Update app.py

app.py

@@ -1,36 +1,33 @@
+"""Document Q&A with Manual RAG - CPU inference on a ZeroGPU Space.
 
-"""Document Q&A with Manual RAG — CPU inference on ZeroGPU hardware."""
+The two visible actions stay on CPU. A hidden, registered GPU handler
+is present solely to satisfy ZeroGPU startup validation.
+"""
 
-# Import spaces before any ML libraries
-import spaces
+import spaces  # Import before Gradio, torch, or Hugging Face model libraries.
 
 import os
 import traceback
 
+# Do NOT set CUDA_VISIBLE_DEVICES on ZeroGPU hardware.
 os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
 
 import gradio as gr
 
-from rag_core import (
-    answer_question,
-    display_evidence,
-    load_retriever,
-)
-
+from rag_core import answer_question, display_evidence, load_retriever
 
-# -----------------------------------------
-# ZeroGPU startup compatibility
-# -----------------------------------------
 
 @spaces.GPU(duration=10)
 def zerogpu_registration():
-    """Registered but intentionally unused GPU function."""
+    """Registered but intentionally unused ZeroGPU handler.
+
+    The host requires at least one @spaces.GPU decorated Gradio event
+    on a ZeroGPU Space. Users do not click this handler, and the two
+    visible buttons do all work on CPU.
+    """
     return "ZeroGPU registration handler (unused)."
 
 
-# -----------------------------------------
-# Custom styling
-# -----------------------------------------
 
 CUSTOM_CSS = """
 .gradio-container {
@@ -45,16 +42,11 @@ CUSTOM_CSS = """
     border-radius: 20px;
     background: linear-gradient(
         125deg,
-        color-mix(
-            in srgb,
-            var(--primary-500) 13%,
-            var(--block-background-fill)
-        ) 0%,
+        color-mix(in srgb, var(--primary-500) 13%, var(--block-background-fill)) 0%,
         var(--block-background-fill) 75%
     );
     margin-bottom: 18px;
 }
-
 #rag-hero .eyebrow {
     font-size: 0.76rem;
     text-transform: uppercase;
@@ -63,7 +55,6 @@ CUSTOM_CSS = """
     color: var(--primary-500);
     margin-bottom: 9px;
 }
-
 #rag-hero h1 {
     font-size: clamp(1.65rem, 3vw, 2.35rem);
     font-weight: 760;
@@ -71,7 +62,6 @@ CUSTOM_CSS = """
     letter-spacing: -0.035em;
     margin: 0 0 10px;
 }
-
 #rag-hero p {
     color: var(--body-text-color-subdued);
     font-size: 0.99rem;
@@ -79,369 +69,28 @@ CUSTOM_CSS = """
     line-height: 1.6;
     margin: 0;
 }
-
 #rag-hero .meta {
     font-size: 0.82rem;
     color: var(--body-text-color-subdued);
     margin-top: 15px;
 }
-
-#rag-input textarea {
-    font-size: 1.03rem !important;
-}
-
-#rag-actions button {
-    min-height: 48px !important;
-    font-weight: 650;
-}
-
-#rag-status {
-    margin: 4px 0 8px;
-}
-
-#rag-answer textarea {
-    font-size: 1.04rem !important;
-    line-height: 1.6 !important;
-}
-
-#rag-evidence textarea {
-    font-size: 0.94rem !important;
-    line-height: 1.55 !important;
-}
-
+#rag-input textarea { font-size: 1.03rem !important; }
+#rag-actions button { min-height: 48px !important; font-weight: 650; }
+#rag-status { margin: 4px 0 8px; }
+#rag-status p { font-size: 0.9rem; color: var(--body-text-color-subdued); }
+#rag-result-heading h3 { font-size: 1.13rem; margin: 12px 0 0; }
+#rag-answer textarea { font-size: 1.04rem !important; line-height: 1.6 !important; }
+#rag-evidence textarea { font-size: 0.94rem !important; line-height: 1.55 !important; }
+#rag-footnote p { font-size: 0.82rem; color: var(--body-text-color-subdued); }
 @media (max-width: 680px) {
-    #rag-hero {
-        padding: 20px 18px;
-        border-radius: 16px;
-    }
-
-    .gradio-container {
-        padding: 12px 10px 22px !important;
-    }
+    #rag-hero { padding: 20px 18px; border-radius: 16px; }
+    .gradio-container { padding: 12px 10px 22px !important; }
 }
 """
 
 
-# -----------------------------------------
-# Format results
-# -----------------------------------------
-
 def format_result(result):
-
+    """Return answer, complete source evidence, and a simple status line."""
     answer = result.get("answer", "")
     evidence = display_evidence(result)
-
-    hits = result.get("retrieved") or []
-    count = len(hits)
-
-    top_score = hits[0]["score"] if hits else None
-
-    score_note = (
-        f" | Top cosine similarity: {top_score:.3f}"
-        if top_score is not None
-        else ""
-    )
-
-    status = result.get("status", "")
-
-    if status == "retrieval_only":
-        answer = (
-            "Evidence retrieved. Select Generate Answer "
-            "to create a response."
-        )
-        message = (
-            f"**Evidence retrieved** | "
-            f"{count} source passages{score_note}"
-        )
-
-    elif status == "generated":
-        message = (
-            f"**Answer generated on CPU** | "
-            f"{count} source passages{score_note}"
-        )
-
-    elif status == "low_similarity_refusal":
-        message = (
-            "**No answer generated** | "
-            "Evidence is below the similarity threshold."
-            f"{score_note}"
-        )
-
-    else:
-        message = (
-            f"**Completed on CPU** | "
-            f"{count} source passages{score_note}"
-        )
-
-    return answer, evidence, message
-
-
-# -----------------------------------------
-# CPU evidence retrieval
-# -----------------------------------------
-
-def retrieve_evidence(question):
-
-    try:
-        result = answer_question(
-            question,
-            retrieval_only=True
-        )
-
-        return format_result(result)
-
-    except ValueError as error:
-        return (
-            "Please check your question.",
-            "",
-            f"**Input error:** {error}"
-        )
-
-    except Exception:
-        traceback.print_exc()
-
-        return (
-            "Unable to retrieve evidence.",
-            "",
-            "**Retrieval error.** Check the logs."
-        )
-
-
-# -----------------------------------------
-# CPU answer generation
-# -----------------------------------------
-
-def generate_answer(question):
-
-    try:
-        result = answer_question(
-            question,
-            retrieval_only=False
-        )
-
-        return format_result(result)
-
-    except ValueError as error:
-        return (
-            "Please check your question.",
-            "",
-            f"**Input error:** {error}"
-        )
-
-    except Exception:
-        traceback.print_exc()
-
-        return (
-            "Unable to generate an answer.",
-            "",
-            "**Generation error.** Check the logs."
-        )
-
-
-# -----------------------------------------
-# Build interface
-# -----------------------------------------
-
-def make_demo():
-
-    with gr.Blocks(
-        title="Document Q&A with Manual RAG"
-    ) as demo:
-
-        gr.HTML(
-            """
-            <section id="rag-hero">
-                <div class="eyebrow">
-                    Manual RAG | CPU Inference
-                </div>
-
-                <h1>
-                    Document Q&amp;A with Manual RAG
-                </h1>
-
-                <p>
-                    Ask a question, retrieve supporting
-                    passages, and generate an answer
-                    using only the available documents.
-                </p>
-
-                <div class="meta">
-                    30 SQuAD/Wikipedia passages
-                    &nbsp; | &nbsp;
-                    MiniLM Retrieval
-                    &nbsp; | &nbsp;
-                    FLAN-T5 Small Generation
-                </div>
-            </section>
-            """
-        )
-
-        question = gr.Textbox(
-            label="Ask a Question",
-            placeholder="Which company produces the iPod?",
-            lines=2,
-            max_lines=3,
-            max_length=300,
-            elem_id="rag-input"
-        )
-
-        with gr.Row(elem_id="rag-actions"):
-
-            answer_button = gr.Button(
-                "Generate Answer",
-                variant="primary",
-                size="lg"
-            )
-
-            retrieve_button = gr.Button(
-                "Retrieve Evidence",
-                variant="secondary",
-                size="lg"
-            )
-
-        status = gr.Markdown(
-            "**Ready** | CPU inference enabled.",
-            elem_id="rag-status"
-        )
-
-        gr.Markdown("### Generated Answer")
-
-        answer = gr.Textbox(
-            label="Answer",
-            placeholder="Your answer will appear here.",
-            lines=4,
-            max_lines=8,
-            interactive=False,
-            buttons=["copy"],
-            elem_id="rag-answer"
-        )
-
-        with gr.Accordion(
-            "Source Passages and Similarity Scores",
-            open=True
-        ):
-
-            evidence = gr.Textbox(
-                show_label=False,
-                placeholder=(
-                    "Retrieved passages will appear here."
-                ),
-                lines=11,
-                max_lines=14,
-                interactive=False,
-                buttons=["copy"],
-                elem_id="rag-evidence"
-            )
-
-        gr.Examples(
-            examples=[
-                ["Which company produces the iPod?"],
-                ["At what age did Frédéric move to Paris?"],
-                [
-                    "What is the live price of Bitcoin "
-                    "in Nigerian naira right now?"
-                ]
-            ],
-            inputs=[question],
-            cache_examples=False,
-            label="Try an Example Question"
-        )
-
-        gr.Markdown(
-            """
-            **Note:** Answers come from the 30
-            available passages, not the live internet.
-            Cosine similarity is not a confidence score.
-            Always verify answers against the sources.
-            """
-        )
-
-        # Visible CPU actions
-        outputs = [
-            answer,
-            evidence,
-            status
-        ]
-
-        answer_button.click(
-            fn=generate_answer,
-            inputs=question,
-            outputs=outputs
-        )
-
-        retrieve_button.click(
-            fn=retrieve_evidence,
-            inputs=question,
-            outputs=outputs
-        )
-
-        # Hidden compatibility event.
-        # ZeroGPU requires a GPU-decorated function
-        # registered directly with a Gradio event.
-
-        hidden_button = gr.Button(
-            "ZeroGPU registration",
-            visible=False,
-            interactive=False
-        )
-
-        hidden_output = gr.Textbox(
-            visible=False
-        )
-
-        hidden_button.click(
-            fn=zerogpu_registration,
-            inputs=[],
-            outputs=hidden_output,
-            api_visibility="private"
-        )
-
-    return demo
-
-
-# -----------------------------------------
-# Initialize at module level
-# -----------------------------------------
-
-print(
-    "Starting Manual RAG: CPU inference "
-    "on ZeroGPU hardware",
-    flush=True
-)
-
-shape = load_retriever()
-
-print(
-    f"Retriever initialized: {shape}",
-    flush=True
-)
-
-# Register all Gradio events before launch
-demo = make_demo()
-
-demo.queue(
-    default_concurrency_limit=1,
-    max_size=20
-)
-
-
-# -----------------------------------------
-# Launch application
-# -----------------------------------------
-
-if __name__ == "__main__":
-
-    demo.launch(
-        server_name="0.0.0.0",
-        server_port=int(
-            os.getenv("PORT", "7860")
-        ),
-        share=False,
-        show_error=True,
-        ssr_mode=False,
-        theme=gr.themes.Soft(
-            primary_hue="indigo",
-            neutral_hue="slate"
-        ),
-        css=CUSTOM_CSS
-    )
+    hits = result.get("retrieved") or []