docs: move Notes section lower

app.py

@@ -130,7 +130,7 @@ def load_models_and_knapsack():
     if INSTANOVO is not None and RESIDUE_SET is not None and KNAPSACK is None:
         # Check if knapsack_cache.zip exists and unzip its contents
         knapsack_zip_path = Path("knapsack_cache.zip")
-        if knapsack_zip_path.exists():
+        if knapsack_zip_path.exists() and not KNAPSACK_DIR.exists():
             logger.info(f"Found {knapsack_zip_path}. Extracting contents to {KNAPSACK_DIR}...")
             try:
                 with zipfile.ZipFile(knapsack_zip_path, 'r') as zip_ref:
@@ -708,23 +708,11 @@ with gr.Blocks(
     )
 
     gr.Markdown(
-        f"""
+        """
         # 🚀 _De Novo_ Peptide Sequencing with InstaNovo and InstaNovo+
         Upload your mass spectrometry data file (.mgf, .mzml, or .mzxml) and get peptide sequence predictions.
         Choose your prediction method and decoding options.
 
-         **Notes:**
-         *   Predictions use version `{TRANSFORMER_MODEL_ID}` for the transformer-based InstaNovo model and version `{DIFFUSION_MODEL_ID}` for the diffusion-based InstaNovo+ model.
-         *   The InstaNovo+ model `{DIFFUSION_MODEL_ID}` is an alpha release.
-         * **Predction Modes:**
-             *   **InstaNovo with InstaNovo+ refinement** Runs initial prediction with the selected Transformer method (Greedy/Knapsack), then refines using InstaNovo+.
-             *   **InstaNovo Only:** Uses only the Transformer with the selected decoding method.
-             *   **InstaNovo+ Only:** Predicts directly using the Diffusion model (alpha release).
-        * **Transformer Decoding Methods:**
-             *   **Greedy Search:** use this for optimal performance, has similar performance as Knapsack Beam Search at 5% FDR.
-             *   **Knapsack Beam Search:** use this for the best results and highest peptide recall, but is about 10x slower than Greedy Search.
-         *   Check logs for progress, especially for large files or slower methods.
-
          This Hugging Face Space is powered by a [ZeroGPU ](https://huggingface.co/docs/hub/en/spaces-zerogpu), which is free but **limited to 5 minutes per day per user**—so if you test with your own files, please use only small files.
          """,
          elem_classes="feedback"
@@ -803,6 +791,19 @@ with gr.Blocks(
         label="Example Usage:",
     )
 
+    gr.Markdown(f"""**Notes:**
+         *   Predictions use version `{TRANSFORMER_MODEL_ID}` for the transformer-based InstaNovo model and version `{DIFFUSION_MODEL_ID}` for the diffusion-based InstaNovo+ model.
+         *   The InstaNovo+ model `{DIFFUSION_MODEL_ID}` is an alpha release.
+         * **Predction Modes:**
+             *   **InstaNovo with InstaNovo+ refinement** Runs initial prediction with the selected Transformer method (Greedy/Knapsack), then refines using InstaNovo+.
+             *   **InstaNovo Only:** Uses only the Transformer with the selected decoding method.
+             *   **InstaNovo+ Only:** Predicts directly using the Diffusion model (alpha release).
+        * **Transformer Decoding Methods:**
+             *   **Greedy Search:** use this for optimal performance, has similar performance as Knapsack Beam Search at 5% FDR.
+             *   **Knapsack Beam Search:** use this for the best results and highest peptide recall, but is about 10x slower than Greedy Search.
+         *   Check logs for progress, especially for large files or slower methods.
+    """)
+
     with gr.Accordion("Application Logs", open=True):
         log_display = Log(log_file, dark=True, height=300)