FIX: Claude's max_tookens_to_sample is a required field (#97)

lib/shared/inference/anthropic_completions.rb

@@ -22,9 +22,9 @@ module ::DiscourseAi
 
         payload = { model: model, prompt: prompt }
 
-        payload[:temperature] = temperature if temperature
         payload[:top_p] = top_p if top_p
-        payload[:max_tokens_to_sample] = max_tokens if max_tokens
+        payload[:max_tokens_to_sample] = max_tokens || 2000
+        payload[:temperature] = temperature if temperature
         payload[:stream] = true if block_given?
 
         Net::HTTP.start(

spec/lib/modules/ai_bot/jobs/regular/create_ai_reply_spec.rb

@@ -85,8 +85,8 @@ RSpec.describe Jobs::CreateAiReply do
           deltas,
           model: "claude-v1.3",
           req_opts: {
-            temperature: 0.4,
             max_tokens_to_sample: 3000,
+            temperature: 0.4,
             stream: true,
           },
         )

spec/lib/modules/summarization/models/anthropic_spec.rb

@@ -3,10 +3,10 @@
 require_relative "../../../../support/anthropic_completion_stubs"
 
 RSpec.describe DiscourseAi::Summarization::Models::Anthropic do
-  let(:model) { "claude-v1" }
-  let(:max_tokens) { 720 }
+  subject(:model) { described_class.new(model_name, max_tokens: max_tokens) }
 
-  subject { described_class.new(model, max_tokens: max_tokens) }
+  let(:model_name) { "claude-v1" }
+  let(:max_tokens) { 720 }
 
   let(:content) do
     {
@@ -44,7 +44,7 @@ RSpec.describe DiscourseAi::Summarization::Models::Anthropic do
         )
 
         summarized_chunks =
-          subject.summarize_in_chunks(content[:contents], opts).map { |c| c[:summary] }
+          model.summarize_in_chunks(content[:contents], opts).map { |c| c[:summary] }
 
         expect(summarized_chunks).to contain_exactly("This is summary 1")
       end
@@ -67,7 +67,7 @@ RSpec.describe DiscourseAi::Summarization::Models::Anthropic do
         end
 
         summarized_chunks =
-          subject.summarize_in_chunks(content[:contents], opts).map { |c| c[:summary] }
+          model.summarize_in_chunks(content[:contents], opts).map { |c| c[:summary] }
 
         expect(summarized_chunks).to contain_exactly("This is summary 1", "This is summary 2")
       end
@@ -86,9 +86,7 @@ RSpec.describe DiscourseAi::Summarization::Models::Anthropic do
 
       AnthropicCompletionStubs.stub_response(messages, "<ai>concatenated summary</ai>")
 
-      expect(subject.concatenate_summaries(["summary 1", "summary 2"])).to eq(
-        "concatenated summary",
-      )
+      expect(model.concatenate_summaries(["summary 1", "summary 2"])).to eq("concatenated summary")
     end
   end
 
@@ -110,7 +108,7 @@ RSpec.describe DiscourseAi::Summarization::Models::Anthropic do
 
       AnthropicCompletionStubs.stub_response(instructions, "<ai>truncated summary</ai>")
 
-      expect(subject.summarize_with_truncation(content[:contents], opts)).to eq("truncated summary")
+      expect(model.summarize_with_truncation(content[:contents], opts)).to eq("truncated summary")
     end
   end
 end

spec/lib/modules/summarization/models/discourse_spec.rb

@@ -1,10 +1,10 @@
 # frozen_string_literal: true
 
 RSpec.describe DiscourseAi::Summarization::Models::Discourse do
-  let(:model) { "bart-large-cnn-samsum" }
-  let(:max_tokens) { 20 }
+  subject(:model) { described_class.new(model_name, max_tokens: max_tokens) }
 
-  subject { described_class.new(model, max_tokens: max_tokens) }
+  let(:model_name) { "bart-large-cnn-samsum" }
+  let(:max_tokens) { 20 }
 
   let(:content) do
     {
@@ -22,7 +22,7 @@ RSpec.describe DiscourseAi::Summarization::Models::Discourse do
         :post,
         "#{SiteSetting.ai_summarization_discourse_service_api_endpoint}/api/v1/classify",
       )
-      .with(body: JSON.dump(model: model, content: prompt))
+      .with(body: JSON.dump(model: model_name, content: prompt))
       .to_return(status: 200, body: JSON.dump(summary_text: response))
   end
 
@@ -40,7 +40,7 @@ RSpec.describe DiscourseAi::Summarization::Models::Discourse do
         stub_request(expected_messages(content[:contents], opts), "This is summary 1")
 
         summarized_chunks =
-          subject.summarize_in_chunks(content[:contents], opts).map { |c| c[:summary] }
+          model.summarize_in_chunks(content[:contents], opts).map { |c| c[:summary] }
 
         expect(summarized_chunks).to contain_exactly("This is summary 1")
       end
@@ -60,7 +60,7 @@ RSpec.describe DiscourseAi::Summarization::Models::Discourse do
         end
 
         summarized_chunks =
-          subject.summarize_in_chunks(content[:contents], opts).map { |c| c[:summary] }
+          model.summarize_in_chunks(content[:contents], opts).map { |c| c[:summary] }
 
         expect(summarized_chunks).to contain_exactly("This is summary 1", "This is summary 2")
       end
@@ -73,9 +73,7 @@ RSpec.describe DiscourseAi::Summarization::Models::Discourse do
 
       stub_request(messages, "concatenated summary")
 
-      expect(subject.concatenate_summaries(["summary 1", "summary 2"])).to eq(
-        "concatenated summary",
-      )
+      expect(model.concatenate_summaries(["summary 1", "summary 2"])).to eq("concatenated summary")
     end
   end
 
@@ -87,7 +85,7 @@ RSpec.describe DiscourseAi::Summarization::Models::Discourse do
 
       stub_request("( 1 asd said : this is", "truncated summary")
 
-      expect(subject.summarize_with_truncation(content[:contents], opts)).to eq("truncated summary")
+      expect(model.summarize_with_truncation(content[:contents], opts)).to eq("truncated summary")
     end
   end
 end

spec/lib/modules/summarization/models/open_ai_spec.rb

@@ -3,10 +3,10 @@
 require_relative "../../../../support/openai_completions_inference_stubs"
 
 RSpec.describe DiscourseAi::Summarization::Models::OpenAi do
-  let(:model) { "gpt-3.5-turbo" }
-  let(:max_tokens) { 720 }
+  subject(:model) { described_class.new(model_name, max_tokens: max_tokens) }
 
-  subject { described_class.new(model, max_tokens: max_tokens) }
+  let(:model_name) { "gpt-3.5-turbo" }
+  let(:max_tokens) { 720 }
 
   let(:content) do
     {
@@ -47,7 +47,7 @@ RSpec.describe DiscourseAi::Summarization::Models::OpenAi do
         )
 
         summarized_chunks =
-          subject.summarize_in_chunks(content[:contents], opts).map { |c| c[:summary] }
+          model.summarize_in_chunks(content[:contents], opts).map { |c| c[:summary] }
 
         expect(summarized_chunks).to contain_exactly("This is summary 1")
       end
@@ -70,7 +70,7 @@ RSpec.describe DiscourseAi::Summarization::Models::OpenAi do
         end
 
         summarized_chunks =
-          subject.summarize_in_chunks(content[:contents], opts).map { |c| c[:summary] }
+          model.summarize_in_chunks(content[:contents], opts).map { |c| c[:summary] }
 
         expect(summarized_chunks).to contain_exactly("This is summary 1", "This is summary 2")
       end
@@ -90,9 +90,7 @@ RSpec.describe DiscourseAi::Summarization::Models::OpenAi do
 
       OpenAiCompletionsInferenceStubs.stub_response(messages, "concatenated summary")
 
-      expect(subject.concatenate_summaries(["summary 1", "summary 2"])).to eq(
-        "concatenated summary",
-      )
+      expect(model.concatenate_summaries(["summary 1", "summary 2"])).to eq("concatenated summary")
     end
   end
 
@@ -110,7 +108,7 @@ RSpec.describe DiscourseAi::Summarization::Models::OpenAi do
 
       OpenAiCompletionsInferenceStubs.stub_response(truncated_version, "truncated summary")
 
-      expect(subject.summarize_with_truncation(content[:contents], opts)).to eq("truncated summary")
+      expect(model.summarize_with_truncation(content[:contents], opts)).to eq("truncated summary")
     end
   end
 end

spec/lib/modules/summarization/strategies/fold_content_spec.rb

@@ -4,6 +4,8 @@ require_relative "../../../../support/summarization/dummy_completion_model"
 
 RSpec.describe DiscourseAi::Summarization::Strategies::FoldContent do
   describe "#summarize" do
+    subject(:strategy) { described_class.new(model) }
+
     let(:summarize_text) { "This is a text" }
     let(:model) { DummyCompletionModel.new(model_tokens) }
     let(:model_tokens) do
@@ -11,13 +13,11 @@ RSpec.describe DiscourseAi::Summarization::Strategies::FoldContent do
       DiscourseAi::Tokenizer::BertTokenizer.size("(1 asd said: This is a text ") + 3
     end
 
-    subject { described_class.new(model) }
-
     let(:content) { { contents: [{ poster: "asd", id: 1, text: summarize_text }] } }
 
     context "when the content to summarize fits in a single call" do
       it "does one call to summarize content" do
-        result = subject.summarize(content)
+        result = strategy.summarize(content)
 
         expect(model.summarization_calls).to eq(1)
         expect(result[:summary]).to eq(DummyCompletionModel::SINGLE_SUMMARY)
@@ -28,7 +28,7 @@ RSpec.describe DiscourseAi::Summarization::Strategies::FoldContent do
       it "summarizes each chunk and then concatenates them" do
         content[:contents] << { poster: "asd2", id: 2, text: summarize_text }
 
-        result = subject.summarize(content)
+        result = strategy.summarize(content)
 
         expect(model.summarization_calls).to eq(3)
         expect(result[:summary]).to eq(DummyCompletionModel::CONCATENATED_SUMMARIES)

spec/lib/modules/summarization/strategies/truncate_content_spec.rb

@@ -4,19 +4,19 @@ require_relative "../../../../support/summarization/dummy_completion_model"
 
 RSpec.describe DiscourseAi::Summarization::Strategies::TruncateContent do
   describe "#summarize" do
+    subject(:strategy) { described_class.new(model) }
+
     let(:summarize_text) { "This is a text" }
     let(:model_tokens) { summarize_text.length }
     let(:model) { DummyCompletionModel.new(model_tokens) }
 
-    subject { described_class.new(model) }
-
     let(:content) { { contents: [{ poster: "asd", id: 1, text: summarize_text }] } }
 
     context "when the content to summarize doesn't fit in a single call" do
       it "summarizes a truncated version" do
         content[:contents] << { poster: "asd2", id: 2, text: summarize_text }
 
-        result = subject.summarize(content)
+        result = strategy.summarize(content)
 
         expect(model.summarization_calls).to eq(1)
         expect(result[:summary]).to eq(DummyCompletionModel::SINGLE_SUMMARY)

spec/shared/inference/anthropic_completions_spec.rb

@@ -9,7 +9,7 @@ RSpec.describe DiscourseAi::Inference::AnthropicCompletions do
     response_text = "1. Serenity\\n2. Laughter\\n3. Adventure"
     prompt = "Human: write 3 words\n\n"
     user_id = 183
-    req_opts = { temperature: 0.5, max_tokens_to_sample: 700 }
+    req_opts = { max_tokens_to_sample: 700, temperature: 0.5 }
 
     AnthropicCompletionStubs.stub_response(prompt, response_text, req_opts: req_opts)
 

spec/support/anthropic_completion_stubs.rb

@@ -17,7 +17,11 @@ class AnthropicCompletionStubs
     def stub_response(prompt, response_text, req_opts: {})
       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: "claude-v1", prompt: prompt, max_tokens_to_sample: 2000 }.merge(
+            req_opts,
+          ).to_json,
+        )
         .to_return(status: 200, body: JSON.dump(response(response_text)))
     end