Class: OmniAI::Google::Embed

Inherits:

Embed

• Object
• Embed
• OmniAI::Google::Embed

Defined in:

lib/omniai/google/embed.rb

Overview

A Google embed implementation.

Usage:

input = "..."
response = OmniAI::Google::Embed.process!(input, client: client)
response.embedding [0.0, ...]

Defined Under Namespace

Modules: Model

Constant Summary

DEFAULT_MODEL =

Model::EMBEDDING

BATCH_EMBED_CONTENTS_DESERIALIZER =

proc do |data, *|
  data["embeddings"].map { |embedding| embedding["values"] }
end

PREDICT_EMBEDDINGS_DESERIALIZER =

proc do |data, *|
  data["predictions"].map { |prediction| prediction["embeddings"]["values"] }
end

PREDICT_USAGE_DESERIALIZER =

proc do |data, *|
  tokens = data["predictions"].sum { |prediction| prediction["embeddings"]["statistics"]["token_count"] }

  Usage.new(prompt_tokens: tokens, total_tokens: tokens)
end

EMBED_CONTENT_DESERIALIZER =

proc do |data, *|
  [data["embedding"]["values"]]
end

USAGE_METADATA_DESERIALIZER =

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.

proc do |data, *|
  prompt_tokens = data.dig("usageMetadata", "promptTokenCount")
  total_tokens = data.dig("usageMetadata", "totalTokenCount")

  Usage.new(prompt_tokens:, total_tokens:)
end

BATCH_EMBED_CONTENTS_CONTEXT =

Returns:

• (Context)

Context.build do |context|
  context.deserializers[:embeddings] = BATCH_EMBED_CONTENTS_DESERIALIZER
  context.deserializers[:usage] = USAGE_METADATA_DESERIALIZER
end

PREDICT_CONTEXT =

Returns:

• (Context)

Context.build do |context|
  context.deserializers[:embeddings] = PREDICT_EMBEDDINGS_DESERIALIZER
  context.deserializers[:usage] = PREDICT_USAGE_DESERIALIZER
end

EMBED_CONTENT_CONTEXT =

Returns:

• (Context)

Context.build do |context|
  context.deserializers[:embeddings] = EMBED_CONTENT_DESERIALIZER
  context.deserializers[:usage] = USAGE_METADATA_DESERIALIZER
end