atlasbot: route subjective queries to LLM

services/comms/atlasbot-deployment.yaml

@@ -16,7 +16,7 @@ spec:
       labels:
         app: atlasbot
       annotations:
-        checksum/atlasbot-configmap: manual-atlasbot-68
+        checksum/atlasbot-configmap: manual-atlasbot-69
         vault.hashicorp.com/agent-inject: "true"
         vault.hashicorp.com/role: "comms"
         vault.hashicorp.com/agent-inject-secret-turn-secret: "kv/data/atlas/comms/turn-shared-secret"

services/comms/scripts/atlasbot/bot.py

@@ -1911,19 +1911,6 @@ def cluster_answer(
     history_lines: list[str] | None = None,
 ) -> str:
     metrics_summary = snapshot_context(prompt, snapshot)
-    if _is_insight_query(prompt):
-        candidates = _insight_candidates(inventory, snapshot)
-        used_keys = _recent_insight_keys(history_lines or [])
-        selected = _select_insight(prompt, candidates, used_keys=used_keys)
-        if selected:
-            key, raw_text, confidence = selected
-            formatted = _format_insight_text(key, raw_text)
-            if not formatted:
-                formatted = raw_text
-            prefix = _insight_prefix(prompt)
-            if prefix:
-                formatted = prefix + formatted
-            return _format_confidence(formatted, confidence)
     structured = structured_answer(
         prompt,
         inventory=inventory,
@@ -2422,6 +2409,17 @@ def _history_payload_lines(history_payload: list[Any]) -> list[str]:
     return [line for line in lines if line]
 
 
+def _append_history_context(context: str, history_lines: list[str]) -> str:
+    lines = [line.strip() for line in history_lines if isinstance(line, str) and line.strip()]
+    if not lines:
+        return context
+    snippet = "\n".join(lines[-6:])
+    combined = context + "\nRecent chat:\n" + snippet if context else "Recent chat:\n" + snippet
+    if len(combined) > MAX_CONTEXT_CHARS:
+        combined = combined[: MAX_CONTEXT_CHARS - 3].rstrip() + "..."
+    return combined
+
+
 # Internal HTTP endpoint for cluster answers (website uses this).
 class _AtlasbotHandler(BaseHTTPRequestHandler):
     server_version = "AtlasbotHTTP/1.0"
@@ -2493,15 +2491,25 @@ class _AtlasbotHandler(BaseHTTPRequestHandler):
             )
         fallback = "I don't have enough data to answer that."
         if cluster_query:
-            answer = cluster_answer(
+            facts_answer = cluster_answer(
                 cleaned,
                 inventory=inventory,
                 snapshot=snapshot,
                 workloads=workloads,
                 history_lines=history_lines,
             )
-            if not answer:
-                answer = fallback
+            open_ended = _is_subjective_query(cleaned) or _knowledge_intent(cleaned)
+            if open_ended:
+                llm_context = _append_history_context(context, history_lines)
+                answer = ollama_reply(
+                    ("http", "internal"),
+                    cleaned,
+                    context=llm_context,
+                    fallback=facts_answer or fallback,
+                    use_history=False,
+                )
+            else:
+                answer = facts_answer or fallback
         else:
             llm_prompt = cleaned
             answer = ollama_reply(
@@ -2761,11 +2769,13 @@ def _ollama_call(hist_key, prompt: str, *, context: str, use_history: bool = Tru
         "When a cluster snapshot is provided, never answer about unrelated meanings of 'Atlas' (maps, mythology, Apache Atlas, etc). "
         "Treat 'hottest' as highest utilization (CPU/RAM/NET/IO) rather than temperature. "
         "If you infer or synthesize, say 'Based on the snapshot' and keep it brief. "
+        "For subjective prompts (interesting, favorite, unconventional), pick one or two observations from the context, explain why they stand out in 1-2 sentences, and avoid repeating the same observation as the last response if you can. "
         "Prefer exact repo paths and Kubernetes resource names when relevant. "
         "Never include or request secret values. "
         "Do not suggest commands unless explicitly asked. "
         "Respond in plain sentences; do not return JSON or code fences unless explicitly asked. "
         "Translate metrics into natural language instead of echoing raw label/value pairs. "
+        "Avoid bare lists unless the user asked for a list; weave numbers into sentences. "
         "Do not answer by only listing runbooks; if the question is about Atlas/Othrys, summarize the cluster first and mention docs only if useful. "
         "If the question is not about Atlas/Othrys and no cluster context is provided, answer using general knowledge and say when you are unsure. "
         "If the answer is not grounded in the provided context or tool data, say you do not know. "
@@ -2974,15 +2984,27 @@ def sync_loop(token: str, room_id: str):
                 fallback = "I don't have enough data to answer that."
 
                 if cluster_query:
-                    reply = cluster_answer(
+                    facts_answer = cluster_answer(
                         cleaned_body,
                         inventory=inventory,
                         snapshot=snapshot,
                         workloads=workloads,
                         history_lines=history[hist_key],
                     )
-                    if not reply:
-                        reply = fallback
+                    open_ended = _is_subjective_query(cleaned_body) or _knowledge_intent(cleaned_body)
+                    if open_ended:
+                        llm_context = _append_history_context(context, history[hist_key])
+                        reply = ollama_reply_with_thinking(
+                            token,
+                            rid,
+                            hist_key,
+                            cleaned_body,
+                            context=llm_context,
+                            fallback=facts_answer or fallback,
+                            use_history=False,
+                        )
+                    else:
+                        reply = facts_answer or fallback
                 else:
                     llm_prompt = cleaned_body
                     reply = ollama_reply_with_thinking(