This pull request demonstrates how to create a new Opik project when using LiteLLM for LlamaIndex-based LLM calls. The primary goal is to enable detailed tracing, monitoring, and logging of LLM interactions in a new Opik project_name, particularly when using LiteLLM as an API proxy. This enhancement allows for better debugging, performance analysis, observability when using LightRAG with LiteLLM and Opik.
**Motivation:**
As our application's reliance on Large Language Models (LLMs) grows, robust observability becomes crucial for maintaining system health, optimizing performance, and understanding usage patterns. Integrating Opik provides the following key benefits:
1. **Improved Debugging:** Enables end-to-end tracing of requests through the LlamaIndex and LiteLLM layers, making it easier to identify and resolve issues or performance bottlenecks.
2. **Comprehensive Performance Monitoring:** Allows for the collection of vital metrics such as LLM call latency, token usage, and error rates. This data can be filtered and analyzed within Opik using project names and tags.
3. **Effective Cost Management:** Facilitates tracking of token consumption associated with specific requests or projects, leading to better cost control and optimization.
4. **Deeper Usage Insights:** Provides a clearer understanding of how different components of the application or various projects are utilizing LLM capabilities.
These changes empower developers to seamlessly add observability to their LlamaIndex-based LLM workflows, especially when leveraging LiteLLM, by passing necessary Opik metadata.
**Changes Made:**
1. **`lightrag/llm/llama_index_impl.py`:**
* Modified the `llama_index_complete_if_cache` function:
* The `**kwargs` parameter, which previously handled additional arguments, has been refined. A dedicated `chat_kwargs={}` parameter is now used to pass keyword arguments directly to the `model.achat()` method. This change ensures that vendor-specific parameters, such as LiteLLM's `litellm_params` for Opik metadata, are correctly propagated.
* The logic for retrieving `llm_instance` from `kwargs` was removed as `model` is now a direct parameter, simplifying the function.
* Updated the `llama_index_complete` function:
* Ensured that `**kwargs` (which may include `chat_kwargs` or other parameters intended for `llama_index_complete_if_cache`) are correctly passed down.
2. **`examples/unofficial-sample/lightrag_llamaindex_litellm_demo.py`:**
* This existing demo file was updated to align with the changes in `llama_index_impl.py`.
* The `llm_model_func` now passes an empty `chat_kwargs={}` by default to `llama_index_complete_if_cache` if no specific chat arguments are needed, maintaining compatibility with the updated function signature. This file serves as a baseline example without Opik integration.
3. **`examples/unofficial-sample/lightrag_llamaindex_litellm_opik_demo.py` (New File):**
* A new example script has been added to specifically demonstrate the integration of LightRAG with LlamaIndex, LiteLLM, and Opik for observability.
* The `llm_model_func` in this demo showcases how to construct the `chat_kwargs` dictionary.
* It includes `litellm_params` with a `metadata` field for Opik, containing `project_name` and `tags`. This provides a clear example of how to send observability data to Opik.
* The call to `llama_index_complete_if_cache` within `llm_model_func` passes these `chat_kwargs`, ensuring Opik metadata is included in the LiteLLM request.
These modifications provide a more robust and extensible way to pass parameters to the underlying LLM calls, specifically enabling the integration of observability tools like Opik.
Co-authored-by: Martin Perez-Guevara <8766915+MartinPerez@users.noreply.github.com>
Co-authored-by: Young Jin Kim <157011356+jidodata-ykim@users.noreply.github.com>
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.
https://github.com/HKUDS/LightRAG/pull/864#issuecomment-2669705946
- Created two new example scripts demonstrating LightRAG integration with LlamaIndex:
- `lightrag_llamaindex_direct_demo.py`: Direct OpenAI integration
- `lightrag_llamaindex_litellm_demo.py`: LiteLLM proxy integration
- Both examples showcase different search modes (naive, local, global, hybrid)
- Includes configuration for working directory, models, and API settings
- Demonstrates text insertion and querying using LightRAG with LlamaIndex
- removed wrapper directory and references to it
- Implemented LlamaIndex interface for language model interactions
- Added async chat completion support
- Included embedding generation functionality
- Implemented retry mechanisms for API calls
- Added configuration and message formatting utilities
- Supports OpenAI-style message handling and external settings
