fix(chatbot): handle structured/list message content from newer LLM p…

services/chatbot/src/chatbot/agent_utils.py

@@ -15,6 +15,31 @@
 CHARS_PER_TOKEN = 4
 
 
+def message_content_to_text(content) -> str:
+    """Normalize an LLM message's ``content`` to plain text.
+
+    The OpenAI provider and older models return ``content`` as a ``str``, but
+    newer models (e.g. recent Anthropic Claude models) return it as a list of
+    structured content blocks (text / thinking / tool-use / signed text).
+    Downstream the value is stored in the Mongo chat history and passed to
+    ``Document(page_content=...)``, which requires a ``str`` -- a list makes it
+    raise ``pydantic.ValidationError`` and the whole turn is lost (chat appears
+    to fail and nothing is persisted). Join the text blocks and drop the rest
+    so output from any provider is handled.
+    """
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        parts = []
+        for block in content:
+            if isinstance(block, str):
+                parts.append(block)
+            elif isinstance(block, dict) and block.get("type") == "text":
+                parts.append(block.get("text", ""))
+        return "".join(parts)
+    return "" if content is None else str(content)
+
+
 def collect_long_strings(obj):
     field_info = []
 

services/chatbot/src/chatbot/chat_service.py

@@ -3,7 +3,7 @@
 
 from langgraph.graph.message import Messages
 
-from .agent_utils import trim_messages_to_token_limit
+from .agent_utils import message_content_to_text, trim_messages_to_token_limit
 from .config import Config
 from .extensions import db
 from .langgraph_agent import execute_langgraph_agent
@@ -64,9 +64,12 @@ async def process_user_message(session_id, user_message, api_key, model_name, us
     logger.debug("Response messages count: %d", len(response.get("messages", [])))
 
     reply: Messages = response.get("messages", [{}])[-1]
+    # Newer models can return `content` as a list of structured blocks rather
+    # than a plain string; normalize it before storing / serializing it.
+    reply_text = message_content_to_text(reply.content)
     response_message_id = uuid4().int & (1 << 63) - 1
     history.append(
-        {"id": response_message_id, "role": "assistant", "content": reply.content}
+        {"id": response_message_id, "role": "assistant", "content": reply_text}
     )
     logger.debug(
         "Added assistant response to history - message_id: %s", response_message_id
@@ -75,7 +78,7 @@ async def process_user_message(session_id, user_message, api_key, model_name, us
     add_to_chroma_collection(
         api_key,
         session_id,
-        [{"user": user_message}, {"assistant": reply.content}],
+        [{"user": user_message}, {"assistant": reply_text}],
         Config.LLM_PROVIDER,
         model_name,
     )
@@ -89,4 +92,4 @@ async def process_user_message(session_id, user_message, api_key, model_name, us
         response_message_id,
         len(history),
     )
-    return reply.content, response_message_id
+    return reply_text, response_message_id

services/chatbot/src/chatbot/tests/test_agent_utils.py

@@ -50,6 +50,7 @@
     CHARS_PER_TOKEN,
     _estimate_tokens,
     _message_content,
+    message_content_to_text,
     trim_messages_to_token_limit,
 )
 
@@ -190,3 +191,44 @@ def test_does_not_mutate_original(self):
         original_len = len(msgs)
         trim_messages_to_token_limit(msgs)
         assert len(msgs) == original_len
+
+
+# ---------------------------------------------------------------------------
+# message_content_to_text
+# ---------------------------------------------------------------------------
+
+class TestMessageContentToText:
+    def test_plain_string_passthrough(self):
+        assert message_content_to_text("hello") == "hello"
+
+    def test_list_of_text_blocks(self):
+        # Newer Anthropic models return signed text blocks as a list.
+        content = [{"type": "text", "text": "OK", "signature": "CAQS..."}]
+        assert message_content_to_text(content) == "OK"
+
+    def test_skips_non_text_blocks(self):
+        # thinking / tool_use blocks must be dropped, text kept and joined.
+        content = [
+            {"type": "thinking", "thinking": "...", "signature": "x"},
+            {"type": "tool_use", "name": "sql", "input": {}},
+            {"type": "text", "text": "Answer"},
+        ]
+        assert message_content_to_text(content) == "Answer"
+
+    def test_joins_multiple_text_blocks(self):
+        content = [{"type": "text", "text": "a"}, {"type": "text", "text": "b"}]
+        assert message_content_to_text(content) == "ab"
+
+    def test_list_with_bare_strings(self):
+        assert message_content_to_text(["a", "b"]) == "ab"
+
+    def test_none_returns_empty_string(self):
+        assert message_content_to_text(None) == ""
+
+    def test_other_types_coerced_to_str(self):
+        assert message_content_to_text(123) == "123"
+
+    def test_result_is_always_a_string(self):
+        # The contract the callers rely on (Document(page_content=...) needs str).
+        for value in ["x", [{"type": "text", "text": "y"}], None, 1, []]:
+            assert isinstance(message_content_to_text(value), str)