Added support for structured output (JSON mode) from the OpenAI API in `openai.py` and `azure_openai.py`.
When `response_format` is used to request structured data, the new logic checks for the `message.parsed` attribute. If it exists, it's serialized into a JSON string as the final content. If not, the code falls back to the existing `message.content` handling, ensuring backward compatibility.
The stream and timeout parameters were moved from **kwargs to explicit
parameters in a previous commit, but were not being passed to the OpenAI
API, causing streaming responses to fail and fall back to non-streaming
mode.Fixes the issue where stream=True was being silently ignored, resulting
in unexpected non-streaming behavior.
This contribution adds optional Langfuse support for LLM observability and tracing.
Langfuse provides a drop-in replacement for the OpenAI client that automatically
tracks all LLM interactions without requiring code changes.
Features:
- Optional Langfuse integration with graceful fallback
- Automatic LLM request/response tracing
- Token usage tracking
- Latency metrics
- Error tracking
- Zero code changes required for existing functionality
Implementation:
- Modified lightrag/llm/openai.py to conditionally use Langfuse's AsyncOpenAI
- Falls back to standard OpenAI client if Langfuse is not installed
- Logs observability status on import
Configuration:
To enable Langfuse tracing, install the observability extras and set environment variables:
```bash
pip install lightrag-hku[observability]
export LANGFUSE_PUBLIC_KEY="your_public_key"
export LANGFUSE_SECRET_KEY="your_secret_key"
export LANGFUSE_HOST="https://cloud.langfuse.com" # or your self-hosted instance
```
If Langfuse is not installed or environment variables are not set, LightRAG
will use the standard OpenAI client without any functionality changes.
Changes:
- Modified lightrag/llm/openai.py (added optional Langfuse import)
- Updated pyproject.toml with optional 'observability' dependencies
Dependencies (optional):
- langfuse>=3.8.1
- Add enable_cot parameter to all LLM APIs
- Implement CoT for OpenAI with <think> tags
- Log warnings for unsupported providers
- Enable CoT in query operations
- Handle streaming and non-streaming CoT
• Remove global temperature parameter
• Add provider-specific temp configs
• Update env example with new settings
• Fix Bedrock temperature handling
• Clean up splash screen display
Replace regex-based JSON extraction with json-repair for better handling of malformed LLM responses. Remove deprecated JSON parsing utilities and clean up keyword_extraction parameter across LLM providers.
- Remove locate_json_string_body_from_string() and convert_response_to_json()
- Use json-repair.loads() in extract_keywords_only() for robust parsing
- Clean up LLM interfaces and remove unused parameters
- Add json-repair dependency
This parameter is no longer used. Its removal simplifies the API and clarifies that token length management is handled by upstream text chunking logic rather than the embedding wrapper.
This avoids issues of resource cleanup (too many open files) when dealing with massively parallel calls to the openai API since RAII in python is highly unreliable in such contexts.
- Corrects the syntax of retry_if_exception_type decorators to ensure proper exception handling and retry behavior
- Implements proper resource cleanup for async iterators to prevent memory leaks and potential SIGSEGV errors
Add support for custom client configurations in the OpenAI integration,
allowing for more flexible configuration of the AsyncOpenAI client.
This includes:
- Create a reusable helper function `create_openai_async_client`
- Add proper documentation for client configuration options
- Ensure consistent parameter precedence across the codebase
- Update the embedding function to support client configurations
- Add example script demonstrating custom client configuration usage
The changes maintain backward compatibility while providing a cleaner
and more maintainable approach to configuring OpenAI clients.
