omniai-google 3.18.0 → 3.18.2

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 10f903f10526a5e5b609434b87ab652142e7298e5fa881fbd9bba342baeae696
-  data.tar.gz: 0b9a117af8acd962fc427a88472f7d5f749e9cfa4848371e723eee191b034ded
+  metadata.gz: c07c0c6fcf73437ef30346dd7123a6f78855631ce160d590dc0331b48a24097a
+  data.tar.gz: 60c2e3f6f9130116a8a34d65935588fffb04e8922c7c3df85dc01ce61734c955
 SHA512:
-  metadata.gz: b5ba5463d9357d4cf12910d26beb99f4300b123c2efa44b9d0b52e1410acfc698cbf8e601e6a2f730664e946b9178522d736c72cabc661524c2adc01b494f4fc
-  data.tar.gz: 99d73d1e2a2ba26424e3cb0ad302fea0ad51e944864ac8d3e14d8770220618386f20c59b1b1ec62a7d50bae3ad15343b1140e2b80b14faebeeb235dbc3460aa1
+  metadata.gz: 180ec343c726b29490fd2a0720b1ce2b1d321bf263badfc113908d7b7d5f1aa3857f395fdc9c415fc59cd4ce925fc5583ef2e0842388316da47137f19a1790b5
+  data.tar.gz: b3a3d233bbcbe4a753bca419edc63482e2aba1143d3323e8be7dffdd6a23678e423e7db0e4f69af8a8b585cd9683537425424065c995fd088a93efb8feab895e

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/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!
 
@@ -80,14 +77,7 @@ module OmniAI
         operation_data = response.parse
         operation_name = operation_data["name"]
 
-        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
+        raise OmniAI::Error, "No operation name returned from batch recognition request" unless operation_name
 
         # Poll for completion
         result = poll_operation!(operation_name)
@@ -95,10 +85,11 @@ module OmniAI
         # 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
 
         {
@@ -237,7 +237,7 @@ module OmniAI
           # Check for errors
           if operation_data["error"]
             error_message = operation_data.dig("error", "message") || "Unknown error"
-            raise HTTPError, "Operation failed: #{error_message}"
+            raise OmniAI::Error, "Operation failed: #{error_message}"
           end
 
           # Check if done
@@ -288,6 +288,11 @@ module OmniAI
         file_result = batch_results.values.first
         return empty_transcript_data unless file_result
 
+        # A done operation can still carry a per-file failure (e.g. unsupported encoding)
+        if (error = file_result["error"])
+          raise OmniAI::Error, "Transcription failed (code #{error['code']}): #{error['message']}"
+        end
+
         transcript_segments = file_result.dig("transcript", "results")
         return empty_transcript_data unless transcript_segments&.any?
 
@@ -475,7 +480,7 @@ 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 +497,8 @@ 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 = "3.18.0"
+    VERSION = "3.18.2"
   end
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: omniai-google
 version: !ruby/object:Gem::Version
-  version: 3.18.0
+  version: 3.18.2
 platform: ruby
 authors:
 - Kevin Sylvestre