Skip to content

armgit / uzu

Public Imported from GitHub
0

A high-performance inference engine for AI models

Report length or content_filter for incomplete responses replies (#1013)

Leoy86acbf7a7c77
diff --git a/crates/legacy/backend-remote/src/openai/bridging/responses/finish_reason.rs b/crates/legacy/backend-remote/src/openai/bridging/responses/finish_reason.rs
index 200836065..66fbe31c7 100644
--- a/crates/legacy/backend-remote/src/openai/bridging/responses/finish_reason.rs
+++ b/crates/legacy/backend-remote/src/openai/bridging/responses/finish_reason.rs
@@ -1,10 +1,100 @@
-use async_openai::types::responses::Status;
+use async_openai::types::responses::{Response, Status};
use shoji::types::session::chat::ChatReplyFinishReason;
-pub fn build(status: &Status) -> Option<ChatReplyFinishReason> {
- match status {
+pub fn build(response: &Response) -> Option<ChatReplyFinishReason> {
+ match &response.status {
Status::Completed | Status::Failed => Some(ChatReplyFinishReason::Stop),
Status::Cancelled => Some(ChatReplyFinishReason::Cancelled),
+ Status::Incomplete => match response.incomplete_details.as_ref().map(|details| details.reason.as_str()) {
+ Some("max_output_tokens") => Some(ChatReplyFinishReason::Length),
+ Some("content_filter") => Some(ChatReplyFinishReason::Rejected),
+ _ => None,
+ },
_ => None,
}
}
+
+#[cfg(test)]
+mod tests {
+ use async_openai::types::responses::{Response, ResponseStreamEvent};
+ use shoji::types::session::chat::ChatReplyFinishReason;
+
+ use super::build;
+ use crate::openai::bridging::responses::stream_chunk;
+
+ /// Builds a `Response` from the provider's wire shape rather than from the
+ /// struct, so the field names the mapping depends on stay pinned to what the
+ /// API actually sends.
+ fn response(
+ status: &str,
+ incomplete_reason: Option<&str>,
+ ) -> Response {
+ let value = serde_json::json!({
+ "id": "resp_test",
+ "object": "response",
+ "created_at": 1_700_000_000_u64,
+ "model": "test-model",
+ "output": [],
+ "status": status,
+ "incomplete_details": incomplete_reason.map(|reason| serde_json::json!({ "reason": reason })),
+ });
+ serde_json::from_value(value).expect("response fixture should deserialize")
+ }
+
+ #[test]
+ fn terminal_statuses_map_to_stop() {
+ assert_eq!(build(&response("completed", None)), Some(ChatReplyFinishReason::Stop));
+ assert_eq!(build(&response("failed", None)), Some(ChatReplyFinishReason::Stop));
+ }
+
+ #[test]
+ fn cancelled_maps_to_cancelled() {
+ assert_eq!(build(&response("cancelled", None)), Some(ChatReplyFinishReason::Cancelled));
+ }
+
+ #[test]
+ fn non_terminal_statuses_have_no_finish_reason() {
+ assert_eq!(build(&response("in_progress", None)), None);
+ assert_eq!(build(&response("queued", None)), None);
+ }
+
+ #[test]
+ fn truncated_response_maps_to_length() {
+ assert_eq!(build(&response("incomplete", Some("max_output_tokens"))), Some(ChatReplyFinishReason::Length));
+ }
+
+ #[test]
+ fn content_filtered_response_maps_to_rejected() {
+ assert_eq!(build(&response("incomplete", Some("content_filter"))), Some(ChatReplyFinishReason::Rejected));
+ }
+
+ #[test]
+ fn unrecognised_incomplete_reason_has_no_finish_reason() {
+ assert_eq!(build(&response("incomplete", None)), None);
+ assert_eq!(build(&response("incomplete", Some("something_else"))), None);
+ }
+
+ /// Drives the real event path, so a truncated reply cannot silently surface as
+ /// a natural stop again.
+ #[test]
+ fn incomplete_event_emits_length_finish_reason() {
+ let event: ResponseStreamEvent = serde_json::from_value(serde_json::json!({
+ "type": "response.incomplete",
+ "sequence_number": 3,
+ "response": {
+ "id": "resp_test",
+ "object": "response",
+ "created_at": 1_700_000_000_u64,
+ "model": "test-model",
+ "output": [],
+ "status": "incomplete",
+ "incomplete_details": { "reason": "max_output_tokens" },
+ },
+ }))
+ .expect("response.incomplete event fixture should deserialize");
+
+ let chunk = stream_chunk::build(event).expect("a response.incomplete event yields a chunk");
+
+ assert_eq!(chunk.finish_reason, Some(ChatReplyFinishReason::Length));
+ }
+}
diff --git a/crates/legacy/backend-remote/src/openai/bridging/responses/stream_chunk.rs b/crates/legacy/backend-remote/src/openai/bridging/responses/stream_chunk.rs
index f1a9fba28..e39c20b1e 100644
--- a/crates/legacy/backend-remote/src/openai/bridging/responses/stream_chunk.rs
+++ b/crates/legacy/backend-remote/src/openai/bridging/responses/stream_chunk.rs
@@ -49,7 +49,7 @@ pub fn build(event: ResponseStreamEvent) -> Option<StreamChunk> {
let (tokens_input, tokens_input_cached, tokens_output) =
event.response.usage.as_ref().map(usage_parts).unwrap_or_default();
Some(StreamChunk {
- finish_reason: finish_reason::build(&event.response.status),
+ finish_reason: finish_reason::build(&event.response),
tokens_input,
tokens_input_cached,
tokens_output,
@@ -60,7 +60,7 @@ pub fn build(event: ResponseStreamEvent) -> Option<StreamChunk> {
let (tokens_input, tokens_input_cached, tokens_output) =
event.response.usage.as_ref().map(usage_parts).unwrap_or_default();
Some(StreamChunk {
- finish_reason: finish_reason::build(&event.response.status),
+ finish_reason: finish_reason::build(&event.response),
tokens_input,
tokens_input_cached,
tokens_output,
@@ -71,7 +71,7 @@ pub fn build(event: ResponseStreamEvent) -> Option<StreamChunk> {
let (tokens_input, tokens_input_cached, tokens_output) =
event.response.usage.as_ref().map(usage_parts).unwrap_or_default();
Some(StreamChunk {
- finish_reason: finish_reason::build(&event.response.status),
+ finish_reason: finish_reason::build(&event.response),
tokens_input,
tokens_input_cached,
tokens_output,