Feature Request: Support node_properties/relationship_properties in LLMGraphTransformer for LLMs without native function calling

[dhiaaeddine16]

The LLMGraphTransformer class currently blocks usage of node_properties and relationship_properties parameters when using LLMs that lack native function calling capabilities. This restriction forces users to choose between:

• Using lightweight/OSS LLMs without structured property enforcement, or
• Requiring expensive/closed models with function calling support.

[pbrady]

I believe this would be addressed by #39 .

[dhiaaeddine16]

Issue #39 is indeed useful for models that support function calling since it can help fix some formatting errors when specifying the method for with_structured_output. However, it's important to note that not all models support function calling due to provider constraints. For example, when using ToGher AI with the model meta-llama/Llama-3.3-70B-Instruct-Turbo-Free, you might encounter the following error:

BadRequestError: Error code: 400 - {'id': '', 'error': {'message': 'meta-llama/Llama-3.3-70B-Instruct-Turbo-Free is not supported for JSON mode/function calling', 'type': 'invalid_request_error', 'param': None, 'code': 'constraints_model'}}

Additionally, setting the parameter ignore_tool_usage to False forces the system to avoid using function calling.

Given these constraints, there is a clear need to implement a new prompt or mechanism that can gracefully handle models without function calling support. This new prompt should enable structured outputs without relying on function calling, ensuring compatibility with a broader range of models. Additionally, it must include the extraction of both node_properties and relationship_properties, so that users can still obtain detailed and structured data even when using models that do not natively support function calling.

[dhiaaeddine16]

take a look at: LLMGraphTransformer