mirror of
https://github.com/discourse/discourse-ai.git
synced 2025-02-20 18:35:55 +00:00
4f1a3effe0
Both endpoints provide OpenAI-compatible servers. The only difference is that Vllm doesn't support passing tools as a separate parameter. Even if the tool param is supported, it ultimately relies on the model's ability to handle native functions, which is not the case with the models we have today. As a part of this change, we are dropping support for StableBeluga/Llama2 models. They don't have a chat_template, meaning the new API can translate them. These changes let us remove some of our existing dialects and are a first step in our plan to support any LLM by defining them as data-driven concepts. I rewrote the "translate" method to use a template method and extracted the tool support strategies into its classes to simplify the code. Finally, these changes bring support for Ollama when running in dev mode. It only works with Mistral for now, but it will change soon..
128 lines
3.3 KiB
Ruby
128 lines
3.3 KiB
Ruby
# frozen_string_literal: true
|
|
|
|
module DiscourseAi
|
|
module Completions
|
|
module Dialects
|
|
class Gemini < Dialect
|
|
class << self
|
|
def can_translate?(model_name)
|
|
%w[gemini-pro gemini-1.5-pro].include?(model_name)
|
|
end
|
|
|
|
def tokenizer
|
|
DiscourseAi::Tokenizer::OpenAiTokenizer ## TODO Replace with GeminiTokenizer
|
|
end
|
|
end
|
|
|
|
def native_tool_support?
|
|
true
|
|
end
|
|
|
|
def translate
|
|
# Gemini complains if we don't alternate model/user roles.
|
|
noop_model_response = { role: "model", parts: { text: "Ok." } }
|
|
messages = super
|
|
|
|
interleving_messages = []
|
|
previous_message = nil
|
|
|
|
messages.each do |message|
|
|
if previous_message
|
|
if (previous_message[:role] == "user" || previous_message[:role] == "function") &&
|
|
message[:role] == "user"
|
|
interleving_messages << noop_model_response.dup
|
|
end
|
|
end
|
|
interleving_messages << message
|
|
previous_message = message
|
|
end
|
|
|
|
interleving_messages
|
|
end
|
|
|
|
def tools
|
|
return if prompt.tools.blank?
|
|
|
|
translated_tools =
|
|
prompt.tools.map do |t|
|
|
tool = t.slice(:name, :description)
|
|
|
|
if t[:parameters]
|
|
tool[:parameters] = t[:parameters].reduce(
|
|
{ type: "object", required: [], properties: {} },
|
|
) do |memo, p|
|
|
name = p[:name]
|
|
memo[:required] << name if p[:required]
|
|
|
|
memo[:properties][name] = p.except(:name, :required, :item_type)
|
|
|
|
memo[:properties][name][:items] = { type: p[:item_type] } if p[:item_type]
|
|
memo
|
|
end
|
|
end
|
|
|
|
tool
|
|
end
|
|
|
|
[{ function_declarations: translated_tools }]
|
|
end
|
|
|
|
def max_prompt_tokens
|
|
if model_name == "gemini-1.5-pro"
|
|
# technically we support 1 million tokens, but we're being conservative
|
|
800_000
|
|
else
|
|
16_384 # 50% of model tokens
|
|
end
|
|
end
|
|
|
|
protected
|
|
|
|
def calculate_message_token(context)
|
|
self.class.tokenizer.size(context[:content].to_s + context[:name].to_s)
|
|
end
|
|
|
|
def system_msg(msg)
|
|
{ role: "user", parts: { text: msg[:content] } }
|
|
end
|
|
|
|
def model_msg(msg)
|
|
{ role: "model", parts: { text: msg[:content] } }
|
|
end
|
|
|
|
def user_msg(msg)
|
|
{ role: "user", parts: { text: msg[:content] } }
|
|
end
|
|
|
|
def tool_call_msg(msg)
|
|
call_details = JSON.parse(msg[:content], symbolize_names: true)
|
|
|
|
{
|
|
role: "model",
|
|
parts: {
|
|
functionCall: {
|
|
name: msg[:name] || call_details[:name],
|
|
args: call_details[:arguments],
|
|
},
|
|
},
|
|
}
|
|
end
|
|
|
|
def tool_msg(msg)
|
|
{
|
|
role: "function",
|
|
parts: {
|
|
functionResponse: {
|
|
name: msg[:name] || msg[:id],
|
|
response: {
|
|
content: msg[:content],
|
|
},
|
|
},
|
|
},
|
|
}
|
|
end
|
|
end
|
|
end
|
|
end
|
|
end
|