Improve passing max tokens to anthropic (#864)

* Clean up passing `max_tokens` to Anthropic constructor and chat method

* changelog entry

Author: andreibondarev

CHANGELOG.md

@@ -11,6 +11,7 @@
 
 ## [Unreleased]
 - [FEATURE] [https://github.com/patterns-ai-core/langchainrb/pull/858] Assistant, when using Anthropic, now also accepts image_url in the message.
+- [FEATURE] [https://github.com/patterns-ai-core/langchainrb/pull/859] Clean up passing `max_tokens` to Anthropic constructor and chat method
 
 ## [0.19.0] - 2024-10-23
 - [BREAKING] [https://github.com/patterns-ai-core/langchainrb/pull/840] Rename `chat_completion_model_name` parameter to `chat_model` in Langchain::LLM parameters.

lib/langchain/llm/anthropic.rb

@@ -15,14 +15,14 @@ class Anthropic < Base
       temperature: 0.0,
       completion_model: "claude-2.1",
       chat_model: "claude-3-5-sonnet-20240620",
-      max_tokens_to_sample: 256
+      max_tokens: 256
     }.freeze
 
     # Initialize an Anthropic LLM instance
     #
     # @param api_key [String] The API key to use
     # @param llm_options [Hash] Options to pass to the Anthropic client
-    # @param default_options [Hash] Default options to use on every call to LLM, e.g.: { temperature:, completion_model:, chat_model:, max_tokens_to_sample: }
+    # @param default_options [Hash] Default options to use on every call to LLM, e.g.: { temperature:, completion_model:, chat_model:, max_tokens: }
     # @return [Langchain::LLM::Anthropic] Langchain::LLM::Anthropic instance
     def initialize(api_key:, llm_options: {}, default_options: {})
       depends_on "anthropic"
@@ -32,7 +32,7 @@ def initialize(api_key:, llm_options: {}, default_options: {})
       chat_parameters.update(
         model: {default: @defaults[:chat_model]},
         temperature: {default: @defaults[:temperature]},
-        max_tokens: {default: @defaults[:max_tokens_to_sample]},
+        max_tokens: {default: @defaults[:max_tokens]},
         metadata: {},
         system: {}
       )
@@ -55,7 +55,7 @@ def initialize(api_key:, llm_options: {}, default_options: {})
     def complete(
       prompt:,
       model: @defaults[:completion_model],
-      max_tokens_to_sample: @defaults[:max_tokens_to_sample],
+      max_tokens: @defaults[:max_tokens],
       stop_sequences: nil,
       temperature: @defaults[:temperature],
       top_p: nil,
@@ -64,12 +64,12 @@ def complete(
       stream: nil
     )
       raise ArgumentError.new("model argument is required") if model.empty?
-      raise ArgumentError.new("max_tokens_to_sample argument is required") if max_tokens_to_sample.nil?
+      raise ArgumentError.new("max_tokens argument is required") if max_tokens.nil?
 
       parameters = {
         model: model,
         prompt: prompt,
-        max_tokens_to_sample: max_tokens_to_sample,
+        max_tokens_to_sample: max_tokens,
         temperature: temperature
       }
       parameters[:stop_sequences] = stop_sequences if stop_sequences

spec/langchain/llm/anthropic_spec.rb

@@ -5,6 +5,30 @@
 RSpec.describe Langchain::LLM::Anthropic do
   let(:subject) { described_class.new(api_key: "123") }
 
+  describe "#initialize" do
+    context "when default_options are passed" do
+      let(:default_options) { {max_tokens: 512} }
+
+      subject { described_class.new(api_key: "123", default_options: default_options) }
+
+      it "sets the defaults options" do
+        expect(subject.defaults[:max_tokens]).to eq(512)
+      end
+
+      it "get passed to consecutive chat() call" do
+        subject
+        expect(subject.client).to receive(:messages).with(parameters: hash_including(default_options)).and_return({})
+        subject.chat(messages: [{role: "user", content: "Hello json!"}])
+      end
+
+      it "can be overridden" do
+        subject
+        expect(subject.client).to receive(:messages).with(parameters: hash_including({max_tokens: 1024})).and_return({})
+        subject.chat(messages: [{role: "user", content: "Hello json!"}], max_tokens: 1024)
+      end
+    end
+  end
+
   describe "#complete" do
     let(:completion) { "How high is the sky?" }
     let(:fixture) { File.read("spec/fixtures/llm/anthropic/complete.json") }
@@ -17,7 +41,7 @@
             model: described_class::DEFAULTS[:completion_model],
             prompt: completion,
             temperature: described_class::DEFAULTS[:temperature],
-            max_tokens_to_sample: described_class::DEFAULTS[:max_tokens_to_sample]
+            max_tokens_to_sample: described_class::DEFAULTS[:max_tokens]
           })
           .and_return(response)
       end
@@ -40,7 +64,7 @@
             model: described_class::DEFAULTS[:completion_model],
             prompt: completion,
             temperature: described_class::DEFAULTS[:temperature],
-            max_tokens_to_sample: described_class::DEFAULTS[:max_tokens_to_sample]
+            max_tokens_to_sample: described_class::DEFAULTS[:max_tokens]
           })
           .and_return(JSON.parse(fixture))
       end
@@ -63,7 +87,7 @@
             model: described_class::DEFAULTS[:chat_model],
             messages: messages,
             temperature: described_class::DEFAULTS[:temperature],
-            max_tokens: described_class::DEFAULTS[:max_tokens_to_sample],
+            max_tokens: described_class::DEFAULTS[:max_tokens],
             stop_sequences: ["beep"]
           })
           .and_return(response)

spec/langchain/llm/openai_spec.rb

@@ -91,9 +91,15 @@
 
       it "get passed to consecutive chat() call" do
         subject
-        expect(subject.client).to receive(:chat).with(parameters: hash_including({response_format: {type: "json_object"}})).and_return({})
+        expect(subject.client).to receive(:chat).with(parameters: hash_including(default_options)).and_return({})
         subject.chat(messages: [{role: "user", content: "Hello json!"}])
       end
+
+      it "can be overridden" do
+        subject
+        expect(subject.client).to receive(:chat).with(parameters: hash_including({response_format: {type: "text"}})).and_return({})
+        subject.chat(messages: [{role: "user", content: "Hello json!"}], response_format: {type: "text"})
+      end
     end
   end