- Implement parallel chunk processing with configurable concurrency
- Add direct NVIDIA API integration bypassing LangChain for better control
- Optimize for DGX Spark unified memory with batch processing
- Use concurrency of 4 for Ollama, 2 for other providers
- Add proper error handling and user stop capability
- Update NVIDIA model to Llama 3.3 Nemotron Super 49B v1.5
- Improve prompt engineering for triple extraction
- Update LangChain service to use Llama 3.3 Nemotron Super 49B v1.5
- Adjust temperature to 0.6 for better response quality
- Increase timeout to 120s for larger model
- Add top_p, frequency_penalty, and presence_penalty parameters
- Remove deprecated response_format configuration
- Add queryWithLLM method to BackendService
- Retrieves top K triples from graph and uses LLM to generate answers
- Supports configurable LLM model and provider selection
- Uses research-backed prompt structure for KG-enhanced RAG
- Includes fallback handling for LLM errors
