About language

[mc112611]

what language does GraphRAG support? Can I use Japanese document to do RAG？

[eyast]

I believe so. GraphRAG makes calls to an OpenAI or AzureOpenAI LLM, and they support Japanese in their tokenizers. Keep in mind that the instructions (and examples) provided to the LLM to do entity extraction and summarization are all in English. You might want to modify these prompts.

For example, when you intiate the indexer, you will see prompts in prompts\entity_extraction.txt. Line 20 says "Return output in English". You might want to modify this so that the result is returned in Japanese.

[mc112611]

I believe so. GraphRAG makes calls to an OpenAI or AzureOpenAI LLM, and they support Japanese in their tokenizers. Keep in mind that the instructions (and examples) provided to the LLM to do entity extraction and summarization are all in English. You might want to modify these prompts.

For example, when you intiate the indexer, you will see prompts in prompts\entity_extraction.txt. Line 20 says "Return output in English". You might want to modify this so that the result is returned in Japanese.

Thank you so much ， hope this project will support local LLM， not only chatGPT。

[jgbradley1]

Thank you so much ， hope this project will support local LLM， not only chatGPT。

We welcome contributions from the community 😉 .

[natoverse]

Consolidating language support issues here: #696

[ahouse73]

Thank you so much ， hope this project will support local LLM， not only chatGPT。

You can use local LLMs, Ollama supports the OpenAI api specification. If you set the base url of the settings.yaml to the Ollama api url you can use the model.

Example for the LLM:

llm:
  api_key: None
  type: openai_chat 
  model: mistral:latest
  model_supports_json: true
  api_base: http://localhost:11434/v1

[mc112611]

非常感谢，希望这个项目将支持当地的LLM，而不仅仅是chatGPT .

你可以使用本地的LLM，Ollama支持OpenAI api规范。如果您设置了settings.yaml到Ollama api url，您可以使用该模型。

LLM的示例:

llm:
  api_key: None
  type: openai_chat 
  model: mistral:latest
  model_supports_json: true
  api_base: http://localhost:11434/v1

This is very helpful. Thank you