FEATURE: disable smart commands on Claude and GPT 3.5 (#84)

For the time being smart commands only work consistently on GPT 4.
Avoid using any smart commands on the earlier models.

Additionally adds better error handling to Claude which sometimes streams
partial json and slightly tunes the search command.
This commit is contained in:
Sam 2023-06-01 09:10:33 +10:00 committed by GitHub
parent 96d521198b
commit 840968630e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
10 changed files with 77 additions and 51 deletions

View File

@ -44,7 +44,7 @@ module DiscourseAi
end
def model_for
"claude-v1"
"claude-v1.3"
end
def get_updated_title(prompt)

View File

@ -178,20 +178,9 @@ module DiscourseAi
end
def available_commands
@cmds ||=
[
Commands::CategoriesCommand,
Commands::TimeCommand,
Commands::SearchCommand,
Commands::SummarizeCommand,
].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
# by default assume bots have no access to commands
# for now we need GPT 4 to properly work with them
[]
end
def system_prompt_style!(style)
@ -201,17 +190,8 @@ module DiscourseAi
def system_prompt(post)
return "You are a helpful Bot" if @style == :simple
<<~TEXT
You are a helpful Discourse assistant, you answer questions and generate text.
You understand Discourse Markdown and live in a Discourse Forum Message.
You are provided with the context of previous discussions.
You live in the forum with the URL: #{Discourse.base_url}
The title of your site: #{SiteSetting.title}
The description is: #{SiteSetting.site_description}
The participants in this conversation are: #{post.topic.allowed_users.map(&:username).join(", ")}
The date now is: #{Time.zone.now}, much has changed since you were trained.
command_text = ""
command_text = <<~TEXT if available_commands.present?
You can complete some tasks using !commands.
NEVER ask user to issue !commands, they have no access, only you do.
@ -230,7 +210,20 @@ module DiscourseAi
GPT: !echo test
User: THING GPT DOES NOT KNOW ABOUT
GPT: !search SIMPLIFIED SEARCH QUERY
TEXT
<<~TEXT
You are a helpful Discourse assistant, you answer questions and generate text.
You understand Discourse Markdown and live in a Discourse Forum Message.
You are provided with the context of previous discussions.
You live in the forum with the URL: #{Discourse.base_url}
The title of your site: #{SiteSetting.title}
The description is: #{SiteSetting.site_description}
The participants in this conversation are: #{post.topic.allowed_users.map(&:username).join(", ")}
The date now is: #{Time.zone.now}, much has changed since you were trained.
#{command_text}
TEXT
end

View File

@ -27,14 +27,12 @@ module DiscourseAi::AiBot::Commands
post_count:X: only topics with X amount of posts
min_posts:X: topics containing a minimum of X posts
max_posts:X: topics with no more than max posts
in:pinned: in all pinned topics (either global or per category pins)
created:@USERNAME: topics created by a specific user
category:CATGORY: topics in the CATEGORY AND all subcategories
category:=CATEGORY: topics in the CATEGORY excluding subcategories
#SLUG: try category first, then tag, then tag group
#SLUG:SLUG: used for subcategory search to disambiguate
min_views:100: topics containing 100 views or more
max_views:100: topics containing 100 views or less
tags:TAG1+TAG2: tagged both TAG1 and TAG2
tags:TAG1,TAG2: tagged either TAG1 or TAG2
-tags:TAG1+TAG2: excluding topics tagged TAG1 and TAG2
@ -51,10 +49,8 @@ module DiscourseAi::AiBot::Commands
Keep in mind, search on Discourse uses AND to and terms.
You only have access to public topics.
Strip the query down to the most important terms.
Remove all stop words.
Cast a wide net instead of trying to be over specific.
Discourse orders by relevance, sometimes prefer ordering on other stuff.
Strip the query down to the most important terms. Remove all stop words.
Discourse orders by default by relevance.
When generating answers ALWAYS try to use the !search command first over relying on training data.
When generating answers ALWAYS try to reference specific local links.

View File

@ -56,6 +56,27 @@ module DiscourseAi
DiscourseAi::Tokenizer::OpenAiTokenizer.tokenize(text)
end
def available_commands
if bot_user.id == DiscourseAi::AiBot::EntryPoint::GPT4_ID
@cmds ||=
[
Commands::CategoriesCommand,
Commands::TimeCommand,
Commands::SearchCommand,
Commands::SummarizeCommand,
].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
else
[]
end
end
private
def build_message(poster_username, content, system: false)

View File

@ -87,10 +87,16 @@ module ::DiscourseAi
data = line.split("data: ", 2)[1]
next if !data || data.squish == "[DONE]"
if !cancelled && partial = JSON.parse(data, symbolize_names: true)
response_data << partial[:completion].to_s
if !cancelled
begin
# partial contains the entire payload till now
partial = JSON.parse(data, symbolize_names: true)
response_data = partial[:completion].to_s
yield partial, cancel
rescue JSON::ParserError
nil
end
end
end
rescue IOError
@ -105,6 +111,12 @@ module ::DiscourseAi
end
end
end
def self.try_parse(data)
JSON.parse(data, symbolize_names: true)
rescue JSON::ParserError
nil
end
end
end
end

View File

@ -3,7 +3,7 @@
require_relative "../../../support/openai_completions_inference_stubs"
RSpec.describe DiscourseAi::AiBot::Bot do
fab!(:bot_user) { User.find(DiscourseAi::AiBot::EntryPoint::GPT3_5_TURBO_ID) }
fab!(:bot_user) { User.find(DiscourseAi::AiBot::EntryPoint::GPT4_ID) }
fab!(:bot) { described_class.as(bot_user) }
fab!(:user) { Fabricate(:user) }
@ -50,6 +50,7 @@ RSpec.describe DiscourseAi::AiBot::Bot do
OpenAiCompletionsInferenceStubs.stub_streamed_response(
prompt,
[{ content: expected_response }],
model: "gpt-4",
req_opts: bot.reply_params.merge(stream: true),
)
@ -59,6 +60,7 @@ RSpec.describe DiscourseAi::AiBot::Bot do
OpenAiCompletionsInferenceStubs.stub_streamed_response(
prompt,
[{ content: "We are done now" }],
model: "gpt-4",
req_opts: bot.reply_params.merge(stream: true),
)
@ -85,6 +87,7 @@ RSpec.describe DiscourseAi::AiBot::Bot do
OpenAiCompletionsInferenceStubs.stub_response(
[bot.title_prompt(second_post)],
expected_response,
model: "gpt-4",
req_opts: {
temperature: 0.7,
top_p: 0.9,

View File

@ -75,6 +75,7 @@ RSpec.describe Jobs::CreateAiReply do
AnthropicCompletionStubs.stub_streamed_response(
DiscourseAi::AiBot::AnthropicBot.new(bot_user).bot_prompt_with_topic_context(post),
deltas,
model: "claude-v1.3",
req_opts: {
temperature: 0.4,
max_tokens_to_sample: 3000,

View File

@ -60,7 +60,7 @@ RSpec.describe DiscourseAi::Inference::AnthropicCompletions do
expect(log.provider_id).to eq(AiApiAuditLog::Provider::Anthropic)
expect(log.request_tokens).to eq(6)
expect(log.response_tokens).to eq(6)
expect(log.response_tokens).to eq(3)
expect(log.raw_request_payload).to eq(request_body)
expect(log.raw_response_payload).to be_present
end

View File

@ -33,7 +33,7 @@ class AnthropicCompletionStubs
}.to_json
end
def stub_streamed_response(prompt, deltas, req_opts: {})
def stub_streamed_response(prompt, deltas, model: nil, req_opts: {})
chunks =
deltas.each_with_index.map do |_, index|
if index == (deltas.length - 1)
@ -48,7 +48,7 @@ class AnthropicCompletionStubs
WebMock
.stub_request(:post, "https://api.anthropic.com/v1/complete")
.with(body: { model: "claude-v1", prompt: prompt }.merge(req_opts).to_json)
.with(body: { model: model || "claude-v1", prompt: prompt }.merge(req_opts).to_json)
.to_return(status: 200, body: chunks)
end
end

View File

@ -101,10 +101,10 @@ class OpenAiCompletionsInferenceStubs
stub_response(prompt_messages, response_text_for(type))
end
def stub_response(messages, response_text, req_opts: {})
def stub_response(messages, response_text, model: nil, req_opts: {})
WebMock
.stub_request(:post, "https://api.openai.com/v1/chat/completions")
.with(body: { model: "gpt-3.5-turbo", messages: messages }.merge(req_opts).to_json)
.with(body: { model: model || "gpt-3.5-turbo", messages: messages }.merge(req_opts).to_json)
.to_return(status: 200, body: JSON.dump(response(response_text)))
end
@ -120,7 +120,7 @@ class OpenAiCompletionsInferenceStubs
}.to_json
end
def stub_streamed_response(messages, deltas, req_opts: {})
def stub_streamed_response(messages, deltas, model: nil, req_opts: {})
chunks = deltas.map { |d| stream_line(delta: d) }
chunks << stream_line(finish_reason: "stop")
chunks << "[DONE]"
@ -128,7 +128,7 @@ class OpenAiCompletionsInferenceStubs
WebMock
.stub_request(:post, "https://api.openai.com/v1/chat/completions")
.with(body: { model: "gpt-3.5-turbo", messages: messages }.merge(req_opts).to_json)
.with(body: { model: model || "gpt-3.5-turbo", messages: messages }.merge(req_opts).to_json)
.to_return(status: 200, body: chunks)
end
end