fix: GoogleGemini generation_config param (#665)

* fix generation_config param

* sync all parameters to match gemini API

* Migrate from active_support to in-house utils

---------

Co-authored-by: Andrei Bondarev <andrei@sourcelabs.io>

lib/langchain/llm/google_gemini.rb

@@ -18,7 +18,9 @@ def initialize(api_key:, default_options: {})
 
       chat_parameters.update(
         model: {default: @defaults[:chat_completion_model_name]},
-        temperature: {default: @defaults[:temperature]}
+        temperature: {default: @defaults[:temperature]},
+        generation_config: {default: nil},
+        safety_settings: {default: nil}
       )
       chat_parameters.remap(
         messages: :contents,
@@ -42,13 +44,25 @@ def chat(params = {})
       raise ArgumentError.new("messages argument is required") if Array(params[:messages]).empty?
 
       parameters = chat_parameters.to_params(params)
-      parameters[:generation_config] = {temperature: parameters.delete(:temperature)} if parameters[:temperature]
+      parameters[:generation_config] ||= {}
+      parameters[:generation_config][:temperature] ||= parameters[:temperature] if parameters[:temperature]
+      parameters.delete(:temperature)
+      parameters[:generation_config][:top_p] ||= parameters[:top_p] if parameters[:top_p]
+      parameters.delete(:top_p)
+      parameters[:generation_config][:top_k] ||= parameters[:top_k] if parameters[:top_k]
+      parameters.delete(:top_k)
+      parameters[:generation_config][:max_output_tokens] ||= parameters[:max_tokens] if parameters[:max_tokens]
+      parameters.delete(:max_tokens)
+      parameters[:generation_config][:response_mime_type] ||= parameters[:response_format] if parameters[:response_format]
+      parameters.delete(:response_format)
+      parameters[:generation_config][:stop_sequences] ||= parameters[:stop] if parameters[:stop]
+      parameters.delete(:stop)
 
       uri = URI("https://generativelanguage.googleapis.com/v1beta/models/#{parameters[:model]}:generateContent?key=#{api_key}")
 
       request = Net::HTTP::Post.new(uri)
       request.content_type = "application/json"
-      request.body = parameters.to_json
+      request.body = Langchain::Utils::HashTransformer.deep_transform_keys(parameters) { |key| Langchain::Utils::HashTransformer.camelize_lower(key.to_s).to_sym }.to_json
 
       response = Net::HTTP.start(uri.hostname, uri.port, use_ssl: uri.scheme == "https") do |http|
         http.request(request)

lib/langchain/utils/hash_transformer.rb

@@ -0,0 +1,25 @@
+module Langchain
+  module Utils
+    class HashTransformer
+      # Converts a string to camelCase
+      def self.camelize_lower(str)
+        str.split("_").inject([]) { |buffer, e| buffer.push(buffer.empty? ? e : e.capitalize) }.join
+      end
+
+      # Recursively transforms the keys of a hash to camel case
+      def self.deep_transform_keys(hash, &block)
+        case hash
+        when Hash
+          hash.each_with_object({}) do |(key, value), result|
+            new_key = block.call(key)
+            result[new_key] = deep_transform_keys(value, &block)
+          end
+        when Array
+          hash.map { |item| deep_transform_keys(item, &block) }
+        else
+          hash
+        end
+      end
+    end
+  end
+end

spec/langchain/llm/google_gemini_spec.rb

@@ -3,6 +3,22 @@
 RSpec.describe Langchain::LLM::GoogleGemini do
   let(:subject) { described_class.new(api_key: "123") }
 
+  describe "#initialize" do
+    it "initializes with default options" do
+      expect(subject.api_key).to eq("123")
+      expect(subject.defaults[:chat_completion_model_name]).to eq("gemini-1.5-pro-latest")
+      expect(subject.defaults[:embeddings_model_name]).to eq("text-embedding-004")
+      expect(subject.defaults[:temperature]).to eq(0.0)
+    end
+
+    it "merges default options with provided options" do
+      custom_options = {chat_completion_model_name: "custom-model", temperature: 2.0}
+      google_gemini_with_custom_options = described_class.new(api_key: "123", default_options: custom_options)
+      expect(google_gemini_with_custom_options.defaults[:chat_completion_model_name]).to eq("custom-model")
+      expect(google_gemini_with_custom_options.defaults[:temperature]).to eq(2.0)
+    end
+  end
+
   describe "#embed" do
     let(:embedding) { [0.013168523, -0.008711934, -0.046782676] }
     let(:raw_embedding_response) { double(body: File.read("spec/fixtures/llm/google_gemini/embed.json")) }
@@ -23,11 +39,42 @@
   describe "#chat" do
     let(:messages) { [{role: "user", parts: [{text: "How high is the sky?"}]}] }
     let(:raw_chat_completions_response) { double(body: File.read("spec/fixtures/llm/google_gemini/chat.json")) }
+    let(:params) { {messages: messages, model: "gemini-1.5-pro-latest", system: "system instruction", tool_choice: "AUTO", tools: [{name: "tool1"}], temperature: 1.1, response_format: "application/json", stop: ["A", "B"], generation_config: {temperature: 1.7, top_p: 1.3, response_schema: {"type" => "object", "description" => "sample schema"}}, safety_settings: [{category: "HARM_CATEGORY_UNSPECIFIED", threshold: "BLOCK_ONLY_HIGH"}]} }
 
     before do
       allow(Net::HTTP).to receive(:start).and_return(raw_chat_completions_response)
     end
 
+    it "raises an error if messages are not provided" do
+      expect { subject.chat({}) }.to raise_error(ArgumentError, "messages argument is required")
+    end
+
+    it "correctly processes and sends parameters" do
+      expect(Net::HTTP::Post).to receive(:new) do |uri|
+        expect(uri.to_s).to include("https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-pro-latest:generateContent?key=123")
+      end.and_call_original
+
+      allow_any_instance_of(Net::HTTP::Post).to receive(:body=) do |request, body|
+        parsed_body = JSON.parse(body)
+
+        expect(parsed_body["model"]).to eq("gemini-1.5-pro-latest")
+        expect(parsed_body["contents"]).to eq([{"parts" => [{"text" => "How high is the sky?"}], "role" => "user"}])
+        expect(parsed_body["systemInstruction"]).to eq({"parts" => [{"text" => "system instruction"}]})
+        expect(parsed_body["toolConfig"]).to eq({"functionCallingConfig" => {"mode" => "AUTO"}})
+        expect(parsed_body["tools"]).to eq({"functionDeclarations" => [{"name" => "tool1"}]})
+        expect(parsed_body["temperature"]).to eq(nil)
+        expect(parsed_body["generationConfig"]["temperature"]).to eq(1.7)
+        expect(parsed_body["topP"]).to eq(nil)
+        expect(parsed_body["generationConfig"]["topP"]).to eq(1.3)
+        expect(parsed_body["responseFormat"]).to eq(nil)
+        expect(parsed_body["generationConfig"]["responseMimeType"]).to eq("application/json")
+        expect(parsed_body["generationConfig"]["responseSchema"]).to eq({"type" => "object", "description" => "sample schema"})
+        expect(parsed_body["safetySettings"]).to eq([{"category" => "HARM_CATEGORY_UNSPECIFIED", "threshold" => "BLOCK_ONLY_HIGH"}])
+      end
+
+      subject.chat(params)
+    end
+
     it "returns valid llm response object" do
       response = subject.chat(messages: messages)
 

spec/langchain/utils/hash_transformer_spec.rb

@@ -0,0 +1,67 @@
+RSpec.describe Langchain::Utils::HashTransformer do
+  describe ".camelize_lower" do
+    it "converts snake_case to camelCase" do
+      expect(described_class.camelize_lower("example_key")).to eq("exampleKey")
+      expect(described_class.camelize_lower("nested_key_example")).to eq("nestedKeyExample")
+    end
+
+    it "handles strings without underscores" do
+      expect(described_class.camelize_lower("example")).to eq("example")
+    end
+
+    it "handles empty strings" do
+      expect(described_class.camelize_lower("")).to eq("")
+    end
+  end
+
+  describe ".deep_transform_keys" do
+    it "transforms keys of a simple hash" do
+      hash = {example_key: "value", another_key: "another_value"}
+      result = described_class.deep_transform_keys(hash) { |key| described_class.camelize_lower(key.to_s).to_sym }
+
+      expect(result).to eq({exampleKey: "value", anotherKey: "another_value"})
+    end
+
+    it "transforms keys of a nested hash" do
+      hash = {example_key: {nested_key: "value"}}
+      result = described_class.deep_transform_keys(hash) { |key| described_class.camelize_lower(key.to_s).to_sym }
+
+      expect(result).to eq({exampleKey: {nestedKey: "value"}})
+    end
+
+    it "transforms keys of an array of hashes" do
+      hash = {array_key: [{nested_key: "value"}, {another_key: "another_value"}]}
+      result = described_class.deep_transform_keys(hash) { |key| described_class.camelize_lower(key.to_s).to_sym }
+
+      expect(result).to eq({arrayKey: [{nestedKey: "value"}, {anotherKey: "another_value"}]})
+    end
+
+    it "handles arrays of non-hash elements" do
+      hash = {array_key: ["string", 123, :symbol]}
+      result = described_class.deep_transform_keys(hash) { |key| described_class.camelize_lower(key.to_s).to_sym }
+
+      expect(result).to eq({arrayKey: ["string", 123, :symbol]})
+    end
+
+    it "handles non-hash, non-array values" do
+      hash = {simple_key: "value"}
+      result = described_class.deep_transform_keys(hash) { |key| described_class.camelize_lower(key.to_s).to_sym }
+
+      expect(result).to eq({simpleKey: "value"})
+    end
+
+    it "handles empty hashes" do
+      hash = {}
+      result = described_class.deep_transform_keys(hash) { |key| described_class.camelize_lower(key.to_s).to_sym }
+
+      expect(result).to eq({})
+    end
+
+    it "handles deeply nested structures" do
+      hash = {level_one: {level_two: {level_three: {nested_key: "value"}}}}
+      result = described_class.deep_transform_keys(hash) { |key| described_class.camelize_lower(key.to_s).to_sym }
+
+      expect(result).to eq({levelOne: {levelTwo: {levelThree: {nestedKey: "value"}}}})
+    end
+  end
+end