Table of Contents

1. Provider Data Retention
   1. OpenAI Responses API
   2. OpenRouter
      1. Zero data retention
   3. Overriding a setting

Provider Data Retention

Two providers store the prompts they are sent by default. Raif sends an explicit parameter to each one, so a host app’s posture does not depend on a setting Raif cannot see.

Each setting can be overridden globally, per class, or per request.

Setting	Default	Sends	Applies to
open_ai_store_responses	false	store	Raif::Llms::OpenAiResponses
open_router_data_collection	"deny"	provider.data_collection	Raif::Llms::OpenRouter
open_router_zdr	false	provider.zdr	Raif::Llms::OpenRouter

Neither provider’s parameter amounts to zero retention on its own. Read the caveat in each section below before you treat one as a guarantee.

OpenAI Responses API

The Responses API defaults store to true. OpenAI then keeps the response object for 30 days, where the prompt is readable in the organization’s dashboard and retrievable by response id. Raif sends store: false instead.

Raif.configure do |config|
  # Set to true for dashboard visibility and provider-side response retrieval.
  config.open_ai_store_responses = false
end

The Chat Completions API already defaults store to false, so Raif::Llms::OpenAiCompletions sends nothing and this setting does not apply to it.

store: false is not zero retention. OpenAI generates abuse monitoring logs for all API usage and holds them for up to 30 days, whatever store says. Only an organization approved for Zero Data Retention or Modified Abuse Monitoring escapes that, and under Zero Data Retention OpenAI treats store as false regardless of what the request asks for - so setting this to true there has no effect.

OpenRouter

OpenRouter routes each request to an upstream inference provider. Some of those providers store prompts non-transiently and train on them. OpenRouter defaults data_collection to "allow", which filters nothing, so without an explicit preference the only thing narrowing the routing is the account-level privacy setting on openrouter.ai, which the host app cannot read from its own code.

Raif sends provider: { data_collection: "deny" }, which routes only to providers that do not collect user data.

Raif.configure do |config|
  config.open_router_data_collection = "deny"
end

"deny" narrows the set of endpoints OpenRouter will route to. A model whose only endpoints collect data returns a no-endpoints error rather than a completion.

"allow" cannot widen the routing past the account-level setting. Request preferences narrow within it rather than override it, so an account that denies data collection stays denied whatever a request asks for. Set the account-level setting to match your intent as well.

Zero data retention

data_collection is about training and non-transient storage, not retention. OpenRouter states it has no routing rules based on providers’ data retention policies, so a provider that holds prompts for a fixed window without training on them still satisfies "deny".

provider: { zdr: true } is the control that restricts routing to endpoints with a Zero Data Retention policy. Raif sends it only when open_router_zdr is true:

Raif.configure do |config|
  config.open_router_zdr = true
end

It defaults to false because it is a far narrower filter than data_collection. Many models have no ZDR endpoint at all and will return a no-endpoints error once it is on. Check the models you use before enabling it.

The parameter is omitted rather than sent as false, since OpenRouter treats false and absent alike. A per-request zdr is OR’d with the account-wide and guardrail ZDR settings, so it can only add enforcement, never remove it.

Overriding a setting

A Raif::Task, Raif::Conversation or Raif::Agent subclass can override either setting for every request it makes:

class Raif::Tasks::PublicSummary < Raif::Task
  self.open_ai_store_responses = true

  def build_prompt
    "Summarize the following press release: ..."
  end
end

Raif::Llm#chat takes the same three keywords for a one-off override:

Raif.llm(:open_ai_responses_gpt_4o).chat(
  message: "Summarize the following press release: ...",
  open_ai_store_responses: true
)

An override is persisted on the Raif::ModelCompletion, in its request_settings jsonb column. An absent key means “use the Raif.config value”, so only a deliberate override is recorded. Raif::ModelCompletion::REQUEST_SETTING_KEYS lists every key the column may carry, and both keys and values are validated.

Persisting matters for batch inference: submission reloads its completions from the database, in a later process, and builds the provider request from what it finds there.