discourse-ai/lib/modules/ai_bot/open_ai_bot.rb

176 lines
5.2 KiB
Ruby
Raw Normal View History

# frozen_string_literal: true
module DiscourseAi
module AiBot
class OpenAiBot < Bot
def self.can_reply_as?(bot_user)
open_ai_bot_ids = [
DiscourseAi::AiBot::EntryPoint::GPT4_ID,
DiscourseAi::AiBot::EntryPoint::GPT3_5_TURBO_ID,
]
open_ai_bot_ids.include?(bot_user.id)
end
def prompt_limit
# note GPT counts both reply and request tokens in limits...
# also allow for an extra 500 or so spare tokens
#
# 2500 are the max reply tokens
# Then we have 450 or so for the full function suite
# 100 additional for growth around function calls
if bot_user.id == DiscourseAi::AiBot::EntryPoint::GPT4_ID
8192 - 3050
else
16_384 - 3050
end
end
def reply_params
# technically we could allow GPT-3.5 16k more tokens
# but lets just keep it here for now
{ temperature: 0.4, top_p: 0.9, max_tokens: 2500 }
end
def submit_prompt(
prompt,
prefer_low_cost: false,
temperature: nil,
top_p: nil,
max_tokens: nil,
&blk
)
params =
reply_params.merge(
temperature: temperature,
top_p: top_p,
max_tokens: max_tokens,
) { |key, old_value, new_value| new_value.nil? ? old_value : new_value }
model = model_for(low_cost: prefer_low_cost)
params[:functions] = available_functions if available_functions.present?
DiscourseAi::Inference::OpenAiCompletions.perform!(prompt, model, **params, &blk)
end
def tokenize(text)
DiscourseAi::Tokenizer::OpenAiTokenizer.tokenize(text)
end
def available_functions
# note if defined? can be a problem in test
# this can never be nil so it is safe
return @available_functions if @available_functions
functions = []
functions =
available_commands.map do |command|
function =
DiscourseAi::Inference::OpenAiCompletions::Function.new(
name: command.name,
description: command.desc,
)
command.parameters.each do |parameter|
function.add_parameter(
name: parameter.name,
type: parameter.type,
description: parameter.description,
required: parameter.required,
)
end
function
end
@available_functions = functions
end
def available_commands
return @cmds if @cmds
all_commands =
[
Commands::CategoriesCommand,
Commands::TimeCommand,
Commands::SearchCommand,
Commands::SummarizeCommand,
Commands::ReadCommand,
].tap do |cmds|
cmds << Commands::TagsCommand if SiteSetting.tagging_enabled
cmds << Commands::ImageCommand if SiteSetting.ai_stability_api_key.present?
if SiteSetting.ai_google_custom_search_api_key.present? &&
SiteSetting.ai_google_custom_search_cx.present?
cmds << Commands::GoogleCommand
end
end
allowed_commands = SiteSetting.ai_bot_enabled_chat_commands.split("|")
@cmds = all_commands.filter { |klass| allowed_commands.include?(klass.name) }
end
def model_for(low_cost: false)
return "gpt-4" if bot_user.id == DiscourseAi::AiBot::EntryPoint::GPT4_ID && !low_cost
"gpt-3.5-turbo-16k"
end
def clean_username(username)
if username.match?(/\0[a-zA-Z0-9_-]{1,64}\z/)
username
else
# not the best in the world, but this is what we have to work with
# if sites enable unicode usernames this can get messy
username.gsub(/[^a-zA-Z0-9_-]/, "_")[0..63]
end
end
private
def populate_functions(partial, functions)
fn = partial.dig(:choices, 0, :delta, :function_call)
if fn
functions.add_function(fn[:name]) if fn[:name].present?
functions.add_argument_fragment(fn[:arguments]) if fn[:arguments].present?
end
end
def build_message(poster_username, content, function: false, system: false)
is_bot = poster_username == bot_user.username
if function
role = "function"
elsif system
role = "system"
else
role = is_bot ? "assistant" : "user"
end
result = { role: role, content: content }
if function
result[:name] = poster_username
elsif !system && poster_username != bot_user.username && poster_username.present?
# Open AI restrict name to 64 chars and only A-Za-z._ (work around)
result[:name] = clean_username(poster_username)
end
result
end
def get_delta(partial, _context)
partial.dig(:choices, 0, :delta, :content).to_s
end
def get_updated_title(prompt)
DiscourseAi::Inference::OpenAiCompletions.perform!(
prompt,
model_for,
temperature: 0.7,
top_p: 0.9,
max_tokens: 40,
).dig(:choices, 0, :message, :content)
end
end
end
end