Update app.py

app.py

@@ -534,14 +534,15 @@ def retrieve_context(query: str, top_k: int = 10) -> List[Dict[str, Any]]:
         if results:
             dlog("PUBMED", "PubMed search hit")
             return results
-        if _biomechish(q):
-            wiki = wiki_summary_allow(q, sentences=3)
-            if wiki:
-                dlog("WIKI", "Wikipedia biomechanics fallback hit")
-                return [{"text": wiki, "source": "wikipedia"}]
+        # Wikipedia fallback (unconditional after PubMed miss)
+        wiki = wiki_summary_allow(q, sentences=3)
+        if wiki:
+            dlog("WIKI", "Wikipedia fallback hit")
+            return [{"text": wiki, "source": "wikipedia"}]
         dlog("RETRIEVAL", "No results found")
         return []
 
+
     # FAISS path
     q_emb = embed_model.encode([q], convert_to_numpy=True).astype("float32")
     if _IS_IP:
@@ -582,15 +583,16 @@ def retrieve_context(query: str, top_k: int = 10) -> List[Dict[str, Any]]:
         dlog("PUBMED", "PubMed search hit")
         return results
 
-    if _biomechish(q):
-        wiki = wiki_summary_allow(q, sentences=3)
-        if wiki:
-            dlog("WIKI", "Wikipedia biomechanics fallback hit")
-            return [{"text": wiki, "source": "wikipedia"}]
+    # Wikipedia fallback (unconditional after PubMed miss)
+    wiki = wiki_summary_allow(q, sentences=3)
+    if wiki:
+        dlog("WIKI", "Wikipedia fallback hit")
+        return [{"text": wiki, "source": "wikipedia"}]
 
     dlog("RETRIEVAL", "No results at all")
     return []
 
+
 def build_prompt(chunks: List[Dict[str, Any]], question: str) -> str:
     header = (
         "You are Askstein (orthopedic biomechanics). Use ONLY the [Context] to answer. "
@@ -610,6 +612,19 @@ def _decode_generated(out_ids, in_len: int) -> str:
     gen = out_ids[0][in_len:]
     return tokenizer_lm.decode(gen, skip_special_tokens=True).lstrip(". \n").strip()
 
+    @lru_cache(maxsize=None)
+def direct_llm_fallback(question: str) -> str:
+    sys_prompt = (
+        "You are Askstein (orthopedic biomechanics). If you lack enough domain context, say you don’t know. "
+        "Avoid discussing non-musculoskeletal systems (cardiology, neurology)."
+    )
+    llm_prompt = f"{sys_prompt}\n\nQuestion: {question}\nAnswer:"
+    inputs = tokenizer_lm(llm_prompt, return_tensors="pt").to(device)
+    out = _generate(inputs, grounded=False)
+    in_len = inputs["input_ids"].shape[-1]
+    return _post_clean(_decode_generated(out, in_len))
+
+
 def _synthesize_answer(chunks: List[Dict[str, Any]], question: str) -> str:
     prompt = build_prompt(chunks, question)
     inputs = tokenizer_lm(prompt, return_tensors="pt").to(device)