
+            debug_errors=["No HF_TOKEN secret is set on this Space."],
         )
 
     try:
         from huggingface_hub import InferenceClient
-
-        client = InferenceClient(model=model_id, token=hf_token)
-        messages = [
-            {"role": "system", "content": SYSTEM_PROMPT},
-            {
-                "role": "user",
-                "content": f"CONTEXT:\n{context_block}\n\nQUESTION: {query}",
-            },
-        ]
-        completion = client.chat_completion(messages=messages, max_tokens=max_tokens, temperature=0.3)
-        text = completion.choices[0].message.content
-        return AssistantResponse(
-            answer=text,
-            used_llm=True,
-            sources=sources,
-            latency_s=time.time() - start,
-            model_id=model_id,
-        )
-    except Exception as e:  # noqa: BLE001 -- deliberately broad: any API/network issue -> fallback
+    except ImportError as e:
         answer = _extractive_fallback(query, sources)
-        answer += f"\n\n_(LLM call failed: {type(e).__name__}. Showing retrieval-only answer.)_"
         return AssistantResponse(
-            answer=answer,
-            used_llm=False,
-            sources=sources,
-            latency_s=time.time() - start,
-            model_id="extractive-fallback",
+            answer=answer, used_llm=False, sources=sources,
+            latency_s=time.time() - start, model_id="extractive-fallback",
+            debug_errors=[f"huggingface_hub not importable: {e}"],
         )
+
+    messages = [
+        {"role": "system", "content": SYSTEM_PROMPT},
+        {"role": "user", "content": f"CONTEXT:\n{context_block}\n\nQUESTION: {query}"},
+    ]
+
+    errors = []
+    for candidate in MODEL_CANDIDATES:
+        try:
+            client = InferenceClient(model=candidate, token=hf_token)
+            completion = client.chat_completion(messages=messages, max_tokens=max_tokens, temperature=0.3)
+            text = completion.choices[0].message.content
+            if text and text.strip():
+                return AssistantResponse(
+                    answer=text,
+                    used_llm=True,
+                    sources=sources,
+                    latency_s=time.time() - start,
+                    model_id=candidate,
+                )
+            errors.append(f"{candidate}: empty response")
+        except Exception as e:  # noqa: BLE001 -- try the next candidate model
+            errors.append(f"{candidate}: {type(e).__name__}: {e}")
+
+    # All candidates failed -- fall back, but surface the real errors so the
+    # deployment is debuggable directly from the UI.
+    answer = _extractive_fallback(query, sources)
+    return AssistantResponse(
+        answer=answer,
+        used_llm=False,
+        sources=sources,
+        latency_s=time.time() - start,
+        model_id="extractive-fallback",
+        debug_errors=errors,
+    )
+
+
+def test_connection(retriever: Optional[KBRetriever] = None) -> AssistantResponse:
+    """Runs a single canned query through the full pipeline -- used by the
+    'Test LLM connection' diagnostics button in the app."""
+    retriever = retriever or KBRetriever()
+    return answer_query("What is a WMS?", retriever, k=1, max_tokens=60)