Do not parse SSE events if not streaming

When response of Anthropic model includes `data:` and even if
the request isn't streaming, `parse_sse_events` will still try to
parse the response. It can fail since the format of the response
is not really in the proper SSE format expected. We also don't need
to parse SSE events if we're not streaming.

To fix that, we check if call for `perform_completion_request` and
`perform_messages_request` have `stream` as `true`. If not, then
we skip calling `parse_sse_events`.

Changelog: fixed
EE: true

ee/lib/gitlab/llm/anthropic/client.rb

@@ -101,6 +101,7 @@ def messages_stream(messages:, **options)
         def perform_completion_request(prompt:, options:)
           logger.info(message: "Performing request to Anthropic", options: options)
           timeout = options.delete(:timeout) || DEFAULT_TIMEOUT
+          stream = options.fetch(:stream, false)
 
           Gitlab::HTTP.post(
             "#{url}/v1/complete",
@@ -108,8 +109,10 @@ def perform_completion_request(prompt:, options:)
             body: request_body(prompt: prompt, options: options).to_json,
             timeout: timeout,
             allow_local_requests: true,
-            stream_body: options.fetch(:stream, false)
+            stream_body: stream
           ) do |fragment|
+            next unless stream
+
             parse_sse_events(fragment).each do |parsed_event|
               yield parsed_event if block_given?
             end
@@ -119,6 +122,7 @@ def perform_completion_request(prompt:, options:)
         def perform_messages_request(messages:, options:)
           logger.info(message: "Performing request to Anthropic", options: options)
           timeout = options.delete(:timeout) || DEFAULT_TIMEOUT
+          stream = options.fetch(:stream, false)
 
           response = Gitlab::HTTP.post(
             "#{url}/v1/messages",
@@ -126,8 +130,10 @@ def perform_messages_request(messages:, options:)
             body: request_body_for_messages(messages: messages, options: options).to_json,
             timeout: timeout,
             allow_local_requests: true,
-            stream_body: options.fetch(:stream, false)
+            stream_body: stream
           ) do |fragment|
+            next unless stream
+
             parse_sse_events(fragment).each do |parsed_event|
               yield parsed_event if block_given?
             end

ee/spec/support/shared_examples/lib/gitlab/llm/anthropic/client_shared_examples.rb

@@ -33,7 +33,7 @@
 
   let(:expected_response) do
     {
-      'completion' => 'Completion Response',
+      'completion' => 'data: { response: Completion Response }',
       'stop' => nil,
       'stop_reason' => 'max_tokens',
       'truncated' => false,
@@ -108,11 +108,11 @@
           stub_const("Gitlab::Llm::Concerns::ExponentialBackoff::INITIAL_DELAY", 0.0)
         end
 
-        it_behaves_like 'tracks events for AI requests', 2, 4
+        it_behaves_like 'tracks events for AI requests', 2, 9
       end
     end
 
-    it_behaves_like 'tracks events for AI requests', 2, 4
+    it_behaves_like 'tracks events for AI requests', 2, 9
 
     it 'logs the response' do
       complete
@@ -353,7 +353,7 @@
         'content' => [
           {
             'type' => 'text',
-            'text' => 'Completion Response'
+            'text' => 'data: { response: Completion Response }'
           }
         ],
         'stop_reason' => 'end_turn',
@@ -426,11 +426,11 @@
           stub_const("Gitlab::Llm::Concerns::ExponentialBackoff::INITIAL_DELAY", 0.0)
         end
 
-        it_behaves_like 'tracks events for AI requests', 2, 4
+        it_behaves_like 'tracks events for AI requests', 2, 9
       end
     end
 
-    it_behaves_like 'tracks events for AI requests', 2, 4
+    it_behaves_like 'tracks events for AI requests', 2, 9
 
     it 'logs the response' do
       messages_complete