fix: handle structured content format in LLM response parsing

Some LLM providers (notably Gemini, DeepSeek via OpenAI-compatible
proxies) return ai_message.content as a list of content parts:
[{'type': 'text', 'text': '...', 'extras': {...}}]

The current code uses str() on non-string content, which produces the
Python repr of the entire list — not valid JSON. This breaks
PydanticOutputParser.parse() with OutputParserException.

This adds extract_text_content() to properly unwrap text from both
string and structured content formats, applied in ask.py, chat.py,
and prompt.py.

Fixes #329

open_notebook/graphs/ask.py

@@ -12,6 +12,7 @@ from typing_extensions import TypedDict
 from open_notebook.ai.provision import provision_langchain_model
 from open_notebook.domain.notebook import vector_search
 from open_notebook.utils import clean_thinking_content
+from open_notebook.utils.text_utils import extract_text_content
 
 
 class SubGraphState(TypedDict):
@@ -62,11 +63,7 @@ async def call_model_with_messages(state: ThreadState, config: RunnableConfig) -
     ai_message = await model.ainvoke(system_prompt)
 
     # Clean the thinking content from the response
-    message_content = (
-        ai_message.content
-        if isinstance(ai_message.content, str)
-        else str(ai_message.content)
-    )
+    message_content = extract_text_content(ai_message.content)
     cleaned_content = clean_thinking_content(message_content)
 
     # Parse the cleaned JSON content
@@ -109,11 +106,7 @@ async def provide_answer(state: SubGraphState, config: RunnableConfig) -> dict:
         max_tokens=2000,
     )
     ai_message = await model.ainvoke(system_prompt)
-    ai_content = (
-        ai_message.content
-        if isinstance(ai_message.content, str)
-        else str(ai_message.content)
-    )
+    ai_content = extract_text_content(ai_message.content)
     return {"answers": [clean_thinking_content(ai_content)]}
 
 
@@ -126,11 +119,7 @@ async def write_final_answer(state: ThreadState, config: RunnableConfig) -> dict
         max_tokens=2000,
     )
     ai_message = await model.ainvoke(system_prompt)
-    final_content = (
-        ai_message.content
-        if isinstance(ai_message.content, str)
-        else str(ai_message.content)
-    )
+    final_content = extract_text_content(ai_message.content)
     return {"final_answer": clean_thinking_content(final_content)}
 
 

open_notebook/graphs/chat.py

@@ -14,6 +14,7 @@ from open_notebook.ai.provision import provision_langchain_model
 from open_notebook.config import LANGGRAPH_CHECKPOINT_FILE
 from open_notebook.domain.notebook import Notebook
 from open_notebook.utils import clean_thinking_content
+from open_notebook.utils.text_utils import extract_text_content
 
 
 class ThreadState(TypedDict):
@@ -69,11 +70,7 @@ def call_model_with_messages(state: ThreadState, config: RunnableConfig) -> dict
     ai_message = model.invoke(payload)
 
     # Clean thinking content from AI response (e.g., <think>...</think> tags)
-    content = (
-        ai_message.content
-        if isinstance(ai_message.content, str)
-        else str(ai_message.content)
-    )
+    content = extract_text_content(ai_message.content)
     cleaned_content = clean_thinking_content(content)
     cleaned_message = ai_message.model_copy(update={"content": cleaned_content})
 

open_notebook/graphs/prompt.py

@@ -7,7 +7,7 @@ from langgraph.graph import END, START, StateGraph
 from typing_extensions import TypedDict
 
 from open_notebook.ai.provision import provision_langchain_model
-from open_notebook.utils.text_utils import clean_thinking_content
+from open_notebook.utils.text_utils import clean_thinking_content, extract_text_content
 
 
 class PatternChainState(TypedDict):
@@ -33,7 +33,7 @@ async def call_model(state: dict, config: RunnableConfig) -> dict:
     response = await chain.ainvoke(payload)
 
     # Clean thinking tags from response (handles extended thinking models)
-    output = clean_thinking_content(str(response.content))
+    output = clean_thinking_content(extract_text_content(response.content))
     return {"output": output}
 
 

open_notebook/utils/text_utils.py

@@ -117,3 +117,29 @@ def clean_thinking_content(content: str) -> str:
     """
     _, cleaned_content = parse_thinking_content(content)
     return cleaned_content
+
+
+def extract_text_content(content) -> str:
+    """Extract text from LLM response content.
+
+    Handles both plain string responses and structured content formats
+    (e.g. Gemini's envelope format):
+    [{'type': 'text', 'text': '...', 'extras': {...}}]
+
+    Args:
+        content: The content from an AI message, either a string or a list of parts.
+
+    Returns:
+        The extracted text content as a string.
+    """
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        text_parts = []
+        for part in content:
+            if isinstance(part, dict) and "text" in part:
+                text_parts.append(part["text"])
+            elif isinstance(part, str):
+                text_parts.append(part)
+        return "".join(text_parts)
+    return str(content)