atlasbot: steer metric focus with route hints

atlasbot/engine/answerer.py

@@ -201,6 +201,8 @@ class AnswerEngine:
             classify.setdefault("needs_snapshot", True)
             classify.setdefault("answer_style", "direct")
             classify.setdefault("follow_up", False)
+            classify.setdefault("focus_entity", "unknown")
+            classify.setdefault("focus_metric", "unknown")
             _debug_log("route_parsed", {"classify": classify, "normalized": normalized})
             cluster_terms = (
                 "atlas",
@@ -227,14 +229,27 @@ class AnswerEngine:
             )
             if any(term in normalized.lower() for term in cluster_terms):
                 classify["needs_snapshot"] = True
+            lowered_norm = normalized.lower()
+            if (
+                ("namespace" in lowered_norm and ("pod" in lowered_norm or "pods" in lowered_norm))
+                or re.search(r"\bmost\s+pods\b", lowered_norm)
+                or re.search(r"\bpods\s+running\b", lowered_norm)
+            ):
+                classify["question_type"] = "metric"
+                classify["needs_snapshot"] = True
             if re.search(r"\b(how many|count|number of|list)\b", normalized.lower()):
                 classify["question_type"] = "metric"
             hottest_terms = ("hottest", "highest", "lowest", "most")
-            metric_terms = ("cpu", "ram", "memory", "net", "network", "io", "disk", "load", "usage")
-            lowered_question = normalized.lower()
+            metric_terms = ("cpu", "ram", "memory", "net", "network", "io", "disk", "load", "usage", "pod", "pods", "namespace")
+            lowered_question = f"{question} {normalized}".lower()
             if any(term in lowered_question for term in hottest_terms) and any(term in lowered_question for term in metric_terms):
                 classify["question_type"] = "metric"
 
+            if not classify.get("follow_up") and state and state.claims:
+                follow_terms = ("there", "that", "those", "these", "it", "them", "that one", "this", "former", "latter")
+                if any(term in lowered_question for term in follow_terms) or len(normalized.split()) <= 6:
+                    classify["follow_up"] = True
+
             if classify.get("follow_up") and state and state.claims:
                 if observer:
                     observer("followup", "answering follow-up")
@@ -257,6 +272,8 @@ class AnswerEngine:
             sub_questions = _select_subquestions(parts, normalized, plan.max_subquestions)
             _debug_log("decompose_parsed", {"sub_questions": sub_questions})
             keyword_tokens = _extract_keywords(question, normalized, sub_questions=sub_questions, keywords=keywords)
+            focus_entity = str(classify.get("focus_entity") or "unknown").lower()
+            focus_metric = str(classify.get("focus_metric") or "unknown").lower()
 
             snapshot_context = ""
             if classify.get("needs_snapshot"):
@@ -282,12 +299,15 @@ class AnswerEngine:
                                     hardware_facts = [seg.strip() for seg in line.split(" | ") if seg.strip().startswith("hardware_usage_avg:")]
                                     break
                 metric_facts = [line for line in key_facts if re.search(r"\d", line)]
-                if hardware_facts:
+                if focus_entity == "node" and hottest_facts:
+                    metric_facts = hottest_facts
+                    key_facts = _merge_fact_lines(metric_facts, key_facts)
+                elif hardware_facts:
                     metric_facts = hardware_facts
                     key_facts = _merge_fact_lines(metric_facts, key_facts)
                 if classify.get("question_type") in {"metric", "diagnostic"}:
                     lowered_q = f"{question} {normalized}".lower()
-                    if any(tok in lowered_q for tok in ("hardware", "class", "type", "rpi", "jetson", "amd64", "arm64")) and any(
+                    if focus_entity != "node" and any(tok in lowered_q for tok in ("hardware", "class", "type", "rpi", "jetson", "amd64", "arm64")) and any(
                         tok in lowered_q for tok in ("average", "avg", "mean", "ram", "memory", "cpu", "load")
                     ):
                         hw_top = None
@@ -300,7 +320,7 @@ class AnswerEngine:
                         if hw_top:
                             metric_facts = [hw_top]
                             key_facts = _merge_fact_lines(metric_facts, key_facts)
-                if hottest_facts and not hardware_facts:
+                if hottest_facts and not hardware_facts and focus_entity != "class":
                     metric_facts = hottest_facts
                     key_facts = _merge_fact_lines(metric_facts, key_facts)
                 if classify.get("question_type") in {"metric", "diagnostic"} and not hottest_facts and not hardware_facts:
@@ -312,15 +332,24 @@ class AnswerEngine:
                         if self._settings.debug_pipeline:
                             _debug_log("metric_facts_selected", {"facts": metric_facts})
                 if classify.get("question_type") in {"metric", "diagnostic"} and not metric_facts:
-                    for line in summary_lines:
-                        if "hardware_usage_top:" in line:
-                            metric_facts = [seg.strip() for seg in line.split(" | ") if seg.strip().startswith("hardware_usage_top:")]
-                            break
-                    if not metric_facts:
+                    lowered_q = f"{question} {normalized}".lower()
+                    if "namespace" in lowered_q and "pod" in lowered_q:
                         for line in summary_lines:
-                            if "hardware_usage_avg:" in line:
-                                metric_facts = [seg.strip() for seg in line.split(" | ") if seg.strip().startswith("hardware_usage_avg:")]
+                            if line.startswith("namespaces_top:"):
+                                metric_facts = [line]
                                 break
+                    if not metric_facts:
+                        hardware_tokens = ("hardware", "class", "type", "rpi", "jetson", "amd64", "arm64")
+                        if any(tok in lowered_q for tok in hardware_tokens):
+                            for line in summary_lines:
+                                if "hardware_usage_top:" in line:
+                                    metric_facts = [seg.strip() for seg in line.split(" | ") if seg.strip().startswith("hardware_usage_top:")]
+                                    break
+                            if not metric_facts:
+                                for line in summary_lines:
+                                    if "hardware_usage_avg:" in line:
+                                        metric_facts = [seg.strip() for seg in line.split(" | ") if seg.strip().startswith("hardware_usage_avg:")]
+                                        break
                     if metric_facts:
                         key_facts = _merge_fact_lines(metric_facts, key_facts)
                 if self._settings.debug_pipeline:
@@ -513,6 +542,61 @@ class AnswerEngine:
             if classify.get("question_type") in {"metric", "diagnostic"} and metric_facts:
                 reply = _metric_fact_guard(reply, metric_facts, keyword_tokens)
 
+            if classify.get("question_type") in {"metric", "diagnostic"}:
+                lowered_q = f"{question} {normalized}".lower()
+                if any(tok in lowered_q for tok in ("how many", "count", "number of")) and any(
+                    tok in lowered_q for tok in ("jetson", "rpi4", "rpi5", "amd64", "arm64", "rpi")
+                ):
+                    hw_line = next((line for line in summary_lines if line.startswith("hardware:")), None)
+                    hw_nodes_line = next((line for line in summary_lines if line.startswith("hardware_nodes:")), None)
+                    if hw_line:
+                        def _find_value(key: str, line: str) -> str | None:
+                            match = re.search(rf"{re.escape(key)}=([^;|]+)", line)
+                            return match.group(1).strip() if match else None
+
+                        target = None
+                        if "jetson" in lowered_q:
+                            target = "jetson"
+                        elif "rpi5" in lowered_q:
+                            target = "rpi5"
+                        elif "rpi4" in lowered_q:
+                            target = "rpi4"
+                        elif "amd64" in lowered_q:
+                            target = "amd64"
+                        elif "arm64" in lowered_q:
+                            target = "arm64"
+                        elif "rpi" in lowered_q:
+                            target = "rpi"
+                        if target:
+                            count = _find_value(target, hw_line)
+                            nodes = _find_value(target, hw_nodes_line or "")
+                            if count:
+                                if nodes and "(" not in count:
+                                    reply = f"From the latest snapshot: {target}={count} ({nodes})."
+                                else:
+                                    reply = f"From the latest snapshot: {target}={count}."
+                if ("io" in lowered_q or "i/o" in lowered_q) and ("node" in lowered_q or "nodes" in lowered_q):
+                    io_facts = _extract_hottest_facts(summary_lines, lowered_q)
+                    io_line = next((fact for fact in io_facts if fact.startswith("hottest_io_node")), None)
+                    if io_line:
+                        reply = f"From the latest snapshot: {io_line}."
+                if "namespace" in lowered_q and "pod" in lowered_q:
+                    ns_line = None
+                    for line in summary_lines:
+                        if line.startswith("namespaces_top:"):
+                            ns_line = line
+                            break
+                    cpu_line = None
+                    if any(tok in lowered_q for tok in ("cpu", "hottest", "highest cpu", "highest")):
+                        cpu_facts = _extract_hottest_facts(summary_lines, lowered_q)
+                        cpu_line = next((fact for fact in cpu_facts if fact.startswith("hottest_cpu_node")), None)
+                    if ns_line:
+                        if cpu_line:
+                            reply = f"From the latest snapshot: {cpu_line}; {ns_line}."
+                        else:
+                            reply = f"From the latest snapshot: {ns_line}."
+                        # do not fall through to other overrides
+
             if classify.get("question_type") in {"metric", "diagnostic"}:
                 lowered_q = f"{question} {normalized}".lower()
                 if any(tok in lowered_q for tok in ("hardware", "class", "type", "rpi", "jetson", "amd64", "arm64")) and any(
@@ -724,6 +808,7 @@ class AnswerEngine:
                 tag="followup_fix",
             )
         reply = await self._dedup_reply(reply, plan, call_llm, tag="dedup_followup")
+        reply = _strip_followup_meta(reply)
         return reply
 
     async def _select_claims(
@@ -771,6 +856,26 @@ class AnswerEngine:
         self._store.cleanup()
 
 
+def _strip_followup_meta(reply: str) -> str:
+    cleaned = reply.strip()
+    if not cleaned:
+        return cleaned
+    prefixes = [
+        "The draft is correct based on the provided context.",
+        "The draft is correct based on the context.",
+        "The draft is correct based on the provided evidence.",
+        "The draft is correct.",
+        "Based on the provided context,",
+        "Based on the context,",
+        "Based on the provided evidence,",
+    ]
+    for prefix in prefixes:
+        if cleaned.lower().startswith(prefix.lower()):
+            cleaned = cleaned[len(prefix) :].lstrip(" .")
+            break
+    return cleaned
+
+
 def _build_meta(
     mode: str,
     call_count: int,
@@ -1091,6 +1196,8 @@ def _extract_hottest_facts(lines: list[str], question: str) -> list[str]:
     if not facts:
         return []
     wanted = []
+    if ("io" in lowered or "i/o" in lowered) and ("disk" in lowered or "storage" in lowered):
+        return [fact for fact in facts if fact.startswith("hottest_io_node")]
     if any(term in lowered for term in ("cpu", "processor")):
         wanted.append("hottest_cpu_node")
     if any(term in lowered for term in ("ram", "memory")):
@@ -1156,6 +1263,13 @@ def _metric_candidate_lines(lines: list[str], keywords: list[str] | None, limit:
         "p95",
         "percent",
         "pressure",
+        "pod",
+        "pods",
+        "namespace",
+        "anomaly",
+        "anomalies",
+        "pvc",
+        "storage",
     }
     candidates: list[str] = []
     for line in lines:

atlasbot/llm/prompts.py

@@ -30,7 +30,8 @@ ROUTE_SYSTEM = (
 
 ROUTE_PROMPT = (
     "Return JSON with fields: needs_snapshot (bool), needs_kb (bool), needs_tool (bool), "
-    "answer_style (direct|insightful), follow_up (bool), question_type (metric|diagnostic|planning|open_ended)."
+    "answer_style (direct|insightful), follow_up (bool), question_type (metric|diagnostic|planning|open_ended), "
+    "focus_entity (node|class|namespace|service|cluster|unknown), focus_metric (cpu|ram|net|io|disk|load|pods|storage|unknown)."
 )
 
 DECOMPOSE_SYSTEM = (