A platform for community discussion. Free, open, simple.
FIX: Bill decision models for a single pass instead of once per question (#45407)
Rafael dos Santos Silvaa36f7fa4252e
diff --git a/plugins/discourse-ai/lib/decisions/client.rb b/plugins/discourse-ai/lib/decisions/client.rbindex ed552594d8b..4d221a888b1 100644--- a/plugins/discourse-ai/lib/decisions/client.rb+++ b/plugins/discourse-ai/lib/decisions/client.rb@@ -3,9 +3,9 @@module DiscourseAimodule Decisionsclass Client- # Billing a provider that omits usage must not make its decisions free. System One- # engines evaluate the state once per question, adding a fixed prompt to each pass.- ESTIMATED_TOKENS_PER_QUESTION = 85+ # Billing a provider that omits usage must not make its decisions free. Engines+ # answer every question in a single pass over the state, behind a fixed prompt.+ ESTIMATED_PROMPT_TOKENS = 85ESTIMATED_TOKENS_PER_IMAGE = 1_000def initialize(decision_model, timeout: 30, max_attempts: 3)@@ -80,14 +80,13 @@ module DiscourseAidef estimated_input_tokens(state, questions, images)tokenizer = DiscourseAi::Tokenizer::OpenAiTokenizer- state_tokens =- tokenizer.size(state.is_a?(String) ? state : state.to_json) +- images.size * ESTIMATED_TOKENS_PER_IMAGE- questions.sum do |question|- criteria = [*question.choices, *question.levels].join(" ")- state_tokens + tokenizer.size("#{question.instructions} #{criteria}") +- ESTIMATED_TOKENS_PER_QUESTION- end+ question_tokens =+ questions.sum do |question|+ criteria = [*question.choices, *question.levels].join(" ")+ tokenizer.size("#{question.instructions} #{criteria}")+ end+ tokenizer.size(state.is_a?(String) ? state : state.to_json) ++ images.size * ESTIMATED_TOKENS_PER_IMAGE + question_tokens + ESTIMATED_PROMPT_TOKENSenddef validate!(state, questions, feature_name, images)diff --git a/plugins/discourse-ai/spec/lib/decisions/client_spec.rb b/plugins/discourse-ai/spec/lib/decisions/client_spec.rbindex f8d81f020af..44883dbb93f 100644--- a/plugins/discourse-ai/spec/lib/decisions/client_spec.rb+++ b/plugins/discourse-ai/spec/lib/decisions/client_spec.rb@@ -595,7 +595,7 @@ RSpec.describe DiscourseAi::Decisions::Client doexpect(allocation.reload.daily_used).to eq(296)end- it "estimates tokens per question, counting state and images in each, when the provider omits usage" do+ it "estimates tokens for a single pass over state, images and questions when the provider omits usage" doseeded_model.update!(vision_enabled: true)stub_request(:post, seeded_model.url).to_return(body: response_body.except(:usage).to_json)@@ -606,12 +606,8 @@ RSpec.describe DiscourseAi::Decisions::Client dofeature_name: "workflow",)- per_question_floor =- described_class::ESTIMATED_TOKENS_PER_IMAGE + described_class::ESTIMATED_TOKENS_PER_QUESTION- expect(allocation.reload.daily_used).to be_between(- questions.size * per_question_floor,- questions.size * (per_question_floor + 20),- )+ floor = described_class::ESTIMATED_TOKENS_PER_IMAGE + described_class::ESTIMATED_PROMPT_TOKENS+ expect(allocation.reload.daily_used).to be_between(floor, floor + 30)endit "does not deduct credits when the request fails" do