From a8e9bc53de64ad98e22c19837454fac7699a0803 Mon Sep 17 00:00:00 2001
From: Brad Stein <Brad.Stein@gmail.com>
Date: Sun, 1 Feb 2026 05:09:32 -0300
Subject: [PATCH] atlasbot: keep raw keywords for metric fallback

---
 atlasbot/engine/answerer.py | 18 ++++++++++++------
 1 file changed, 12 insertions(+), 6 deletions(-)

diff --git a/atlasbot/engine/answerer.py b/atlasbot/engine/answerer.py
index a4b862c..29efcb2 100644
--- a/atlasbot/engine/answerer.py
+++ b/atlasbot/engine/answerer.py
@@ -184,7 +184,7 @@ class AnswerEngine:
             normalized = str(normalize.get("normalized") or question).strip() or question
             keywords = normalize.get("keywords") or []
             _debug_log("normalize_parsed", {"normalized": normalized, "keywords": keywords})
-            keyword_tokens = _extract_keywords(normalized, sub_questions=[], keywords=keywords)
+            keyword_tokens = _extract_keywords(question, normalized, sub_questions=[], keywords=keywords)
 
             if observer:
                 observer("route", "routing")
@@ -250,7 +250,7 @@ class AnswerEngine:
             parts = _parse_json_list(decompose_raw)
             sub_questions = _select_subquestions(parts, normalized, plan.max_subquestions)
             _debug_log("decompose_parsed", {"sub_questions": sub_questions})
-            keyword_tokens = _extract_keywords(normalized, sub_questions=sub_questions, keywords=keywords)
+            keyword_tokens = _extract_keywords(question, normalized, sub_questions=sub_questions, keywords=keywords)
 
             snapshot_context = ""
             if classify.get("needs_snapshot"):
@@ -407,9 +407,10 @@ class AnswerEngine:
                 )
             if classify.get("question_type") in {"metric", "diagnostic"} and metric_facts and not re.search(r"\\d", reply):
                 best_line = None
-                lowered = normalized.lower()
+                lowered_keywords = [kw.lower() for kw in keyword_tokens if kw]
                 for line in metric_facts:
-                    if any(token in line.lower() for token in lowered.split()):
+                    line_lower = line.lower()
+                    if any(kw in line_lower for kw in lowered_keywords):
                         best_line = line
                         break
                 best_line = best_line or metric_facts[0]
@@ -987,7 +988,12 @@ def _needs_focus_fix(question: str, reply: str, classify: dict[str, Any]) -> boo
     return any(marker in reply.lower() for marker in extra_markers)
 
 
-def _extract_keywords(normalized: str, sub_questions: list[str], keywords: list[Any] | None) -> list[str]:
+def _extract_keywords(
+    raw_question: str,
+    normalized: str,
+    sub_questions: list[str],
+    keywords: list[Any] | None,
+) -> list[str]:
     stopwords = {
         "the",
         "and",
@@ -1011,7 +1017,7 @@ def _extract_keywords(normalized: str, sub_questions: list[str], keywords: list[
         "now",
     }
     tokens: list[str] = []
-    for source in [normalized, *sub_questions]:
+    for source in [raw_question, normalized, *sub_questions]:
         for part in re.split(r"[^a-zA-Z0-9_-]+", source.lower()):
             if len(part) < 3 or part in stopwords:
                 continue