UX: search auto-populates and auto-ranks (no Enter needed); perf: precomputed embeddings with weights-md5 guard (cold start ~90s -> ~15s), single-pass lowercase search blobs

app.py

@@ -88,6 +88,22 @@ target_df.rename(
 
 BATCH_SIZE = 1024
 
+# Precomputed lowercase search blobs: one plain-substring scan per keystroke
+# instead of five case-insensitive regex scans. The newline separator prevents
+# accidental cross-field matches (queries never contain newlines).
+_DISEASE_SEARCH_BLOB = (
+    disease_df["name"].astype(str)
+    + "\n" + disease_df["diseaseId"].astype(str)
+    + "\n" + disease_df["synonyms"].astype(str)
+    + "\n" + disease_df["ExactSynonyms"].astype(str)
+    + "\n" + disease_df["description"].astype(str)
+).str.lower()
+_TARGET_SEARCH_BLOB = (
+    target_df["approvedSymbol"].astype(str)
+    + "\n" + target_df["targetId"].astype(str)
+    + "\n" + target_df["approvedName"].astype(str)
+).str.lower()
+
 
 def _load_weights_file() -> str:
     last_error = None
@@ -166,6 +182,38 @@ def _load_model_with_weights():
     return loaded_model
 
 
+def _weights_md5() -> str:
+    import hashlib
+    try:
+        return hashlib.md5(Path(_load_weights_file()).read_bytes()).hexdigest()
+    except Exception:
+        return ""
+
+
+def _load_packaged_embeddings(loaded_model):
+    """Load precomputed candidate/disease embeddings if present AND produced
+    from the exact weights file we just loaded (md5 guard). Returns
+    (cand_embs, dis_embs) or None -> caller computes live. Cuts the cold-start
+    embedding precompute (~1 min on the free CPU tier) to a file read."""
+    path = DATA_DIR / "embeddings.npz"
+    if not path.exists():
+        return None
+    try:
+        data = np.load(path, allow_pickle=False)
+        if str(data["weights_md5"]) != _weights_md5():
+            print("Packaged embeddings are for different weights; recomputing live.")
+            return None
+        cand, dis = data["candidate_embs"], data["disease_embs"]
+        if len(cand) != len(target_df) or len(dis) != len(disease_df):
+            print("Packaged embeddings shape mismatch; recomputing live.")
+            return None
+        print(f"Loaded packaged embeddings: candidates {cand.shape}, diseases {dis.shape}")
+        return cand.astype(np.float32), dis.astype(np.float32)
+    except Exception as error:
+        print(f"Packaged embeddings unreadable ({error}); recomputing live.")
+        return None
+
+
 def _precompute_candidate_embeddings(loaded_model):
     print("Precomputing candidate embeddings (batched)...")
     target_texts = target_df["target_text"].astype(str).to_numpy()
@@ -211,10 +259,18 @@ def _precompute_disease_embeddings(loaded_model):
 @lru_cache(maxsize=1)
 def _get_runtime_cached():
     loaded_model = _load_model_with_weights()
+    packaged = _load_packaged_embeddings(loaded_model)
+    if packaged is not None:
+        cand_embs, dis_embs = packaged
+        _PACKAGED_DISEASE_EMBS.append(dis_embs)
+        return loaded_model, cand_embs
     cand_embs = _precompute_candidate_embeddings(loaded_model)
     return loaded_model, cand_embs
 
 
+_PACKAGED_DISEASE_EMBS: list = []
+
+
 def get_runtime():
     with RUNTIME_LOCK:
         return _get_runtime_cached()
@@ -223,6 +279,8 @@ def get_runtime():
 @lru_cache(maxsize=1)
 def _get_disease_runtime_cached():
     loaded_model, _ = get_runtime()
+    if _PACKAGED_DISEASE_EMBS:
+        return loaded_model, _PACKAGED_DISEASE_EMBS[0]
     dis_embs = _precompute_disease_embeddings(loaded_model)
     return loaded_model, dis_embs
 
@@ -937,13 +995,7 @@ def search_targets(query):
         return gr.update(choices=[], value=None)
     query = query.strip()
     lowered = query.lower()
-    mask = (
-        target_df["approvedSymbol"].str.contains(query, case=False, na=False)
-        | target_df["targetId"].str.contains(query, case=False, na=False)
-        | target_df["approvedName"]
-        .astype(str)
-        .str.contains(query, case=False, na=False)
-    )
+    mask = _TARGET_SEARCH_BLOB.str.contains(lowered, regex=False, na=False)
     matches = target_df.loc[mask].copy()
     if matches.empty:
         return gr.update(choices=[], value=None)
@@ -962,7 +1014,7 @@ def search_targets(query):
         )
         for _, row in matches.iterrows()
     ]
-    return gr.update(choices=choices, value=None)
+    return gr.update(choices=choices, value=choices[0][1])
 
 
 # ============================================
@@ -976,17 +1028,7 @@ def search_diseases(query):
 
     query = query.strip()
     lowered_query = query.lower()
-    mask = (
-        disease_df["name"].str.contains(query, case=False, na=False)
-        | disease_df["diseaseId"].str.contains(query, case=False, na=False)
-        | disease_df["synonyms"].astype(str).str.contains(query, case=False, na=False)
-        | disease_df["ExactSynonyms"]
-        .astype(str)
-        .str.contains(query, case=False, na=False)
-        | disease_df["description"]
-        .astype(str)
-        .str.contains(query, case=False, na=False)
-    )
+    mask = _DISEASE_SEARCH_BLOB.str.contains(lowered_query, regex=False, na=False)
 
     matches = disease_df.loc[mask].copy()
     if matches.empty:
@@ -1007,7 +1049,9 @@ def search_diseases(query):
         for _, row in matches.iterrows()
     ]
 
-    return gr.update(choices=choices, value=None)
+    # Auto-select the top match so the user never has to open the dropdown;
+    # selecting a value does NOT trigger ranking (only Rank/sliders/select do).
+    return gr.update(choices=choices, value=choices[0][1])
 
 
 def launch():
@@ -1091,13 +1135,6 @@ def launch():
 
                 gr.Examples(examples=examples, inputs=search_box)
 
-                search_box.input(
-                    fn=search_diseases, inputs=search_box, outputs=did_dropdown
-                )
-                search_box.submit(
-                    fn=search_diseases, inputs=search_box, outputs=did_dropdown
-                )
-
                 inputs = [
                     search_box,
                     did_dropdown,
@@ -1114,7 +1151,16 @@ def launch():
                     out_file_full,
                 ]
 
+                # Typing populates AND ranks: per-disease scoring is lru-cached,
+                # so keystrokes that keep the same top match cost only a render.
+                search_box.input(
+                    fn=search_diseases, inputs=search_box, outputs=did_dropdown
+                ).then(fn=run_disease_query, inputs=inputs, outputs=outputs)
+                search_box.submit(
+                    fn=search_diseases, inputs=search_box, outputs=did_dropdown
+                ).then(fn=run_disease_query, inputs=inputs, outputs=outputs)
                 btn.click(fn=run_disease_query, inputs=inputs, outputs=outputs)
+                did_dropdown.select(fn=run_disease_query, inputs=inputs, outputs=outputs)
                 topk.change(fn=run_disease_query, inputs=inputs, outputs=outputs)
                 filter_mode.change(fn=run_disease_query, inputs=inputs, outputs=outputs)
                 sort_by.change(fn=run_disease_query, inputs=inputs, outputs=outputs)
@@ -1151,15 +1197,16 @@ def launch():
                     )
                 )
                 t_out_file = gr.File(label="Download full disease ranking (CSV)")
+                t_inputs = [t_search_box, tid_dropdown, t_topk, t_min_score]
+                t_outputs = [target_summary, t_out_df, t_out_file]
                 t_search_box.input(
                     fn=search_targets, inputs=t_search_box, outputs=tid_dropdown
-                )
+                ).then(fn=run_target_query, inputs=t_inputs, outputs=t_outputs)
                 t_search_box.submit(
                     fn=search_targets, inputs=t_search_box, outputs=tid_dropdown
-                )
-                t_inputs = [t_search_box, tid_dropdown, t_topk, t_min_score]
-                t_outputs = [target_summary, t_out_df, t_out_file]
+                ).then(fn=run_target_query, inputs=t_inputs, outputs=t_outputs)
                 t_btn.click(fn=run_target_query, inputs=t_inputs, outputs=t_outputs)
+                tid_dropdown.select(fn=run_target_query, inputs=t_inputs, outputs=t_outputs)
                 t_topk.change(fn=run_target_query, inputs=t_inputs, outputs=t_outputs)
                 t_min_score.change(
                     fn=run_target_query, inputs=t_inputs, outputs=t_outputs

data/proc/embeddings.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4bea154adf07e9a408e1c4efba502bd7963fc1842cd0f63e47f5c956ec5e392d
+size 84362056