Add an exception for GoogleGemini and GoogleVertexAI when using the OutputFixingParser

lib/langchain/output_parsers/output_fixing_parser.rb

@@ -47,12 +47,7 @@ def parse(completion)
       parser.parse(completion)
     rescue OutputParserException => e
       new_completion = llm.chat(
-        messages: [{role: "user",
-                    content: prompt.format(
-                      instructions: parser.get_format_instructions,
-                      completion: completion,
-                      error: e
-                    )}]
+        messages: chat_messages(completion, e)
       ).completion
       parser.parse(new_completion)
     end
@@ -70,6 +65,38 @@ def self.from_llm(llm:, parser:, prompt: nil)
 
     private
 
+    def chat_messages(completion, e)
+      # For Google LLMs, use the parts format
+      if llm.is_a?(Langchain::LLM::GoogleGemini) || llm.is_a?(Langchain::LLM::GoogleVertexAI)
+        return [
+          {
+            role: "user", 
+            parts: [
+              {
+                text: prompt.format(
+                  instructions: parser.get_format_instructions,
+                  completion: completion,
+                  error: e
+                )
+              }
+            ]
+          }
+        ]
+      end
+
+      # For other LLMs, use the standard content format
+      [
+        {
+          role: "user", 
+          content: prompt.format(
+            instructions: parser.get_format_instructions,
+            completion: completion,
+            error: e
+          )
+        }
+      ]
+    end
+
     private_class_method def self.naive_fix_prompt
       Langchain::Prompt.load_from_path(
         file_path: Langchain.root.join("langchain/output_parsers/prompts/naive_fix_prompt.yaml")

spec/lib/langchain/output_parsers/fix_spec.rb

@@ -186,5 +186,26 @@
       expect { parser.parse("Whoops I don't understand") }.to raise_error(Langchain::OutputParsers::OutputParserException)
       expect(parser.llm).to have_received(:chat).once
     end
+
+    context "with a Gemini based model" do
+      let(:llm_example) { Langchain::LLM::GoogleGemini.new(api_key: "123") }
+
+      it "parses when the llm is a Gemini based model" do
+        parser = described_class.new(**kwargs_example.merge(prompt: fix_prompt_template_example))
+        expect(parser.llm).to receive(:chat).with({ 
+          messages: [
+            {
+              role: "user",
+              parts: [
+                {
+                  text: match(fix_prompt_matcher_example)
+                }
+              ]
+            }
+          ]
+        }).and_return(double(completion: json_text_response))
+        expect(parser.parse("Whoops I don't understand")).to eq(json_response)
+      end
+    end
   end
 end