Class: PromptBuilder::Serializers::Gemini::Request

Inherits:
Base
  • Object
show all
Defined in:
lib/prompt_builder/serializers/gemini/request.rb

Overview

Request serializer for the Google Gemini API format.

Unsupported Open Responses features

These session fields are not supported and are silently omitted from the serialized output:

  • background — Gemini has no background/async mode on the generate endpoint
  • include — response-field inclusion is an Open Responses-only concept
  • max_tool_calls — per-request tool-call caps are not supported
  • metadata — arbitrary metadata is not supported
  • parallel_tool_calls — parallel tool call control is not supported
  • prompt_cache_key / prompt_cache_retention — explicit prompt cache keys are not supported
  • safety_identifier — no equivalent user-safety field on the generate endpoint
  • stream_options — stream event options are not supported
  • truncation — server-side context truncation is not supported

Input content restrictions:

  • InputImage content is only supported in user messages (assistant images are omitted)
  • InputImage with image_url requires either base64 data or a URL; content without image_url or file_id raises
  • InputFile content is only supported in user messages (assistant files are omitted)
  • InputFile requires media_type when file_data is provided, or a recognized extension on filename / file_url
  • InputVideo requires video_url
  • RefusalContent is dropped silently (a parsed Chat Completions refusal can stay in session history without breaking subsequent request_payload calls)
  • redacted_thinking and unknown reasoning blocks are silently skipped
  • Reasoning items with summary blocks have the summary skipped
  • FunctionCallOutput array contents must be text-only; other content is omitted
  • Compaction and ItemReference items are silently skipped

Features in Gemini not available through Open Responses

The following Gemini parameters cannot be set through the Open Responses canonical format:

  • thinkingConfig.thinkingBudget — use reasoning.budget_tokens instead
  • thinkingConfig.thinkingLevel — use reasoning.effort instead
  • thinkingConfig.includeThoughts — use reasoning.summary = "auto" instead
  • topK — top-K sampling parameter
  • seed — for reproducible outputs (model-dependent)
  • stopSequences — custom stop sequences
  • candidateCount — requesting multiple response candidates
  • responseModalities — selecting TEXT/IMAGE/AUDIO output channels
  • responseJsonSchema — newer JSON Schema variant of responseSchema
  • safetySettings — configurable harm-category safety thresholds
  • mediaResolution — controls token cost of image/video inputs
  • audioTimestamp, speechConfig — audio output configuration
  • enableEnhancedCivicAnswers
  • routingConfig, modelSelectionConfig
  • Video metadata controls (+videoMetadata+ offset, FPS) on +Part+s
  • Audio input model capability controls (audio can be sent as InputFile with an audio MIME type)
  • Built-in Gemini tools: googleSearch / googleSearchRetrieval, codeExecution, urlContext, computerUse, fileSearch, googleMaps, mcpServers
  • functionCallingConfig.mode = VALIDATED
  • toolConfig.retrievalConfig, includeServerSideToolInvocations
  • cachedContent — referencing a CachedContent resource by name
  • Top-level labels (Vertex flavor only)

Constant Summary collapse

SUPPORTED_REASONING_EFFORTS = 
%w[minimal low medium high].freeze
SUPPORTED_REASONING_SUMMARIES = 
%w[auto].freeze
SUPPORTED_SERVICE_TIERS = 
%w[unspecified standard flex priority].freeze

Method Summary

Methods inherited from Base

parse_response, request_payload