Deepseek

[killrawr]

https://github.com/deepseek-ai/DeepSeek-V3

Are there plans to integrate Deepseek into the LLPhant source?

[josecanciani]

From what I can see, you can already use deepseek-r1 model from Ollama, the only difference with other models is that you get the chain of thought in the reply, between <think></think> tags.

I guess we need Ollama to be updated so the "think" section is returned in a different part of the message, but in the meantime, LLPhant could start refactoring the "generate*Chat" methods to be able to return more than just the main text response.

For now, we can just remove this manually from the response. Here's what I've done:

    private function createMessage(string $response): Message {
        $message = $response;
        $thoughts = '';
        $startOfThoughts = '<think>' . PHP_EOL;
        if (mb_strpos($message, $startOfThoughts) === 0) {
            $endOfThoughts = PHP_EOL . '</think>' . PHP_EOL;
            list ($thoughts, $message) = explode(
                $endOfThoughts,
                mb_substr($message, mb_strlen($startOfThoughts)),
                2
            );
        }
        // I'm using a custom class to store the message, you can just return an array with the two texts
        return Message::createMessage(trim($message), ChatRole::Assistant, $thoughts);
    }

[MaximeThoonsen]

@killrawr the api seems to be compatible with openai. Do you want to make a pull request?

Something like : https://github.com/LLPhant/LLPhant/blob/main/src/Chat/MistralAIChat.php