atlasbot: fix metric fact detection

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"):
@@ -260,7 +260,7 @@ class AnswerEngine:
                 scored = await _score_chunks(call_llm, chunks, normalized, sub_questions, plan)
                 selected = _select_chunks(chunks, scored, plan, keyword_tokens)
                 key_facts = _key_fact_lines(summary_lines, keyword_tokens)
-                metric_facts = [line for line in key_facts if re.search(r"\\d", line)]
+                metric_facts = [line for line in key_facts if re.search(r"\d", line)]
                 if self._settings.debug_pipeline:
                     scored_preview = sorted(
                         [{"id": c["id"], "score": scored.get(c["id"], 0.0), "summary": c["summary"]} for c in chunks],
@@ -405,15 +405,19 @@ class AnswerEngine:
                     model=plan.model,
                     tag="focus_fix",
                 )
-            if classify.get("question_type") in {"metric", "diagnostic"} and metric_facts and not re.search(r"\\d", reply):
+            if classify.get("question_type") in {"metric", "diagnostic"} and metric_facts:
                 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]
-                reply = f"From the latest snapshot: {best_line}."
+                reply_numbers = set(re.findall(r"\d+(?:\.\d+)?", reply))
+                fact_numbers = set(re.findall(r"\d+(?:\.\d+)?", " ".join(metric_facts)))
+                if not reply_numbers or (fact_numbers and not (reply_numbers & fact_numbers)):
+                    reply = f"From the latest snapshot: {best_line}."
 
             if plan.use_critic:
                 if observer:
@@ -443,6 +447,9 @@ class AnswerEngine:
                 if note:
                     reply = f"{reply}\n\n{note}"
 
+            if classify.get("question_type") in {"metric", "diagnostic"} and metric_facts:
+                reply = _metric_fact_guard(reply, metric_facts, keyword_tokens)
+
             scores = await self._score_answer(normalized, reply, plan, call_llm)
             claims = await self._extract_claims(normalized, reply, summary, call_llm)
         except LLMLimitReached:
@@ -873,16 +880,45 @@ def _key_fact_lines(lines: list[str], keywords: list[str] | None, limit: int = 6
     lowered = [kw.lower() for kw in keywords if kw]
     if not lowered:
         return []
+    focused = _focused_keywords(lowered)
+    primary = focused or lowered
     matches: list[str] = []
     for line in lines:
         line_lower = line.lower()
-        if any(kw in line_lower for kw in lowered):
+        if any(kw in line_lower for kw in primary):
             matches.append(line)
         if len(matches) >= limit:
             break
+    if len(matches) < limit and focused:
+        for line in lines:
+            if len(matches) >= limit:
+                break
+            if line in matches:
+                continue
+            line_lower = line.lower()
+            if any(kw in line_lower for kw in lowered):
+                matches.append(line)
     return matches
 
 
+def _metric_fact_guard(reply: str, metric_facts: list[str], keywords: list[str]) -> str:
+    if not metric_facts:
+        return reply
+    best_line = None
+    lowered_keywords = [kw.lower() for kw in keywords if kw]
+    for line in metric_facts:
+        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]
+    reply_numbers = set(re.findall(r"\d+(?:\.\d+)?", reply))
+    fact_numbers = set(re.findall(r"\d+(?:\.\d+)?", " ".join(metric_facts)))
+    if not reply_numbers or (fact_numbers and not (reply_numbers & fact_numbers)):
+        return f"From the latest snapshot: {best_line}."
+    return reply
+
+
 def _lexicon_context(summary: dict[str, Any]) -> str:
     if not isinstance(summary, dict):
         return ""
@@ -987,7 +1023,12 @@ 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 +1052,7 @@ 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