mirror of
https://github.com/discourse/discourse.git
synced 2026-08-07 13:19:19 +08:00
0b73d7ac1aBackfill jobs were fetching the same LlmModel on every loop iteration, causing N+1 queries. For a batch of 100 items, this meant 100+ identical SQL queries to populate the same LlmModel object. ## Changes - Cache LlmModel once per job execution in summaries and translation backfills - Pass cached model through to summarization and translation methods via optional parameters - Add tests verifying models are queried once per batch instead of once per item ## Impact **Before:** For 100 items, 100+ `LlmModel.find_by()` queries **After:** For 100 items, 1-2 queries total All changes are backward compatible - new parameters are optional with nil defaults, so existing callers continue to work unchanged.
84 lines
2.7 KiB
Ruby
Vendored
84 lines
2.7 KiB
Ruby
Vendored
# frozen_string_literal: true
|
|
|
|
module DiscourseAi
|
|
module Summarization
|
|
class << self
|
|
def topic_summary(topic, llm_model: nil)
|
|
return nil if !SiteSetting.ai_summarization_enabled
|
|
if (
|
|
ai_persona = AiPersona.find_by_id_from_cache(SiteSetting.ai_summarization_persona)
|
|
).blank?
|
|
return nil
|
|
end
|
|
|
|
persona_klass = ai_persona.class_instance
|
|
llm_model ||= find_summarization_model(persona_klass)
|
|
return nil if llm_model.blank?
|
|
|
|
DiscourseAi::Summarization::FoldContent.new(
|
|
build_bot(persona_klass, llm_model),
|
|
DiscourseAi::Summarization::Strategies::TopicSummary.new(topic),
|
|
)
|
|
end
|
|
|
|
def topic_gist(topic, llm_model: nil)
|
|
return nil if !SiteSetting.ai_summarization_enabled
|
|
if (
|
|
ai_persona = AiPersona.find_by_id_from_cache(SiteSetting.ai_summary_gists_persona)
|
|
).blank?
|
|
return nil
|
|
end
|
|
|
|
persona_klass = ai_persona.class_instance
|
|
llm_model ||= find_summarization_model(persona_klass)
|
|
return nil if llm_model.blank?
|
|
|
|
DiscourseAi::Summarization::FoldContent.new(
|
|
build_bot(persona_klass, llm_model),
|
|
DiscourseAi::Summarization::Strategies::HotTopicGists.new(topic),
|
|
)
|
|
end
|
|
|
|
def chat_channel_summary(channel, time_window_in_hours, llm_model: nil)
|
|
return nil if !SiteSetting.ai_summarization_enabled
|
|
if (
|
|
ai_persona = AiPersona.find_by_id_from_cache(SiteSetting.ai_summarization_persona)
|
|
).blank?
|
|
return nil
|
|
end
|
|
|
|
persona_klass = ai_persona.class_instance
|
|
llm_model ||= find_summarization_model(persona_klass)
|
|
return nil if llm_model.blank?
|
|
|
|
DiscourseAi::Summarization::FoldContent.new(
|
|
build_bot(persona_klass, llm_model),
|
|
DiscourseAi::Summarization::Strategies::ChatMessages.new(channel, time_window_in_hours),
|
|
persist_summaries: false,
|
|
)
|
|
end
|
|
|
|
# Priorities are:
|
|
# 1. Persona's default LLM
|
|
# 2. SiteSetting.ai_default_llm_model (or newest LLM if not set)
|
|
def find_summarization_model(persona_klass)
|
|
model_id = persona_klass.default_llm_id || SiteSetting.ai_default_llm_model
|
|
|
|
if model_id.present?
|
|
LlmModel.find_by(id: model_id)
|
|
else
|
|
LlmModel.last
|
|
end
|
|
end
|
|
|
|
### Private
|
|
|
|
def build_bot(persona_klass, llm_model)
|
|
persona = persona_klass.new
|
|
user = User.find_by(id: persona_klass.user_id) || Discourse.system_user
|
|
|
|
DiscourseAi::Personas::Bot.as(user, persona: persona, model: llm_model)
|
|
end
|
|
end
|
|
end
|
|
end
|
