Sam e817b7dc11
FEATURE: improve tool support (#904)
This re-implements tool support in DiscourseAi::Completions::Llm #generate

Previously tool support was always returned via XML and it would be the responsibility of the caller to parse XML

New implementation has the endpoints return ToolCall objects.

Additionally this simplifies the Llm endpoint interface and gives it more clarity. Llms must implement

decode, decode_chunk (for streaming)

It is the implementers responsibility to figure out how to decode chunks, base no longer implements. To make this easy we ship a flexible json decoder which is easy to wire up.

Also (new)

    Better debugging for PMs, we now have a next / previous button to see all the Llm messages associated with a PM
    Token accounting is fixed for vllm (we were not correctly counting tokens)
2024-11-12 08:14:30 +11:00

365 lines
10 KiB
Ruby

#frozen_string_literal: true
module DiscourseAi
module AiBot
module Personas
class Persona
class << self
def rag_conversation_chunks
10
end
def vision_enabled
false
end
def vision_max_pixels
1_048_576
end
def question_consolidator_llm
nil
end
def force_default_llm
false
end
def allow_chat_channel_mentions
false
end
def allow_chat_direct_messages
false
end
def system_personas
@system_personas ||= {
Personas::General => -1,
Personas::SqlHelper => -2,
Personas::Artist => -3,
Personas::SettingsExplorer => -4,
Personas::Researcher => -5,
Personas::Creative => -6,
Personas::DallE3 => -7,
Personas::DiscourseHelper => -8,
Personas::GithubHelper => -9,
}
end
def system_personas_by_id
@system_personas_by_id ||= system_personas.invert
end
def all(user:)
# listing tools has to be dynamic cause site settings may change
AiPersona.all_personas.filter do |persona|
next false if !user.in_any_groups?(persona.allowed_group_ids)
if persona.system
instance = persona.new
(
instance.required_tools == [] ||
(instance.required_tools - all_available_tools).empty?
)
else
true
end
end
end
def find_by(id: nil, name: nil, user:)
all(user: user).find { |persona| persona.id == id || persona.name == name }
end
def name
I18n.t("discourse_ai.ai_bot.personas.#{to_s.demodulize.underscore}.name")
end
def description
I18n.t("discourse_ai.ai_bot.personas.#{to_s.demodulize.underscore}.description")
end
def all_available_tools
tools = [
Tools::ListCategories,
Tools::Time,
Tools::Search,
Tools::Read,
Tools::DbSchema,
Tools::SearchSettings,
Tools::SettingContext,
Tools::RandomPicker,
Tools::DiscourseMetaSearch,
Tools::GithubFileContent,
Tools::GithubPullRequestDiff,
Tools::GithubSearchFiles,
Tools::WebBrowser,
Tools::JavascriptEvaluator,
]
tools << Tools::GithubSearchCode if SiteSetting.ai_bot_github_access_token.present?
tools << Tools::ListTags if SiteSetting.tagging_enabled
tools << Tools::Image if SiteSetting.ai_stability_api_key.present?
tools << Tools::DallE if SiteSetting.ai_openai_api_key.present?
if SiteSetting.ai_google_custom_search_api_key.present? &&
SiteSetting.ai_google_custom_search_cx.present?
tools << Tools::Google
end
tools
end
end
def id
@ai_persona&.id || self.class.system_personas[self.class]
end
def tools
[]
end
def force_tool_use
[]
end
def forced_tool_count
-1
end
def required_tools
[]
end
def temperature
nil
end
def top_p
nil
end
def options
{}
end
def available_tools
self
.class
.all_available_tools
.filter { |tool| tools.include?(tool) }
.concat(tools.filter(&:custom?))
end
def craft_prompt(context, llm: nil)
system_insts =
system_prompt.gsub(/\{(\w+)\}/) do |match|
found = context[match[1..-2].to_sym]
found.nil? ? match : found.to_s
end
prompt_insts = <<~TEXT.strip
#{system_insts}
#{available_tools.map(&:custom_system_message).compact_blank.join("\n")}
TEXT
question_consolidator_llm = llm
if self.class.question_consolidator_llm.present?
question_consolidator_llm =
DiscourseAi::Completions::Llm.proxy(self.class.question_consolidator_llm)
end
if context[:custom_instructions].present?
prompt_insts << "\n"
prompt_insts << context[:custom_instructions]
end
fragments_guidance =
rag_fragments_prompt(
context[:conversation_context].to_a,
llm: question_consolidator_llm,
user: context[:user],
)&.strip
prompt_insts << fragments_guidance if fragments_guidance.present?
prompt =
DiscourseAi::Completions::Prompt.new(
prompt_insts,
messages: context[:conversation_context].to_a,
topic_id: context[:topic_id],
post_id: context[:post_id],
)
prompt.max_pixels = self.class.vision_max_pixels if self.class.vision_enabled
prompt.tools = available_tools.map(&:signature) if available_tools
prompt
end
def find_tool(partial, bot_user:, llm:, context:)
return nil if !partial.is_a?(DiscourseAi::Completions::ToolCall)
tool_instance(partial, bot_user: bot_user, llm: llm, context: context)
end
protected
def tool_instance(tool_call, bot_user:, llm:, context:)
function_id = tool_call.id
function_name = tool_call.name
return nil if function_name.nil?
tool_klass = available_tools.find { |c| c.signature.dig(:name) == function_name }
return nil if tool_klass.nil?
arguments = {}
tool_klass.signature[:parameters].to_a.each do |param|
name = param[:name]
value = tool_call.parameters[name.to_sym]
if param[:type] == "array" && value
value =
begin
JSON.parse(value)
rescue JSON::ParserError
[value.to_s]
end
elsif param[:type] == "string" && value
value = strip_quotes(value).to_s
elsif param[:type] == "integer" && value
value = strip_quotes(value).to_i
end
if param[:enum] && value && !param[:enum].include?(value)
# invalid enum value
value = nil
end
arguments[name.to_sym] = value if value
end
tool_klass.new(
arguments,
tool_call_id: function_id || function_name,
persona_options: options[tool_klass].to_h,
bot_user: bot_user,
llm: llm,
context: context,
)
end
def strip_quotes(value)
if value.is_a?(String)
if value.start_with?('"') && value.end_with?('"')
value = value[1..-2]
elsif value.start_with?("'") && value.end_with?("'")
value = value[1..-2]
else
value
end
else
value
end
end
def rag_fragments_prompt(conversation_context, llm:, user:)
upload_refs =
UploadReference.where(target_id: id, target_type: "AiPersona").pluck(:upload_id)
return nil if !SiteSetting.ai_embeddings_enabled?
return nil if conversation_context.blank? || upload_refs.blank?
latest_interactions =
conversation_context.select { |ctx| %i[model user].include?(ctx[:type]) }.last(10)
return nil if latest_interactions.empty?
# first response
if latest_interactions.length == 1
consolidated_question = latest_interactions[0][:content]
else
consolidated_question =
DiscourseAi::AiBot::QuestionConsolidator.consolidate_question(
llm,
latest_interactions,
user,
)
end
return nil if !consolidated_question
strategy = DiscourseAi::Embeddings::Strategies::Truncation.new
vector_rep =
DiscourseAi::Embeddings::VectorRepresentations::Base.current_representation(strategy)
reranker = DiscourseAi::Inference::HuggingFaceTextEmbeddings
interactions_vector = vector_rep.vector_from(consolidated_question)
rag_conversation_chunks = self.class.rag_conversation_chunks
candidate_fragment_ids =
vector_rep.asymmetric_rag_fragment_similarity_search(
interactions_vector,
target_type: "AiPersona",
target_id: id,
limit:
(
if reranker.reranker_configured?
rag_conversation_chunks * 5
else
rag_conversation_chunks
end
),
offset: 0,
)
fragments =
RagDocumentFragment.where(upload_id: upload_refs, id: candidate_fragment_ids).pluck(
:fragment,
:metadata,
)
if reranker.reranker_configured?
guidance = fragments.map { |fragment, _metadata| fragment }
ranks =
DiscourseAi::Inference::HuggingFaceTextEmbeddings
.rerank(conversation_context.last[:content], guidance)
.to_a
.take(rag_conversation_chunks)
.map { _1[:index] }
if ranks.empty?
fragments = fragments.take(rag_conversation_chunks)
else
fragments = ranks.map { |idx| fragments[idx] }
end
end
<<~TEXT
<guidance>
The following texts will give you additional guidance for your response.
We included them because we believe they are relevant to this conversation topic.
Texts:
#{
fragments
.map do |fragment, metadata|
if metadata.present?
["# #{metadata}", fragment].join("\n")
else
fragment
end
end
.join("\n")
}
</guidance>
TEXT
end
end
end
end
end