support disabling native tools for open ai

Author: SamSaffron

app/models/llm_model.rb

@@ -26,10 +26,14 @@ def self.provider_params
       },
       open_ai: {
         organization: :text,
+        disable_native_tools: :checkbox,
       },
       google: {
         disable_native_tools: :checkbox,
       },
+      azure: {
+        disable_native_tools: :checkbox,
+      },
       hugging_face: {
         disable_system_prompt: :checkbox,
       },

lib/completions/dialects/chat_gpt.rb

@@ -44,17 +44,31 @@ def is_gpt_o?
           llm_model.provider == "open_ai" && llm_model.name.include?("o1-")
         end
 
+        def disable_native_tools?
+          return @disable_native_tools if defined?(@disable_native_tools)
+          !!@disable_native_tools = llm_model.lookup_custom_param("disable_native_tools")
+        end
+
         private
 
         def tools_dialect
-          @tools_dialect ||= DiscourseAi::Completions::Dialects::OpenAiTools.new(prompt.tools)
+          if disable_native_tools?
+            super
+          else
+            @tools_dialect ||= DiscourseAi::Completions::Dialects::OpenAiTools.new(prompt.tools)
+          end
         end
 
         def system_msg(msg)
+          content = msg[:content]
+          if disable_native_tools? && tools_dialect.instructions.present?
+            content = content + "\n\n" + tools_dialect.instructions
+          end
+
           if is_gpt_o?
-            { role: "user", content: msg[:content] }
+            { role: "user", content: content }
           else
-            { role: "system", content: msg[:content] }
+            { role: "system", content: content }
           end
         end
 
@@ -63,11 +77,19 @@ def model_msg(msg)
         end
 
         def tool_call_msg(msg)
-          tools_dialect.from_raw_tool_call(msg)
+          if disable_native_tools?
+            super
+          else
+            tools_dialect.from_raw_tool_call(msg)
+          end
         end
 
         def tool_msg(msg)
-          tools_dialect.from_raw_tool(msg)
+          if disable_native_tools?
+            super
+          else
+            tools_dialect.from_raw_tool(msg)
+          end
         end
 
         def user_msg(msg)

lib/completions/endpoints/open_ai.rb

@@ -36,6 +36,7 @@ def perform_completion!(
           partial_tool_calls: false,
           &blk
         )
+          @disable_native_tools = dialect.disable_native_tools?
           if dialect.respond_to?(:is_gpt_o?) && dialect.is_gpt_o? && block_given?
             # we need to disable streaming and simulate it
             blk.call "", lambda { |*| }
@@ -69,10 +70,17 @@ def prepare_payload(prompt, model_params, dialect)
             # We'll fallback to guess this using the tokenizer.
             payload[:stream_options] = { include_usage: true } if llm_model.provider == "open_ai"
           end
-          if dialect.tools.present?
-            payload[:tools] = dialect.tools
-            if dialect.tool_choice.present?
-              payload[:tool_choice] = { type: "function", function: { name: dialect.tool_choice } }
+          if !xml_tools_enabled?
+            if dialect.tools.present?
+              payload[:tools] = dialect.tools
+              if dialect.tool_choice.present?
+                payload[:tool_choice] = {
+                  type: "function",
+                  function: {
+                    name: dialect.tool_choice,
+                  },
+                }
+              end
             end
           end
           payload
@@ -121,7 +129,7 @@ def decode_chunk_finish
         end
 
         def xml_tools_enabled?
-          false
+          !!@disable_native_tools
         end
 
         private

spec/lib/completions/endpoints/open_ai_spec.rb

@@ -53,11 +53,10 @@ def stream_line(delta, finish_reason: nil, tool_call: false)
     }.to_json
   end
 
-  def stub_raw(chunks)
-    WebMock.stub_request(:post, "https://api.openai.com/v1/chat/completions").to_return(
-      status: 200,
-      body: chunks,
-    )
+  def stub_raw(chunks, body_blk: nil)
+    stub = WebMock.stub_request(:post, "https://api.openai.com/v1/chat/completions")
+    stub.with(body: body_blk) if body_blk
+    stub.to_return(status: 200, body: chunks)
   end
 
   def stub_streamed_response(prompt, deltas, tool_call: false)
@@ -391,6 +390,59 @@ def request_body(prompt, stream: false, tool_call: false)
   end
 
   describe "#perform_completion!" do
+    context "when using XML tool calls format" do
+      let(:xml_tool_call_response) { <<~XML }
+        <function_calls>
+        <invoke>
+        <tool_name>get_weather</tool_name>
+        <parameters>
+        <location>Sydney</location>
+        <unit>c</unit>
+        </parameters>
+        </invoke>
+        </function_calls>
+      XML
+
+      it "parses XML tool calls" do
+        response = {
+          id: "chatcmpl-6sZfAb30Rnv9Q7ufzFwvQsMpjZh8S",
+          object: "chat.completion",
+          created: 1_678_464_820,
+          model: "gpt-3.5-turbo-0301",
+          usage: {
+            prompt_tokens: 8,
+            completion_tokens: 13,
+            total_tokens: 499,
+          },
+          choices: [
+            {
+              message: {
+                role: "assistant",
+                content: xml_tool_call_response,
+              },
+              finish_reason: "stop",
+              index: 0,
+            },
+          ],
+        }.to_json
+
+        endpoint.llm_model.update!(provider_params: { disable_native_tools: true })
+        body = nil
+        open_ai_mock.stub_raw(response, body_blk: proc { |inner_body| body = inner_body })
+
+        dialect = compliance.dialect(prompt: compliance.generic_prompt(tools: tools))
+        tool_call = endpoint.perform_completion!(dialect, user)
+
+        body_parsed = JSON.parse(body, symbolize_names: true)
+        expect(body_parsed[:tools]).to eq(nil)
+
+        expect(body_parsed[:messages][0][:content]).to include("<function_calls>")
+
+        expect(tool_call.name).to eq("get_weather")
+        expect(tool_call.parameters).to eq({ location: "Sydney", unit: "c" })
+      end
+    end
+
     context "when using regular mode" do
       context "with simple prompts" do
         it "completes a trivial prompt and logs the response" do