Package org.springframework.ai.openai.responses

Class OpenAiResponsesMetadata

java.lang.Object

org.springframework.ai.openai.responses.OpenAiResponsesMetadata

public final class OpenAiResponsesMetadata
extends Object

Metadata keys published by OpenAiResponsesChatModel, and the part attributes it uses.

There are two groups. What a turn reports is published on the generation or response metadata; what a turn needs to be replayed travels in the part attributes and payloads.

The assistant message carries no metadata of its own. What a turn reports beyond its parts describes one response rather than the message, and message metadata is persisted by some chat memory repositories and dropped by others, so nothing that matters may live there.

The openai.responses.* keys are specific to this API. REFUSAL, ANNOTATIONS and REASONING_CONTENT keep the unprefixed names OpenAiChatModel publishes on its assistant message, so code moving between the two beans only has to change where it reads them, not what it looks up.

Since:

2.1.0

Author:

Dimitar Proynov

Field Summary

Fields

Modifier and Type	Field	Description
static final String	ANNOTATIONS	Citations attached to the generated text, e.g. from web or file search.
static final String	CREATED_AT	The response creation timestamp, in seconds since the epoch.
static final String	ENCRYPTED_CONTENT_KIND	The kind of the payload carried by a reasoning part: the encrypted_content of a reasoning item, which has to travel back verbatim for the model to keep its train of thought across a tool call or a follow-up turn.
static final String	HOSTED_TOOL_CALLS	Summary of the tools OpenAI executed server-side within the request, as a list of {type, id, status} maps.
static final String	INCOMPLETE_REASON	Why an incomplete response stopped: max_output_tokens or content_filter.
static final String	ITEM_ID_ATTRIBUTE	Part attribute holding the id of the output item a part was mapped from, e.g.
static final String	ITEM_STATUS_ATTRIBUTE	Part attribute holding the status of the output item a part was mapped from: in_progress, completed or incomplete.
static final String	PHASE_ATTRIBUTE	Part attribute holding the phase of an assistant message item (commentary or final_answer), which GPT-5.5 and later return to mark whether a message was a final answer or an intermediate update.
static final String	PROVIDER	The provider of every payload and every UnknownPart this model produces, and the only one it replays.
static final String	REASONING_CONTENT	Reasoning summary text, one line per reasoning item.
static final String	REFUSAL	The model's refusal text, or an empty string.
static final String	STATUS	The response status: completed, incomplete, failed, ...

Method Summary

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Details

PROVIDER

public static final String PROVIDER

The provider of every payload and every UnknownPart this model produces, and the only one it replays.

The endpoint, not the vendor. Chat Completions and Responses are two OpenAI APIs whose replay payloads are not interchangeable, and a conversation can move between them, so naming the endpoint is what keeps either one from replaying the other's data if their handling of payload-carrying parts ever converges.

See Also:

• Constant Field Values

ENCRYPTED_CONTENT_KIND

public static final String ENCRYPTED_CONTENT_KIND

The kind of the payload carried by a reasoning part: the encrypted_content of a reasoning item, which has to travel back verbatim for the model to keep its train of thought across a tool call or a follow-up turn.

See Also:

• Constant Field Values

ITEM_ID_ATTRIBUTE

public static final String ITEM_ID_ATTRIBUTE

Part attribute holding the id of the output item a part was mapped from, e.g. rs_... for a reasoning item or fc_... for a function call.

This is the item id, which addresses the item's position in the transcript, never the call_id that pairs a call with its result. A reasoning part cannot be replayed without it.

See Also:

• Constant Field Values

ITEM_STATUS_ATTRIBUTE

public static final String ITEM_STATUS_ATTRIBUTE

Part attribute holding the status of the output item a part was mapped from: in_progress, completed or incomplete.

See Also:

• Constant Field Values

PHASE_ATTRIBUTE

public static final String PHASE_ATTRIBUTE

Part attribute holding the phase of an assistant message item (commentary or final_answer), which GPT-5.5 and later return to mark whether a message was a final answer or an intermediate update.

Replayed so the model does not mistake a prior intermediate update for a final answer.

See Also:

• Constant Field Values

STATUS

public static final String STATUS

The response status: completed, incomplete, failed, ...

See Also:

• Constant Field Values

INCOMPLETE_REASON

public static final String INCOMPLETE_REASON

Why an incomplete response stopped: max_output_tokens or content_filter.

See Also:

• Constant Field Values

HOSTED_TOOL_CALLS

public static final String HOSTED_TOOL_CALLS

Summary of the tools OpenAI executed server-side within the request, as a list of {type, id, status} maps. These are never surfaced as tool calls.

The summary describes activity inside a single request, so it is neither replayed on the next turn nor worth persisting with the conversation. The authoritative copy of each item is the UnknownPart it was mapped to, which holds the item verbatim and is replayed as-is.

See Also:

• Constant Field Values

CREATED_AT

public static final String CREATED_AT

The response creation timestamp, in seconds since the epoch.

See Also:

• Constant Field Values

REFUSAL

public static final String REFUSAL

The model's refusal text, or an empty string.

There is no refusal part type, so this is the only place a refusal surfaces. On a stream it is the running total, published on every chunk that extends it, which is what lets it survive aggregation.

See Also:

• Constant Field Values

ANNOTATIONS

public static final String ANNOTATIONS

Citations attached to the generated text, e.g. from web or file search.

The key matches OpenAiChatModel, but the value shape does not: this model publishes a List<Map<String, Object>>, whereas OpenAiChatModel publishes the OpenAI SDK's own annotation objects. Code that casts the entries has to be adjusted when swapping one bean for the other.

See Also:

• Constant Field Values

REASONING_CONTENT

public static final String REASONING_CONTENT

Reasoning summary text, one line per reasoning item. OpenAI never returns raw reasoning text, only summaries and an encrypted blob.

Published under the key OpenAiChatModel and DeepSeek surface reasoning under. It is a flattened view: the authoritative form is the ordered ReasoningParts on the assistant message, which keep each item separate, in place among the tool calls it justified, and carry the encrypted content needed to replay it.

See Also:

• Constant Field Values