Add support for DeepSeek language models in STORM Wiki pipeline

[rmcc3]

Description

This pull request adds support for using DeepSeek language models in the STORM Wiki pipeline, providing an alternative to the existing OpenAI models. The integration allows users to easily switch between OpenAI and DeepSeek models, enhancing the flexibility and capabilities of the STORM Wiki system.

Key Changes

1. Implemented a new DeepSeekModel class in src/lm.py, which is compatible with the existing dspy.OpenAI interface.
2. Created a new example script examples/run_storm_wiki_deepseek.py to demonstrate how to use DeepSeek models with the STORM Wiki pipeline.
3. Updated the STORMWikiRunner class to be model-agnostic, ensuring compatibility with both OpenAI and DeepSeek models.
4. Added topic name sanitization to handle special characters in file names, which fixes a bug which would cause the research to crash when using special characters.
5. Add Unicode normalization to handle non-ASCII characters.

How to Use

Users can now run the STORM Wiki pipeline with DeepSeek models by using the new run_storm_wiki_deepseek.py script. The script allows configuration of various DeepSeek-specific parameters such as model choice, temperature, and top_p sampling. The API key should be set with DEEPSEEK_API_KEY. The DeepSeek base can be set with DEEPSEEK_API_BASE, if needed.

Testing

• Tested the integration with DeepSeek models.
• Verified compatibility with existing retrieval methods.
• Ensured proper handling of topic names with special characters.

Future Considerations

• Consider implementing a unified interface for easily switching between different model providers and retrievers.

[shaoyijia]

Hi @rmcc3 , thank you so much for the prompt response! The quality of the example outputs is pretty good - we are very happy to support DeepSeek models in our project. Could you make the following change so that I can merge this PR?

• Resolve the merge conflict. (We recently have breaking change in order to support installing our project via pip.)
• Add a few lines (see my comment) to DeepSeekModel to ensure the call history is tracked. Currently, llm_call_history.jsonl is empty in your shared output. After adding these lines, it shall include call history in the session.

[rmcc3]

Seems updating to head caused this to close. Working on the change now.

[rmcc3]

llm_call_history.txt

Here is the call history (had to upload as .txt because GitHub does not accept .jsonl)

[shaoyijia]

LGTM, thank you so much!!