vrr/open-notebook
2
0
Fork 0
mirror of https://github.com/lfnovo/open-notebook.git synced 2026-04-29 12:00:00 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

fix: strip <think> tags from chat responses

Browse source 
Add thinking content cleaning to notebook and source chat graphs.
Previously, models that output <think>...</think> tags (like DeepSeek)
or malformed variants without opening tags (like Nemotron) would leak
reasoning content into user-visible responses.

Changes:
- chat.py: Clean AI response content before returning messages
- source_chat.py: Same fix for source-specific chat
- text_utils.py: Handle malformed output where opening <think> tag
  is missing but </think> is present

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
Justin Florentine 2025-12-18 16:31:23 -05:00
parent 65166d4d2a
commit 869664a10b
No known key found for this signature in database
3 changed files with 45 additions and 15 deletions
Download patch file Download diff file Expand all files Collapse all files

37
open_notebook/utils/text_utils.py
View file

@ -11,8 +11,11 @@ from langchain_text_splitters import RecursiveCharacterTextSplitter
from .token_utils import token_count
# Pattern for matching thinking content in AI responses
# Patterns for matching thinking content in AI responses
# Standard pattern: <think>...</think>
THINK_PATTERN = re.compile(r"<think>(.*?)</think>", re.DOTALL)
# Pattern for malformed output: content</think> (missing opening tag)
THINK_PATTERN_NO_OPEN = re.compile(r"^(.*?)</think>", re.DOTALL)
def split_text(txt: str, chunk_size=500):
@ -77,6 +80,9 @@ def parse_thinking_content(content: str) -> Tuple[str, str]:
"""
Parse message content to extract thinking content from <think> tags.
Handles both well-formed tags and malformed output where the opening
<think> tag is missing but </think> is present.
Args:
content (str): The original message content
@ -101,22 +107,31 @@ def parse_thinking_content(content: str) -> Tuple[str, str]:
if len(content) > 100000:
return "", content
# Find all thinking blocks
# Find all well-formed thinking blocks
thinking_matches = THINK_PATTERN.findall(content)
if not thinking_matches:
return "", content
if thinking_matches:
# Join all thinking content with double newlines
thinking_content = "\n\n".join(match.strip() for match in thinking_matches)
# Join all thinking content with double newlines
thinking_content = "\n\n".join(match.strip() for match in thinking_matches)
# Remove all <think>...</think> blocks from the original content
cleaned_content = THINK_PATTERN.sub("", content)
# Remove all <think>...</think> blocks from the original content
cleaned_content = THINK_PATTERN.sub("", content)
# Clean up extra whitespace
cleaned_content = re.sub(r"\n\s*\n\s*\n", "\n\n", cleaned_content).strip()
# Clean up extra whitespace
cleaned_content = re.sub(r"\n\s*\n\s*\n", "\n\n", cleaned_content).strip()
return thinking_content, cleaned_content
return thinking_content, cleaned_content
# Handle malformed output: content</think> (missing opening tag)
# Some models like Nemotron output thinking without the opening <think> tag
malformed_match = THINK_PATTERN_NO_OPEN.match(content)
if malformed_match:
thinking_content = malformed_match.group(1).strip()
# Remove the thinking content and </think> tag
cleaned_content = content[malformed_match.end():].strip()
return thinking_content, cleaned_content
return "", content
def clean_thinking_content(content: str) -> str: