omniai-google 3.15.0 → 3.17.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: dc72ad3bc8cfe3181908f748f705708827ea8f10a3c533b9d1d4352d6f5d9df0
-  data.tar.gz: c29f890a61470dc3ddce22d5753ca4410a04ba91b93faab6381411805e41fdbd
+  metadata.gz: 9fd43c08069868c8d20d91ac6e1028baf97643eda9018a0db146c73ce0828c15
+  data.tar.gz: 04ca6d4c4789b91c194abb44f8d2d178f1e65605644fe3c72a72824e8707a9e2
 SHA512:
-  metadata.gz: bdee9a2925e8ea39873fb8e34b94378aa0549d0b1443a3c20b55bef586dda1208514e042ee63c6780ed6ec8f457b24a7af85f9def73d82c1c77e4a6ca84a6db2
-  data.tar.gz: 3402943f6994d0489fc2382731e41e116667e3e93b09908dc6549c1f890f7ba2bdb397e27af9d4999fb925a7be19489c2f1566dae6e56dcf9cdb61bf731e20c2
+  metadata.gz: 3b2076e4c7f504e899015fbd0c8fb4f3290d347d2191051cdad1a6e079f45b6086bf8b6725693d073c45e02adf85bc43be7aace65a37fce282a163ed338596c6
+  data.tar.gz: ebf586d81abd71f7959f12f85159a27b3a5a53cd4966adcc2e05bec009341378fc306140ca0f94e129a3a4df1875ba2b5433e1b2b738e2c78c6a2eaf24ab0e87

data/README.md

@@ -118,6 +118,30 @@ end
 client.chat('Be poetic.', stream:)
 ```
 
+#### Max Output Tokens
+
+`max_tokens:` caps the response, mapping to Gemini's `generationConfig.maxOutputTokens`:
+
+```ruby
+client.chat("Summarize this page.", model: "gemini-3.7-flash", max_tokens: 8_000)
+```
+
+It can also be set globally, and a per-call value wins:
+
+```ruby
+OmniAI::Google.configure do |config|
+  config.chat_options[:max_tokens] = 8_000
+end
+```
+
+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:
+
+```
+max_tokens: 200  ->  finish_reason :length, 196 output tokens, 115 of them thinking, 81 characters of answer
+```
+
 #### Extended Thinking
 
 Gemini models support extended thinking, which shows the model's reasoning process.

data/lib/omniai/google/chat.rb

@@ -93,9 +93,15 @@ module OmniAI
           }.compact, json: payload)
       end
 
+      # `chat_options` is forwarded verbatim, so a key this class builds itself must be excluded or it is sent
+      # twice — and `max_tokens` is OmniAI's normalized name, not one Gemini knows. Left in, it reaches the
+      # wire as an unknown top-level field and the request fails with
+      # `Invalid JSON payload received. Unknown name "max_tokens"`. `#generation_config` consumes it and
+      # emits `maxOutputTokens` instead.
+      #
       # @return [Hash]
       def payload
-        OmniAI::Google.config.chat_options.merge({
+        OmniAI::Google.config.chat_options.except(:max_tokens).merge({
           system_instruction: @prompt.messages.find(&:system?)&.serialize(context:),
           contents: @prompt.messages.reject(&:system?).map { |message| message.serialize(context:) },
           tools:,
@@ -138,11 +144,26 @@ module OmniAI
 
         data[:temperature] = @temperature if @temperature
         data[:thinkingConfig] = thinking_config if @options[:thinking]
+        data[:maxOutputTokens] = max_tokens if max_tokens
 
         data = data.compact
         data unless data.empty?
       end
 
+      # A per-call `max_tokens:`, falling back to `config.chat_options[:max_tokens]` so a cap can also be set
+      # globally — the same precedence omniai-anthropic applies. The value is passed through unchanged: no
+      # floor is imposed, so the number a caller asks for is the number that reaches the wire.
+      #
+      # Note that Gemini spends this budget on thinking BEFORE emitting an answer, unlike Anthropic's
+      # answer-only ceiling. A cap sized to the expected answer alone will be consumed by thinking on any
+      # request the model reasons about, returning `finishReason: MAX_TOKENS` with little or no text. Size it
+      # as thinking headroom plus expected answer.
+      #
+      # @return [Integer, nil]
+      def max_tokens
+        @options[:max_tokens] || OmniAI::Google.config.chat_options[:max_tokens]
+      end
+
       # @return [String]
       def path
         "#{@client.path}/models/#{@model}:#{operation}"

data/lib/omniai/google/client.rb

@@ -121,9 +121,26 @@ module OmniAI
         !@credentials.nil?
       end
 
+      # Vertex AI is served from three host shapes under `googleapis.com`: the global `aiplatform`, a
+      # region-prefixed `<region>-aiplatform`, and the multi-region `aiplatform.<geo>.rep`. Only the first two
+      # contain the literal "aiplatform.googleapis.com", so a substring test misses the multi-region endpoint.
+      #
+      # Deliberately wider than those three shapes: any labels are accepted between `aiplatform` and
+      # `googleapis.com`, so a future multi-region shape needs no change here. Everything it accepts is still
+      # under `googleapis.com`.
+      #
+      # Matched against the parsed hostname rather than the raw host, so a proxy whose path or query merely
+      # mentions a Vertex host is not treated as Vertex. A host given with no scheme has no hostname to parse, so
+      # the raw value is matched instead; the pattern is anchored, which keeps that safe — note that a schemeless
+      # host carrying a path therefore does not match.
+      VERTEX_HOSTNAME = /\A(?:[a-z0-9-]+-)?aiplatform(?:\.[a-z0-9-]+)*\.googleapis\.com\z/
+
       # @return [Boolean]
       def vertex?
-        @host.include?("aiplatform.googleapis.com")
+        hostname = URI.parse(@host).hostname || @host
+        VERTEX_HOSTNAME.match?(hostname.downcase)
+      rescue URI::InvalidURIError
+        false
       end
 
     private

data/lib/omniai/google/version.rb

@@ -2,6 +2,6 @@
 
 module OmniAI
   module Google
-    VERSION = "3.15.0"
+    VERSION = "3.17.0"
   end
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: omniai-google
 version: !ruby/object:Gem::Version
-  version: 3.15.0
+  version: 3.17.0
 platform: ruby
 authors:
 - Kevin Sylvestre