60759b955c
RubyLLM auto-detecta provider pelo prefixo do nome do modelo (ex: `gemini-*` → provider Gemini → exige `gemini_api_key`). Quando temos config dedicada de embedding (CAPTAIN_EMBEDDING_API_KEY) apontando pra endpoint OpenAI-compatible (ex: Gemini OpenAI-compat em generativelanguage.googleapis.com/v1beta/openai), queremos que o RubyLLM mande a request via OpenAI client mesmo que o nome do modelo bata com outro provider. Solução: passar provider: :openai e assume_model_exists: true ao chamar embed quando dedicated_embedding_config? retornar true. Sem isso, o RubyLLM falha com `Missing configuration for Gemini: gemini_api_key` mesmo com a key correta setada. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
91 lines
3.1 KiB
Ruby
91 lines
3.1 KiB
Ruby
class Captain::Llm::EmbeddingService
|
|
include Integrations::LlmInstrumentation
|
|
|
|
class EmbeddingsError < StandardError; end
|
|
|
|
def initialize(account_id: nil)
|
|
Llm::Config.initialize!
|
|
@account_id = account_id
|
|
@embedding_model = InstallationConfig.find_by(name: 'CAPTAIN_EMBEDDING_MODEL')&.value.presence || LlmConstants::DEFAULT_EMBEDDING_MODEL
|
|
end
|
|
|
|
def self.embedding_model
|
|
InstallationConfig.find_by(name: 'CAPTAIN_EMBEDDING_MODEL')&.value.presence || LlmConstants::DEFAULT_EMBEDDING_MODEL
|
|
end
|
|
|
|
def get_embedding(content, model: @embedding_model)
|
|
return [] if content.blank?
|
|
|
|
instrument_embedding_call(instrumentation_params(content, model)) do
|
|
embed_with_legacy_openai(content, model)
|
|
end
|
|
rescue RubyLLM::Error => e
|
|
Rails.logger.error "Embedding API Error: #{e.message}"
|
|
raise EmbeddingsError, "Failed to create an embedding: #{e.message}"
|
|
end
|
|
|
|
private
|
|
|
|
# Embeddings vão pra OpenAI tradicional por default (o endpoint Codex
|
|
# via ChatGPT OAuth não expõe /embeddings). Override opcional via env vars
|
|
# dedicadas — útil pra trocar provider de embedding (ex: Gemini
|
|
# OpenAI-compatible) sem alterar o provider de chat:
|
|
#
|
|
# CAPTAIN_EMBEDDING_API_KEY — sobrescreve API key
|
|
# CAPTAIN_EMBEDDING_ENDPOINT — sobrescreve base URL (sem /v1 no final)
|
|
# CAPTAIN_EMBEDDING_DIMENSIONS — força reduction (ex: 1536 pra Gemini
|
|
# bater com schema pgvector(1536))
|
|
def embed_with_legacy_openai(content, model)
|
|
settings = embedding_settings
|
|
api_base = settings[:api_base].present? ? "#{settings[:api_base]}/v1" : nil
|
|
embed_options = embed_extra_options
|
|
|
|
# Quando há config dedicada de embedding (CAPTAIN_EMBEDDING_API_KEY etc),
|
|
# forçamos provider :openai pra que o RubyLLM trate como OpenAI-compatible
|
|
# mesmo com modelos cujo nome auto-detectaria outro provider (ex:
|
|
# `gemini-embedding-001` apontado pro endpoint Gemini OpenAI-compat).
|
|
embed_options[:provider] = :openai if dedicated_embedding_config?
|
|
embed_options[:assume_model_exists] = true if dedicated_embedding_config?
|
|
|
|
Llm::Config.with_api_key(settings[:api_key], api_base: api_base) do |ctx|
|
|
ctx.embed(content, model: model, **embed_options).vectors
|
|
end
|
|
end
|
|
|
|
def dedicated_embedding_config?
|
|
installation_config_value('CAPTAIN_EMBEDDING_API_KEY').present?
|
|
end
|
|
|
|
def embedding_settings
|
|
custom_key = installation_config_value('CAPTAIN_EMBEDDING_API_KEY')
|
|
return Captain::Llm::ProviderConfig.legacy_openai_settings if custom_key.blank?
|
|
|
|
{
|
|
api_key: custom_key,
|
|
api_base: installation_config_value('CAPTAIN_EMBEDDING_ENDPOINT')&.chomp('/')
|
|
}
|
|
end
|
|
|
|
def embed_extra_options
|
|
dims = installation_config_value('CAPTAIN_EMBEDDING_DIMENSIONS')
|
|
return {} if dims.blank?
|
|
|
|
{ dimensions: dims.to_i }
|
|
end
|
|
|
|
def installation_config_value(name)
|
|
ENV.fetch(name, nil) ||
|
|
InstallationConfig.find_by(name: name)&.value
|
|
end
|
|
|
|
def instrumentation_params(content, model)
|
|
{
|
|
span_name: 'llm.captain.embedding',
|
|
model: model,
|
|
input: content,
|
|
feature_name: 'embedding',
|
|
account_id: @account_id
|
|
}
|
|
end
|
|
end
|