Add DeepSeek model client

[mxsl-gr]

Hi, this PR is add DeepSeek model client and has passed unit testing.
I can provide my api_key if needed for testing

the PR content:

• DeepSeek chat client
• spring starter
• unit test
• chat client documents

For some reasons, products from OpenAI and others can't be directly used in Chinese Mainland.

DeepSeek is a strong, economical, and efficient Mixture of Experts (MoE) language model, with an API pricing of $0.14/$0.28 per 1 million tokens.

the link: DeepSeek

If they can supported, it will further aid spring-ai to promotion in Chinese Mainland.

If necessary, I can take care of subsequent maintenance since I'm currently using them.

The Moonshot PR is #596
The ZhiPuAI PR is #623
The MiniMax PR is #628

[mxsl-gr]

There are a lot of AI companies, and a lot of AI models, so the code can't be finished by creating clients all the time. Another thing is that deepseek has already said in the document that his API is fully compatible with openAI, that is to say, chatclient set the proxy address can call deepseek smoothly, i think the project maintainer needs to consider this problem: not to add a new model, but to add a new way to call，above are just my own thoughts。

hi, I agree with your point. I have two thoughts.

if the companies provided an API standard compatible with OpenAI, in most cases, you can use it by just replacing the base url and API key, it's great.

while the OpenAI API design is not a standard, there are indeed many followers of its design, especially in mainland China. But this need the followers maintaining compatibility with the OpenAI API design. During the stage when their features are converging, it is predictable as they want to lower the barrier for potential users. But if they continue to develop, their paths may not necessarily converge.
A somewhat inappropriate example is Quarkus and Spring. Quarkus is relatively aggressive in native aspects, but they need to lower the barrier for developers using Spring. So, they provide Spring extensions to work in some cases to achieve low-barrier migration. However, if Spring API gets new designs, Quarkus may not necessarily follow them continuously.

Secondly, the two are not completely compatible. I encountered this with the n parameter. I have a scenario where I need to generate multiple options and then score them, but Deepseek currently does not support the n parameter. I think that if the API does not support it, it’s best not to include it in the configuration. This is something that cannot be achieved if using OpenAI implementation.

in conclusion, I think widely used model providers can still have independent implementations, but it doesn’t hurt to add a new way to call.

[mxsl-gr]

this PR has completed the refactoring, has been squashed and force pushed again.

[markpollack]

In my experience at the model level, I agree that once you get into the details, "api compatability" is surface level. On the other hand, there is indeed a model "zoo" that can add to maintenance and carries risk of them disappearing.

I suspect that over time, once the market decides on winners and losers in the marketplace, there will be consolidation and some models in spring ai will go away. As a concrete example of that, Google's PAaM API is deprecated so we will be removing it (I'll create an issue).

[markpollack]

I'm still not sure what to do here. Other opinions? @mxsl-gr have you tried to use Spring AI's OpenAI support to talk with this model?

[mxsl-gr]

I'm still not sure what to do here. Other opinions? @mxsl-gr have you tried to use Spring AI's OpenAI support to talk with this model?

hi @markpollack , i have been busy with other work recently. we are using a customized version built on my local branch, which merges this branch with other strange features. for example, we have a multi-level model that can handle tasks at different levels to improve speed or reduce costs, as well as a client pool and polling mechanisms to enhance throughput and redundancy.
i just ran the OpenAI model unit tests using DeepSeek's base url and key, and they worked for most scenarios, i think with configuration the OpenAI model can be used in most cases. but there are some properties not supported by DeepSeek, such as the parameter n mentioned earlier. if we ignore those properties, so this PR can be closed.

however, considering that DeepSeek has a significant user base in Mainland China, now even surpassing ZhiPuAI and Minimax. i think having an independent implementation isn't necessarily a bad idea. if that's the case, i might need to adjust this branch, including the implementation of function calls and adding more unit tests.

[markpollack]

Thanks for the update. I did take a look at the github project, it doesn't seem active - https://github.com/deepseek-ai/DeepSeek-V2/graphs/contributors what is your opinion. Does function calling with OpenAI 'compatability' work. In practice, for chat models, I haven't seen much use of 'n' in the wild, say as compared to image models.

[markpollack]

oops, didn't mean to close.

[markpollack]

Also, llama3 has a similar commit history - minimal - I guess all the work is up front! https://github.com/meta-llama/llama3/graphs/contributors

[jingsewu]

Hi, everyone! Do you have any plans for when the branch will be merged into the main branch?

[mxsl-gr]

Hi, everyone! Do you have any plans for when the branch will be merged into the main branch?

hi, for some reason, this PR hasn’t been merged yet.
but I think now is a good timing, our customized version on local branch is still being actively updated. but certain features, like Prefix Completion, are difficult to implement in a non-intrusive way within the current Spring AI project code, I’ll try to simplify and adjust it later.
I also hope it gets merged since DeepSeek is the model I use the most.

[mxsl-gr]

hi, @markpollack

this PR has completed the refactoring, implemented beta features like prefix completion.

the Function Call feature was added at this commit, but is not very stable, because the current version (V2.5 ~ V3) of DeepSeek's modal Function Call is unstable, the tests related to the function call have been temporarily disabled.

All other tests have passed.

References:
DeepSeek Function Call
Chat Prefix Completion

[markpollack]

I’m so very sorry for not being timely. My bad.

Let’s find a time to zoom. I think there is a mixture of approaches, using OpenAI comparability but also allowing for a more dedicated model. This can be something we develop in the spring-ai-community github org. With the GA nearing we need to settle on a supported set of models (we need to support them for commercial support contexts for 3 years or more). A dedicated model can evolve outside out GA timeline and later be incorporated or live side by side as with spring data and and spring cloud community projects. You can commit directly and be in charge without me or anyone else on the team being a blocking factor.

Please email me for a good time to chat.
again , sincere apologies for not relying sooner.

[mxsl-gr]

I’m so very sorry for not being timely. My bad.

Let’s find a time to zoom. I think there is a mixture of approaches, using OpenAI comparability but also allowing for a more dedicated model. This can be something we develop in the spring-ai-community github org. With the GA nearing we need to settle on a supported set of models (we need to support them for commercial support contexts for 3 years or more). A dedicated model can evolve outside out GA timeline and later be incorporated or live side by side as with spring data and and spring cloud community projects. You can commit directly and be in charge without me or anyone else on the team being a blocking factor.

Please email me for a good time to chat. again , sincere apologies for not relying sooner.

hi, @markpollack

i'm sorry for the delayed response, i just got back from a long trip.

i understand the challenge of maintaining multiple models and that you’re suggesting we manage some of them through community projects, similar to how spring-data-elasticsearch relates to spring-data. on that point, I’m fine with the idea.

[eternachen]

Hi, @mxsl-gr @markpollack is there any plan to merge this PR?

[mxsl-gr]

hi, @markpollack

The function call changes and CoT support for deepseek-reasoner have been completed, and all unit tests have passed.

[markpollack]

The PR is basically impossible to rebase as so much has changed. Also note that autoconfig now has a different organization.

I've extracted the files from the PR and merged the core classes into a new module spring-ai-deepseek.

See a5920a1

Please can you make a new PR with the remaining parts from this PR that need to be added?

[mxsl-gr]

The PR is basically impossible to rebase as so much has changed. Also note that autoconfig now has a different organization.

I've extracted the files from the PR and merged the core classes into a new module spring-ai-deepseek.

See a5920a1

Please can you make a new PR with the remaining parts from this PR that need to be added?

Sorry, I missed the changes in the project structure. I’ll take care of it today.