amazon-bedrock-haystack 8.2.0__tar.gz → 8.3.1__tar.gz

{amazon_bedrock_haystack-8.2.0 → amazon_bedrock_haystack-8.3.1}/CHANGELOG.md

@@ -1,5 +1,18 @@
 # Changelog
 
+## [integrations/amazon_bedrock-v8.3.0] - 2026-10-01
+
+### 🚜 Refactor
+
+- Generalize flattened generation_kwargs for tool_choice/thinking/output_config (#4028)
+
+## [integrations/amazon_bedrock-v8.2.0] - 2026-09-24
+
+### 🚀 Features
+
+- Add thinking_display flattened generation_kwarg to Anthropic and Bedrock chat generators (#3993)
+
+
 ## [integrations/amazon_bedrock-v8.1.1] - 2026-09-23
 
 ### 🐛 Bug Fixes

{amazon_bedrock_haystack-8.2.0 → amazon_bedrock_haystack-8.3.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: amazon-bedrock-haystack
-Version: 8.2.0
+Version: 8.3.1
 Summary: An integration of AWS S3 and Bedrock as a Downloader and Generator components.
 Project-URL: Documentation, https://github.com/deepset-ai/haystack-core-integrations/tree/main/integrations/amazon_bedrock#readme
 Project-URL: Issues, https://github.com/deepset-ai/haystack-core-integrations/issues

{amazon_bedrock_haystack-8.2.0 → amazon_bedrock_haystack-8.3.1}/src/haystack_integrations/components/generators/amazon_bedrock/chat/chat_generator.py

@@ -550,11 +550,6 @@ class AmazonBedrockChatGenerator:
             tool_choice["disable_parallel_tool_use"] = disable_parallel_tool_use
             tool_choice.setdefault("type", "auto")  # default value
 
-        tool_choice_type = generation_kwargs.pop("tool_choice_type", None)
-        if tool_choice_type is not None:
-            tool_choice = generation_kwargs.setdefault("tool_choice", {})
-            tool_choice["type"] = tool_choice_type
-
         thinking_budget_tokens = generation_kwargs.pop("thinking_budget_tokens", None)
         if thinking_budget_tokens is not None:
             thinking = generation_kwargs.setdefault("thinking", {})
@@ -571,12 +566,14 @@ class AmazonBedrockChatGenerator:
                 output_config = generation_kwargs.setdefault("output_config", {})
                 output_config["effort"] = adaptive_thinking_effort
 
-        thinking_display = generation_kwargs.pop("thinking_display", None)
-        if thinking_display is not None:
-            thinking = generation_kwargs.setdefault("thinking", {})
-            # `display` is only accepted with enabled or adaptive thinking, so it's dropped when thinking is disabled
-            if thinking.get("type") != "disabled":
-                thinking["display"] = thinking_display
+        for nested_name in ("tool_choice", "thinking", "output_config"):
+            prefix = f"{nested_name}_"
+            for key in [k for k in generation_kwargs if k.startswith(prefix)]:
+                value = generation_kwargs.pop(key)
+                if value is not None:
+                    nested_key = key[len(prefix) :]
+                    nested_dict = generation_kwargs.setdefault(nested_name, {})
+                    nested_dict[nested_key] = value
 
         return generation_kwargs
 

{amazon_bedrock_haystack-8.2.0 → amazon_bedrock_haystack-8.3.1}/src/haystack_integrations/components/generators/amazon_bedrock/chat/utils.py

@@ -478,7 +478,7 @@ def _parse_completion_response(response_body: dict[str, Any], model: str) -> lis
                     "total_tokens": response_body.get("usage", {}).get("totalTokens", 0),
                     "cache_read_input_tokens": response_body.get("usage", {}).get("cacheReadInputTokens", 0),
                     "cache_write_input_tokens": response_body.get("usage", {}).get("cacheWriteInputTokens", 0),
-                    "cache_details": response_body.get("usage", {}).get("CacheDetails", {}),
+                    "cache_details": response_body.get("usage", {}).get("cacheDetails", []),
                 },
             }
             # guardrail trace
@@ -615,10 +615,10 @@ def _convert_event_to_streaming_chunk(
             streaming_chunk = StreamingChunk(
                 content="",
                 index=block_idx,
-                reasoning=ReasoningContent(
-                    reasoning_text=reasoning_text,
-                    extra=extra,
-                ),
+                # redactedContent-only deltas have no text or signature, skip them like the non-streaming path
+                reasoning=ReasoningContent(reasoning_text=reasoning_text, extra=extra)
+                if reasoning_text or extra
+                else None,
                 meta=base_meta,
             )
 
@@ -642,7 +642,7 @@ def _convert_event_to_streaming_chunk(
                 "total_tokens": usage.get("totalTokens", 0),
                 "cache_read_input_tokens": usage.get("cacheReadInputTokens", 0),
                 "cache_write_input_tokens": usage.get("cacheWriteInputTokens", 0),
-                "cache_details": usage.get("cacheDetails", {}),
+                "cache_details": usage.get("cacheDetails", []),
             }
         if "trace" in event_meta:
             chunk_meta["trace"] = event_meta["trace"]

{amazon_bedrock_haystack-8.2.0 → amazon_bedrock_haystack-8.3.1}/tests/test_chat_generator.py

@@ -80,7 +80,8 @@ MODELS_TO_TEST_WITH_THINKING_TOOLS_STREAMING = [
     "global.anthropic.claude-sonnet-4-6",
 ]
 
-MODELS_TO_TEST_WITH_PROMPT_CACHING = ["us.amazon.nova-micro-v1:0"]  # cheap, fast model
+# smallest Claude model; Claude models report cache writes split by TTL
+MODELS_TO_TEST_WITH_PROMPT_CACHING = ["global.anthropic.claude-haiku-4-5-20251001-v1:0"]
 
 
 def hello_world():
@@ -707,23 +708,27 @@ class TestAmazonBedrockChatGenerator:
                 },
             ),
             (
-                # thinking_display is dropped when thinking is disabled through a flattened kwarg
+                # thinking type in_between with effort
                 {
-                    "adaptive_thinking_effort": "none",
+                    "thinking_type": "in_between",
                     "thinking_display": "summarized",
+                    "output_config_effort": "high",
                 },
                 {
-                    "thinking": {"type": "disabled"},
+                    "thinking": {"type": "in_between", "display": "summarized"},
+                    "output_config": {"effort": "high"},
                 },
             ),
             (
-                # thinking_display is dropped when thinking is disabled through an explicit thinking dict
+                # thinking type adaptive with effort
                 {
-                    "thinking": {"type": "disabled"},
+                    "thinking_type": "adaptive",
                     "thinking_display": "summarized",
+                    "output_config_effort": "high",
                 },
                 {
-                    "thinking": {"type": "disabled"},
+                    "thinking": {"type": "adaptive", "display": "summarized"},
+                    "output_config": {"effort": "high"},
                 },
             ),
         ],
@@ -1622,8 +1627,9 @@ class TestAmazonBedrockChatGeneratorInference:
 
         system_message = ChatMessage.from_system("Always respond with: 'Life is beautiful' (and nothing else).")
 
+        # Claude Haiku 4.5 caches only prompts of at least 4,096 tokens
         user_message = ChatMessage.from_user(
-            "User message that should be long enough to cache. " * 100,
+            "User message that should be long enough to cache. " * 450,
             meta={"cachePoint": {"type": "default"}},
         )
         messages = [system_message, user_message]
@@ -1636,7 +1642,10 @@ class TestAmazonBedrockChatGeneratorInference:
         # tests run in parallel based on the workflow matrix, so this request should either hit the cache (read tokens)
         # or populate it (write tokens)
         assert usage["cache_read_input_tokens"] > 1000 or usage["cache_write_input_tokens"] > 1000
-        assert "cache_details" in usage
+        if usage["cache_write_input_tokens"] > 0:
+            # a cache point without `ttl` uses the default 5-minute TTL
+            expected = [{"ttl": "5m", "inputTokens": usage["cache_write_input_tokens"]}]
+            assert usage["cache_details"] == expected, f"Unexpected usage stats: {usage}"
 
     @pytest.mark.parametrize("model_name", MODELS_TO_TEST_WITH_TOOLS[:1])  # just one model is enough
     @pytest.mark.skipif(ToolInvoker is None, reason="ToolInvoker is not available in the installed haystack-ai version")

{amazon_bedrock_haystack-8.2.0 → amazon_bedrock_haystack-8.3.1}/tests/test_chat_generator_utils.py

@@ -768,8 +768,8 @@ class TestAmazonBedrockChatGeneratorUtils:
                 "outputTokens": 20,
                 "totalTokens": 30,
                 "cacheReadInputTokens": 1000,
-                "cacheWriteInputTokens": 0,
-                "cacheDetails": {},
+                "cacheWriteInputTokens": 1500,
+                "cacheDetails": [{"ttl": "5m", "inputTokens": 500}, {"ttl": "1h", "inputTokens": 1000}],
             },
         }
 
@@ -785,8 +785,8 @@ class TestAmazonBedrockChatGeneratorUtils:
                 "completion_tokens": 20,
                 "total_tokens": 30,
                 "cache_read_input_tokens": 1000,
-                "cache_write_input_tokens": 0,
-                "cache_details": {},
+                "cache_write_input_tokens": 1500,
+                "cache_details": [{"ttl": "5m", "inputTokens": 500}, {"ttl": "1h", "inputTokens": 1000}],
             },
             "index": 0,
         }
@@ -828,7 +828,7 @@ class TestAmazonBedrockChatGeneratorUtils:
                 "total_tokens": 40,
                 "cache_read_input_tokens": 0,
                 "cache_write_input_tokens": 0,
-                "cache_details": {},
+                "cache_details": [],
             },
             "index": 0,
         }
@@ -872,7 +872,7 @@ class TestAmazonBedrockChatGeneratorUtils:
                 "total_tokens": 60,
                 "cache_read_input_tokens": 0,
                 "cache_write_input_tokens": 0,
-                "cache_details": {},
+                "cache_details": [],
             },
             "index": 0,
         }
@@ -951,7 +951,7 @@ class TestAmazonBedrockChatGeneratorUtils:
                     "total_tokens": 500,
                     "cache_read_input_tokens": 0,
                     "cache_write_input_tokens": 0,
-                    "cache_details": {},
+                    "cache_details": [],
                 },
             },
         )
@@ -1006,7 +1006,7 @@ class TestAmazonBedrockChatGeneratorUtils:
                 "totalTokens": 558,
                 "cacheReadInputTokens": 0,
                 "cacheWriteInputTokens": 0,
-                "cacheDetails": {},
+                "cacheDetails": [],
             },
             "metrics": {"latencyMs": 4811},
         }
@@ -1040,7 +1040,7 @@ class TestAmazonBedrockChatGeneratorUtils:
                     "total_tokens": 558,
                     "cache_read_input_tokens": 0,
                     "cache_write_input_tokens": 0,
-                    "cache_details": {},
+                    "cache_details": [],
                 },
             },
         )
@@ -1121,7 +1121,7 @@ class TestAmazonBedrockChatGeneratorUtils:
                 "total_tokens": 0,
                 "cache_read_input_tokens": 0,
                 "cache_write_input_tokens": 0,
-                "cache_details": {},
+                "cache_details": [],
             },
             "trace": trace,
         }
@@ -1372,7 +1372,7 @@ class TestAmazonBedrockChatGeneratorUtils:
                         "total_tokens": 435,
                         "cache_read_input_tokens": 0,
                         "cache_write_input_tokens": 0,
-                        "cache_details": {},
+                        "cache_details": [],
                     },
                 },
             )
@@ -1479,7 +1479,7 @@ class TestAmazonBedrockChatGeneratorUtils:
                         "total_tokens": 435,
                         "cache_read_input_tokens": 0,
                         "cache_write_input_tokens": 0,
-                        "cache_details": {},
+                        "cache_details": [],
                     },
                 },
                 component_info=c_info,
@@ -1657,7 +1657,7 @@ class TestAmazonBedrockChatGeneratorUtils:
                         "total_tokens": 516,
                         "cache_read_input_tokens": 0,
                         "cache_write_input_tokens": 0,
-                        "cache_details": {},
+                        "cache_details": [],
                     },
                     "completion_start_time": ANY,
                 },
@@ -1841,7 +1841,7 @@ class TestAmazonBedrockChatGeneratorUtils:
                         "total_tokens": 449,
                         "cache_read_input_tokens": 0,
                         "cache_write_input_tokens": 0,
-                        "cache_details": {},
+                        "cache_details": [],
                     },
                     "completion_start_time": ANY,
                 },
@@ -1929,7 +1929,7 @@ class TestAmazonBedrockChatGeneratorUtils:
                         "total_tokens": 0,
                         "cache_read_input_tokens": 0,
                         "cache_write_input_tokens": 0,
-                        "cache_details": {},
+                        "cache_details": [],
                     },
                     "trace": trace,
                 },
@@ -2296,3 +2296,30 @@ class TestAmazonBedrockChatGeneratorUtils:
         chunk = _convert_event_to_streaming_chunk(event_str, model, component_info)
         assert chunk.tool_calls is not None
         assert chunk.tool_calls[0].arguments == '{"city": "Berlin"}'
+
+    def test_parse_streaming_response_redacted_reasoning_is_not_replayed(self):
+        # OpenAI models on Bedrock (e.g. GPT-6) stream reasoning only as redactedContent
+        model = "us.openai.gpt-6-sol"
+        type_ = (
+            "haystack_integrations.components.generators.amazon_bedrock.chat.chat_generator.AmazonBedrockChatGenerator"
+        )
+        events = [
+            {"messageStart": {"role": "assistant"}},
+            {
+                "contentBlockDelta": {
+                    "delta": {"reasoningContent": {"redactedContent": b"rsn_abc"}},
+                    "contentBlockIndex": 0,
+                }
+            },
+            {"contentBlockStop": {"contentBlockIndex": 0}},
+            {"contentBlockDelta": {"delta": {"text": "No."}, "contentBlockIndex": 1}},
+            {"contentBlockStop": {"contentBlockIndex": 1}},
+            {"messageStop": {"stopReason": "end_turn"}},
+        ]
+
+        replies = _parse_streaming_response(events, lambda chunk: None, model, ComponentInfo(type=type_))
+
+        assert replies[0].text == "No."
+        assert replies[0].reasoning is None
+        formatted_message = _format_messages([replies[0]])[1][0]
+        assert formatted_message == {"role": "assistant", "content": [{"text": "No."}]}