omniai-google 3.10.0 → 3.13.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 2b8dca01238f0fabfede35ce895cb4ec7ae936d9f3296c7accd88aa0284a6bfc
-  data.tar.gz: aa5f4c994eaac71ce5a39e6c3c0e0765eee2da33aacf2753d82b994b8c19b328
+  metadata.gz: 8dab5e45bcb29a2506c505023b188ff9bfab5e000529a84bf691d5fbd576f66d
+  data.tar.gz: 2f77225d19897c818297fdd225721d5aa412b95edcf4f4445c2d309662b3360e
 SHA512:
-  metadata.gz: c5b8f8335542020564f81a03bb69236f3db4c2b0683353e84ca8866d69fc557b52ae353deebd272c480fe3a9f54b190a4c451550c8a1f93f5d4e2638731b48d6
-  data.tar.gz: f5a2f43feda13a60a4fcdf7ae9e4826ffae58f527f550d258a3ba8e520421ff1b601e2f59b6affcb6560a2e59aa58452be1ce74fcbdf48c433736d2e98fdc212
+  metadata.gz: 8c5582f5d5a94cdac496da59f9840799ba3786a26047afe269e06deabf120886e704474b976cebb9b35ee947e31031f8b00479cd8d0f2d6ff0a551878799219c
+  data.tar.gz: 6c9bd1633e7deba0331cdf3134381b25b195e869c3607139797decc04b43d57b424ffbfc83234035b955cac22a37a635f34981adab55dddfc67949b586f27cb0

data/README.md

@@ -87,7 +87,7 @@ completion.text # 'The capital of Canada is Ottawa.'
 
 #### Model
 
-`model` takes an optional string (default is `gemini-3.5-flash`):
+`model` takes an optional string (default is `gemini-3.7-flash`):
 
 ```ruby
 completion = client.chat('How fast is a cheetah?', model: OmniAI::Google::Chat::Model::GEMINI_FLASH)

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

@@ -5,23 +5,51 @@ module OmniAI
     class Chat
       # Overrides usage serialize / deserialize.
       module UsageSerializer
+        # Gemini reports thinking separately from the answer: `candidatesTokenCount` covers the answer only, while
+        # `thoughtsTokenCount` covers internal reasoning. Google bills both as output and `totalTokenCount` includes
+        # both, so `output_tokens` is the sum — matching Anthropic and OpenAI, where reasoning is already folded into
+        # the output count and reported back only as a breakdown.
+        #
         # @param usage [OmniAI::Chat::Usage]
         # @return [Hash]
         def self.serialize(usage, *)
-          {
+          thinking_tokens = usage.thinking_tokens
+          candidates_tokens = usage.output_tokens
+          # `thinking_tokens` is a subset of `output_tokens`, so this cannot go negative for any Usage this gem
+          # builds. Clamp anyway: a hand-constructed Usage that violates the subset invariant should not produce a
+          # negative token count on the wire.
+          candidates_tokens = [candidates_tokens - thinking_tokens, 0].max if candidates_tokens && thinking_tokens
+
+          data = {
             promptTokenCount: usage.input_tokens,
-            candidatesTokenCount: usage.output_tokens,
+            candidatesTokenCount: candidates_tokens,
             totalTokenCount: usage.total_tokens,
           }
+          # Gemini omits the key entirely when nothing was thought; only emit it when there is a value to report.
+          data[:thoughtsTokenCount] = thinking_tokens unless thinking_tokens.nil?
+          data
         end
 
+        # Returns `nil` when the payload carries no token counts at all. A truncated stream still assembles a
+        # `usageMetadata` — Gemini sends one on every chunk, carrying only `trafficType` until the terminal chunk —
+        # so the presence of the key is not the presence of usage. Building a Usage from it would report every
+        # count as `nil`, which arithmetic downstream silently turns into zero.
+        #
+        # The test is strictly "no count is present", never "the counts are falsy": a reported `0` is a count.
+        #
         # @param data [Hash]
-        # @return [OmniAI::Chat::Usage]
+        # @return [OmniAI::Chat::Usage, nil]
         def self.deserialize(data, *)
           input_tokens = data["promptTokenCount"]
-          output_tokens = data["candidatesTokenCount"]
+          candidates_tokens = data["candidatesTokenCount"]
+          thinking_tokens = data["thoughtsTokenCount"]
           total_tokens = data["totalTokenCount"]
-          OmniAI::Chat::Usage.new(input_tokens:, output_tokens:, total_tokens:)
+
+          return if [input_tokens, candidates_tokens, thinking_tokens, total_tokens].all?(&:nil?)
+
+          output_tokens = (candidates_tokens || 0) + (thinking_tokens || 0) if candidates_tokens || thinking_tokens
+
+          OmniAI::Chat::Usage.new(input_tokens:, output_tokens:, total_tokens:, thinking_tokens:)
         end
       end
     end

data/lib/omniai/google/chat.rb

@@ -23,11 +23,13 @@ module OmniAI
         GEMINI_2_5_FLASH = "gemini-2.5-flash"
         GEMINI_3_FLASH = "gemini-3-flash-preview"
         GEMINI_3_5_FLASH = "gemini-3.5-flash"
+        GEMINI_3_6_FLASH = "gemini-3.6-flash"
+        GEMINI_3_7_FLASH = "gemini-3.7-flash"
         GEMINI_PRO = GEMINI_3_1_PRO
-        GEMINI_FLASH = GEMINI_3_5_FLASH
+        GEMINI_FLASH = GEMINI_3_7_FLASH
       end
 
-      DEFAULT_MODEL = Model::GEMINI_3_5_FLASH
+      DEFAULT_MODEL = Model::GEMINI_FLASH
 
       module ResponseMimeType
         JSON = "application/json"

data/lib/omniai/google/client.rb

@@ -39,7 +39,7 @@ module OmniAI
         credentials: OmniAI::Google.config.credentials,
         logger: OmniAI::Google.config.logger,
         host: OmniAI::Google.config.host,
-        version: OmniAI::Google.config.version,
+        version: nil,
         timeout: OmniAI::Google.config.timeout
       )
         if api_key.nil? && credentials.nil?
@@ -51,7 +51,7 @@ module OmniAI
         @project_id = project_id
         @location_id = location_id
         @credentials = Credentials.parse(credentials)
-        @version = version
+        @version = version || default_version
       end
 
       # @raise [OmniAI::Error]
@@ -128,6 +128,19 @@ module OmniAI
 
     private
 
+      # Vertex AI serves `v1`; the Gemini API serves `v1beta`. `OmniAI::Google.config.version` derives from the
+      # *config's* host, not this client's, so a client constructed with a Vertex host but an otherwise default
+      # config inherited `v1beta` and produced `/v1beta/projects/.../locations/...`, which 404s against every
+      # regional endpoint.
+      #
+      # Only the Vertex case is derived from this client's host. Any other custom host — a proxy or gateway in
+      # front of the Gemini API — keeps deferring to the config, so those callers are unaffected.
+      #
+      # @return [String]
+      def default_version
+        vertex? ? Config::Version::STABLE : OmniAI::Google.config.version
+      end
+
       # @return [String] e.g. "Bearer ..."
       def auth
         @credentials.fetch_access_token!

data/lib/omniai/google/embed.rb

@@ -40,6 +40,13 @@ module OmniAI
         [data["embedding"]["values"]]
       end
 
+      # Always returns a Usage, never nil — callers rely on `response.usage` being present here without a nil check.
+      #
+      # This deliberately differs from the chat-side `UsageSerializer`, which returns nil when the payload carries
+      # no token counts at all (a truncated Gemini stream still sends a `usageMetadata` containing only
+      # `trafficType`). Embeddings have no streaming path and so no equivalent partial payload, so there is nothing
+      # to distinguish. Do not make this nil-safe "for consistency" with chat: it would turn an unguarded
+      # `response.usage.total_tokens` in a consumer into a NoMethodError.
       USAGE_METADATA_DESERIALIZER = proc do |data, *|
         prompt_tokens = data.dig("usageMetadata", "promptTokenCount")
         total_tokens = data.dig("usageMetadata", "totalTokenCount")

data/lib/omniai/google/version.rb

@@ -2,6 +2,6 @@
 
 module OmniAI
   module Google
-    VERSION = "3.10.0"
+    VERSION = "3.13.0"
   end
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: omniai-google
 version: !ruby/object:Gem::Version
-  version: 3.10.0
+  version: 3.13.0
 platform: ruby
 authors:
 - Kevin Sylvestre
@@ -57,14 +57,14 @@ dependencies:
     requirements:
     - - "~>"
       - !ruby/object:Gem::Version
-        version: '3.7'
+        version: '3.8'
   type: :runtime
   prerelease: false
   version_requirements: !ruby/object:Gem::Requirement
     requirements:
     - - "~>"
       - !ruby/object:Gem::Version
-        version: '3.7'
+        version: '3.8'
 - !ruby/object:Gem::Dependency
   name: openssl
   requirement: !ruby/object:Gem::Requirement