omniai-google 3.17.0 → 3.18.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 9fd43c08069868c8d20d91ac6e1028baf97643eda9018a0db146c73ce0828c15
-  data.tar.gz: 04ca6d4c4789b91c194abb44f8d2d178f1e65605644fe3c72a72824e8707a9e2
+  metadata.gz: cc3b952ff375db770509f12502a396e2a7cbdd9675467f76f3e83c6eb8e0ce23
+  data.tar.gz: 9f7c3394ad70dd14035ed1fe3dd60b9b2097d4708c8f744601e66bfee5fc09d9
 SHA512:
-  metadata.gz: 3b2076e4c7f504e899015fbd0c8fb4f3290d347d2191051cdad1a6e079f45b6086bf8b6725693d073c45e02adf85bc43be7aace65a37fce282a163ed338596c6
-  data.tar.gz: ebf586d81abd71f7959f12f85159a27b3a5a53cd4966adcc2e05bec009341378fc306140ca0f94e129a3a4df1875ba2b5433e1b2b738e2c78c6a2eaf24ab0e87
+  metadata.gz: daf96548f0e096cd0fa5c7b2cfb2c1dda1e70559dab2f98ba75a69dd00ff00e443969529c634f317b8df84d089b25106640d6478bfd54473034fd082f23a0a25
+  data.tar.gz: 56aa7271fc97ac4002c739eedbddb8834bf440f2d4868732b31f68a250425d2a6f1f6b9fe22a1496f08deb115fd66312813c08fa5396042d9b76ea9a360fafe6

data/README.md

@@ -134,6 +134,8 @@ OmniAI::Google.configure do |config|
 end
 ```
 
+The key must be a **symbol**. `chat_options["max_tokens"]` is neither read nor withheld from the payload, so it reaches Gemini as an unknown top-level field and the request fails with a 400.
+
 The value is passed through unchanged — no floor is imposed, so the number you ask for is the number that reaches the wire. When the cap is hit, `response.finish_reason.reason` is `:length`.
 
 **Size it as thinking headroom plus expected answer.** Unlike Anthropic's answer-only ceiling, Gemini spends this budget on thinking *before* emitting an answer. A cap sized to the expected answer alone gets consumed by thinking on any request the model reasons about:

data/lib/omniai/google/chat/tool_call_result_serializer.rb

@@ -5,17 +5,21 @@ module OmniAI
     class Chat
       # Overrides tool-call result serialize / deserialize.
       module ToolCallResultSerializer
+        # A result whose content is media (e.g. an image) is sent as a multimodal function response part.
+        #
         # @param tool_call_response [OmniAI::Chat::ToolCallResult]
         # @return [Hash]
         def self.serialize(tool_call_response, *)
+          name = tool_call_response.tool_call_id
+          content = tool_call_response.content
+
           result = {
-            functionResponse: {
-              name: tool_call_response.tool_call_id,
-              response: {
-                name: tool_call_response.tool_call_id,
-                content: tool_call_response.content,
-              },
-            },
+            functionResponse:
+              if content.is_a?(OmniAI::Chat::Media)
+                { name:, response: { name: }, parts: [MediaSerializer.serialize(content)] }
+              else
+                { name:, response: { name:, content: } }
+              end,
           }
 
           thought_signature = tool_call_response.options[:thought_signature]

data/lib/omniai/google/transcribe.rb

@@ -69,9 +69,6 @@ module OmniAI
       #
       # @return [OmniAI::Transcribe::Transcription]
       def process_async!
-        # Track if we uploaded the file for cleanup
-        uploaded_gcs_uri = nil
-
         # Start the batch recognition job
         response = request_batch!
 
@@ -82,23 +79,17 @@ module OmniAI
 
         raise HTTPError, "No operation name returned from batch recognition request" unless operation_name
 
-        # Extract GCS URI for cleanup if we uploaded it
-        if operation_data.dig("metadata", "batchRecognizeRequest", "files")
-          file_uri = operation_data.dig("metadata", "batchRecognizeRequest", "files", 0, "uri")
-          # Only mark for cleanup if it's not a user-provided GCS URI
-          uploaded_gcs_uri = file_uri unless @io.is_a?(String) && @io.start_with?("gs://")
-        end
-
         # Poll for completion
         result = poll_operation!(operation_name)
 
         # Extract transcript from completed operation
         transcript_data = extract_batch_transcript(result)
 
-        # Clean up uploaded file if we created it
-        cleanup_gcs_file(uploaded_gcs_uri) if uploaded_gcs_uri
-
         Transcription.parse(model: @model, format: @format, data: transcript_data)
+      ensure
+        # Delete our upload even when the request fails or polling times out
+        cleanup_gcs_file(@uploaded_gcs_uri) if @uploaded_gcs_uri
+        @uploaded_gcs_uri = nil
       end
 
     protected

data/lib/omniai/google/transcribe_helpers.rb

@@ -194,8 +194,8 @@ module OmniAI
         audio_uri = if @io.is_a?(String) && @io.start_with?("gs://")
                       @io
                     else
-                      # Force GCS upload for batch recognition
-                      Bucket.process!(client: @client, io: @io)
+                      # Force GCS upload for batch recognition; process_async! deletes it
+                      @uploaded_gcs_uri ||= Bucket.process!(client: @client, io: @io)
                     end
 
         {
@@ -475,7 +475,7 @@ module OmniAI
 
         file = bucket.file(object_name)
         file&.delete
-      rescue ::Google::Cloud::Error => e
+      rescue StandardError => e # never mask the transcript or the original error
         @client.logger&.warn("Failed to cleanup GCS file #{gcs_uri}: #{e.message}")
       end
 
@@ -492,7 +492,8 @@ module OmniAI
         estimated_duration_seconds = (file_size_bytes * 8.0) / assumed_bitrate_bps
         estimated_processing_seconds = estimated_duration_seconds * 0.3
         total_wait_seconds = estimated_processing_seconds + 90
-        final_wait_seconds = total_wait_seconds.clamp(180, 10_800)
+        # Batch jobs can queue for minutes regardless of audio length
+        final_wait_seconds = total_wait_seconds.clamp(600, 10_800)
         (final_wait_seconds / 15).ceil
       end
     end

data/lib/omniai/google/version.rb

@@ -2,6 +2,6 @@
 
 module OmniAI
   module Google
-    VERSION = "3.17.0"
+    VERSION = "3.18.1"
   end
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: omniai-google
 version: !ruby/object:Gem::Version
-  version: 3.17.0
+  version: 3.18.1
 platform: ruby
 authors:
 - Kevin Sylvestre