Update app.py

app.py

@@ -149,16 +149,19 @@ def _compute_avg_from_file():
         return "No eval_runs.jsonl yet. Run at least one evaluation."
 
 
+
+
 # ---------- UI ----------
 with gr.Blocks() as demo:
-    gr.Markdown("JWST multimodal search — build the index")
+    gr.Markdown("JWST multimodal search  build the index")
 
+    # Build
     n = gr.Slider(50, 1000, value=200, step=10, label="How many images to index")
     build_btn = gr.Button("Build index")
     status = gr.Textbox(label="Status", lines=2)
-
     build_btn.click(build_index, inputs=n, outputs=status)
 
+    # Search
     gr.Markdown("Search the index with text or an example image")
 
     q_text = gr.Textbox(label="Text query", placeholder="e.g., spiral galaxy")
@@ -166,28 +169,66 @@ with gr.Blocks() as demo:
     k      = gr.Slider(1, 12, value=6, step=1, label="Top K")
 
     search_btn = gr.Button("Search")
+    gallery    = gr.Gallery(label="Results", columns=6, height=300)
+    info2      = gr.Textbox(label="Search status", lines=1)
+    search_btn.click(search, inputs=[q_text, q_img, k], outputs=[gallery, info2])
 
-     # ---------- evaluation UI ----------
+    # ---------- Evaluation (guided) ----------
     with gr.Accordion("Evaluation", open=False):
-        eval_query = gr.Textbox(label="Evaluation query", placeholder="Type a query or leave empty and upload an image")
+        gr.Markdown(
+            "### What this does\n"
+            "We evaluate text to image retrieval using Precision at K.\n"
+            "Steps: pick a preset or type a query, click **Run and label**, "
+            "tick the results that match the rule shown, then click **Compute metrics**. "
+            "Each run is saved so you can average later."
+        )
+
+        # Preset prompts with plain English relevance rules
+        PRESETS = {
+            "star with spikes": "Relevant = bright point source with clear 4 to 6 diffraction spikes. Minimal extended glow.",
+            "edge-on galaxy": "Relevant = thin elongated streak. Looks like a narrow line. No round diffuse blob.",
+            "spiral galaxy": "Relevant = visible spiral arms or a spiral outline. Arms can be faint.",
+            "diffuse nebula": "Relevant = fuzzy cloud like structure. No sharp round core.",
+            "ring or annulus": "Relevant = ring or donut shape is the main feature.",
+            "two merging objects": "Relevant = two bright blobs touching or overlapping."
+        }
+
+        with gr.Row():
+            preset = gr.Dropdown(choices=list(PRESETS.keys()), label="Preset query (optional)")
+            eval_k = gr.Slider(1, 12, value=6, step=1, label="K for evaluation")
+
+        eval_query = gr.Textbox(label="Evaluation query (you can edit or type your own)")
         eval_img   = gr.Image(label="Evaluation image (optional)", type="pil")
-        eval_k     = gr.Slider(1, 12, value=6, step=1, label="K for evaluation")
+        rules_md   = gr.Markdown()
 
         run_and_label = gr.Button("Run and label this query")
 
-        eval_gallery  = gr.Gallery(label="Eval top K results", columns=6, height=300)
+        eval_gallery    = gr.Gallery(label="Eval top K results", columns=6, height=300)
         relevant_picker = gr.CheckboxGroup(label="Select indices of relevant results (1..K)")
-        eval_md = gr.Markdown()
+        eval_md         = gr.Markdown()
 
+        # state bag for this panel
         eval_state = gr.State({"result_indices": [], "k": 5, "query": ""})
 
+        def _on_preset_change(name):
+            if name in PRESETS:
+                return gr.update(value=name), PRESETS[name]
+            return gr.update(), ""
+
+        preset.change(fn=_on_preset_change, inputs=preset, outputs=[eval_query, rules_md])
+
+        # uses helper _search_topk_for_eval defined above
         def _run_eval_query(q_txt, q_img_in, k_in, state):
             items, idxs, _ = _search_topk_for_eval(q_txt, q_img_in, k_in)
             state["result_indices"] = idxs
             state["k"] = int(k_in)
             state["query"] = q_txt if (q_txt and q_txt.strip()) else "[image query]"
             choice_labels = [str(i+1) for i in range(len(idxs))]
-            return items, gr.update(choices=choice_labels, value=[]), "Mark relevant then click Compute metrics.", state
+            help_text = PRESETS.get((q_txt or "").strip().lower(), "Mark results that match the concept you typed.")
+            return (items,
+                    gr.update(choices=choice_labels, value=[]),
+                    f"**Relevance rule:** {help_text}\n\nThen click **Compute metrics**.",
+                    state)
 
         run_and_label.click(
             fn=_run_eval_query,
@@ -197,36 +238,26 @@ with gr.Blocks() as demo:
 
         compute_btn = gr.Button("Compute metrics")
 
+        # uses helpers _save_eval_run and _format_eval_summary defined above
         def _compute_pk(selected_indices, state):
-            k = int(state.get("k", 5))
+            k_val = int(state.get("k", 5))
             query = state.get("query", "")
-            # user marks which of the K are relevant; count is the hits
             hits = len(selected_indices)
-            p_at_k = hits / max(k, 1)
+            p_at_k = hits / max(k_val, 1)
             record = {
                 "ts": int(time.time()),
                 "query": query,
-                "k": k,
+                "k": k_val,
                 "relevant_indices": sorted([int(s) for s in selected_indices]),
                 "precision_at_k": p_at_k
             }
             _save_eval_run(record)
-            return _format_eval_summary(query, k, hits, p_at_k)
+            return _format_eval_summary(query, k_val, hits, p_at_k)
 
-        compute_btn.click(
-            fn=_compute_pk,
-            inputs=[relevant_picker, eval_state],
-            outputs=eval_md
-        )
+        compute_btn.click(fn=_compute_pk, inputs=[relevant_picker, eval_state], outputs=eval_md)
 
         avg_btn = gr.Button("Compute average across saved runs")
         avg_md  = gr.Markdown()
-
         avg_btn.click(fn=_compute_avg_from_file, outputs=avg_md)
-    
-    gallery    = gr.Gallery(label="Results", columns=6, height=300)
-    info2      = gr.Textbox(label="Search status", lines=1)
-
-    search_btn.click(search, inputs=[q_text, q_img, k], outputs=[gallery, info2])
 
 demo.launch()