Remove corpus counts from chat chrome

apps/ai-service/adapters/bedrock_converse.py

@@ -9,8 +9,8 @@ One structural difference from the Anthropic path drives the shape of this file:
 Converse has **no** server-side response schema (no `output_config.format`), so
 the JSON envelope `answer.py` parses cannot be enforced by the API. It is asked
 for in the prompt and then isolated here (`_extract_json`) before returning. If
-the model still emits something unparseable, `answer.py` falls back to the
-verbatim source text — losing the rewrite, never the answer.
+the model still emits something unparseable, `answer.py` fails closed with a
+diagnosable abstention instead of presenting a raw dump as a generated answer.
 """
 from __future__ import annotations
 
@@ -92,17 +92,15 @@ class BedrockConverseAnswerGenerator:
                 BEDROCK_RUNTIME_SERVICE,
                 region_name=self._region,
                 config=Config(
-                    connect_timeout=10,
-                    read_timeout=60,
-                    # "adaptive" was tried 2026-08-07 and reverted same day: its
-                    # client-side rate limiter remembers "throttled" across
-                    # requests and paces even unrelated, otherwise-healthy calls
-                    # down after a burst — turned a single answerable turn's
-                    # baseline ~9s into 1-5 MINUTES following this session's own
-                    # heavy adversarial test traffic. "standard" retries each
-                    # call independently, no shared state to get stuck in a bad
-                    # regime. max_attempts alone (3->4) is kept.
-                    retries={"max_attempts": 4, "mode": "standard"},
+                    connect_timeout=5,
+                    # The request budget is checked between calls and cannot
+                    # interrupt boto3 while a call is in flight. A 60s read
+                    # timeout with four SDK attempts allowed one turn to run
+                    # for minutes. Allow one bounded retry: production smoke
+                    # tests showed an isolated Qwen read timeout immediately
+                    # followed by a healthy 3s response for the same request.
+                    read_timeout=20,
+                    retries={"total_max_attempts": 2, "mode": "standard"},
                 ),
             )
         return self._client