sync: from ael backend/space/

rerank.py

@@ -105,6 +105,18 @@ DROP_BELOW = float(os.environ.get("AEL_RERANK_DROP", "0.35"))
 # one, not a target: a short page is fine, an empty one is not.
 MIN_RESULTS = 1
 
+# …but one result is not a page you can research with. When the threshold would
+# leave fewer than this, the *highest-scoring* results are kept instead of the
+# original set: the floor must not resurrect the junk, so it tops the list back
+# up from the model's own ranking rather than falling back to engine order.
+#
+# This is not the "floor of 3" the module's own docstring warns about. That one
+# fell back to the whole original set, so a set of eight with two worth showing
+# served all eight. Topping up from the ranking cannot do that: the results that
+# come back are still the ones the model rated highest, just including a couple
+# more of them.
+MIN_KEEP = int(os.environ.get("AEL_RERANK_MIN_KEEP", "2"))
+
 # How many results the model is allowed to judge. A 421M cross-encoder on a
 # shared CPU does not score eight documents inside an 800ms budget, and the
 # first version's answer to that was to throw the whole set away — which is why
@@ -370,8 +382,24 @@ def rerank(query, results, budget_ms=800):
     # Highest score first. Unscored results keep their relative position at the
     # end rather than being dropped: we have no opinion about them.
     scored.sort(key=lambda t: (t[0] is None, -(t[0] if t[0] is not None else 0.0), t[1]))
-    kept = [r for p, _, r in scored if p is None or p >= DROP_BELOW]
-    dropped = len(scored) - len(kept)
+
+    judged_rows = [row for row in scored if row[0] is not None]
+    below = [row for row in judged_rows if row[0] < DROP_BELOW]
+    above = [row for row in judged_rows if row[0] >= DROP_BELOW]
+    unscored = [row for row in scored if row[0] is None]
+
+    if len(above) + len(unscored) < MIN_KEEP and below and above:
+        # Top up from the model's own ranking — never from the original order.
+        # Guarded on `above` being non-empty: when *nothing* cleared the bar
+        # that is not a thin page, it is a model that did not understand the
+        # query, and it is reported as starvation below instead of being
+        # disguised as two confident results.
+        need = MIN_KEEP - (len(above) + len(unscored))
+        above = above + below[:need]
+        below = below[need:]
+
+    kept = [r for _, _, r in above] + [r for _, _, r in unscored]
+    dropped = len(below)
 
     out["scores"] = [
         {"title": str(r.get("title") or "")[:80], "score": p}