Handling non-JSON payloads with asDecodedServerSentEventsWithJSONData()

[paulhdk]

Question

Hi!

I'm generating code with the OpenAI OpenAPI specification and am iterating over streamed responses from the chat/completions endpoint with asDecodedServerSentEventsWithJSONData().

The problem is that API returns ChatCompletionStreamResponse JSON payloads until the stream is done, which is denoted by a simple [DONE] payload.

Because [DONE] isn't valid JSON, the last chunk of a successful stream will throw a DecodingError.dataCorrupted().

Currently, I'm just handling it with an empty catch block to avoid the crash, but that seems like a hacky workaround. So, I was wondering whether there was a better solution.

The following ideas haven't been successful so far:

• Use a ClientMiddleware to match incoming payloads for the [DONE] bytes and toss the payload if they're found.
• Add another schema to the OpenAPI spec that defines [DONE] and is connect with the ChatCompletionStreamResponse through an anyOf.

Do you have any suggestions?

[czechboy0]

Hi @paulhdk,

I think you're looking for https://swiftpackageindex.com/apple/swift-openapi-runtime/1.5.0/documentation/openapiruntime/_concurrency/asyncsequence/asdecodedserversentevents()

[simonjbeaumont]

I've just downloaded the latest OpenAPI document for OpenAI, and I think we might need some changes. It currently provides application/json as the content-type regardless of whether the streaming parameter is set or not:

            responses:
                "200":
                    description: OK
                    content:
                        application/json:
                            schema:
                                $ref: "#/components/schemas/CreateChatCompletionResponse"

It returns the following body:

data: {"id":"chatcmpl-A6aNCrs7LgMTsQlex5k1tdPIDfPfg","object":"chat.completion.chunk","created":1726133150,"model":"gpt-4o-mini-2024-07-18","system_fingerprint":"fp_483d39d857","choices":[{"index":0,"delta"
:{"role":"assistant","content":"","refusal":null},"logprobs":null,"finish_reason":null}]}

data: {"id":"chatcmpl-A6aNCrs7LgMTsQlex5k1tdPIDfPfg","object":"chat.completion.chunk","created":1726133150,"model":"gpt-4o-mini-2024-07-18","system_fingerprint":"fp_483d39d857","choices":[{"index":0,"delta":{"con
tent":"Yes"},"logprobs":null,"finish_reason":null}]}

data: {"id":"chatcmpl-A6aNCrs7LgMTsQlex5k1tdPIDfPfg","object":"chat.completion.chunk","created":1726133150,"model":"gpt-4o-mini-2024-07-18","system_fingerprint":"fp_483d39d857","choices":[{"index":0,"delta":{"con
tent":","},"logprobs":null,"finish_reason":null}]}

data: {"id":"chatcmpl-A6aNCrs7LgMTsQlex5k1tdPIDfPfg","object":"chat.completion.chunk","created":1726133150,"model":"gpt-4o-mini-2024-07-18","system_fingerprint":"fp_483d39d857","choices":[{"index":0,"delta":{"con
tent":" here"},"logprobs":null,"finish_reason":null}]}

data: {"id":"chatcmpl-A6aNCrs7LgMTsQlex5k1tdPIDfPfg","object":"chat.completion.chunk","created":1726133150,"model":"gpt-4o-mini-2024-07-18","system_fingerprint":"fp_483d39d857","choices":[{"index":0,"delta":{"con
tent":"."},"logprobs":null,"finish_reason":null}]}

data: {"id":"chatcmpl-A6aNCrs7LgMTsQlex5k1tdPIDfPfg","object":"chat.completion.chunk","created":1726133150,"model":"gpt-4o-mini-2024-07-18","system_fingerprint":"fp_483d39d857","choices":[{"index":0,"delta":{},"l
ogprobs":null,"finish_reason":"stop"}]}

data: [DONE]

The problem is that our streaming helpers in OpenAPI runtime operate as extensions on AsyncSequence<ArraySlice<UInt8>>, which means they're only available when the generated code is dealing with a binary response body.

This works fine when the documented content type is text/event-stream but with application/json we don't have access to these helpers.

This appears to be an OpenAPI doc issue because it is actually returning a text/event-stream, which trips the runtime validation in any event.

Swift/ErrorType.swift:253: Fatal error: Error raised at top level: Client error - cause description: 'Unexpected content type, expected: application/json, received: text/event-stream; charset=utf-8', underlying error: Unexpected content type, expected: application/json, received: text/event-stream; ...

Let me play with this a bit more and report back, it might be we just need the doc to be patched. ISTR this used to work :)

[simonjbeaumont]

OK, so I think there's potentially a combination of issues here:

1. The OpenAPI document for this API specifies application/json as the content-type but will actually send text/event-stream based on a boolean parameter.
2. The server is sending a [DONE] at the end of the SSE stream, which doesn't decode using the asDecodedServerSentEventsWithJSONData helper.

(1) will need to be addressed in the upstream OpenAPI doc, by just adding another content-type to the response, which I have done locally. In the meantime, you might be able to swizzle the content-type in a client middleware.

--- a/openapi.yaml
+++ b/openapi.yaml
@@ -61,6 +61,9 @@ paths:
                         application/json:
                             schema:
                                 $ref: "#/components/schemas/CreateChatCompletionResponse"
+                        text/event-stream:
+                            schema:
+                                $ref: "#/components/schemas/CreateChatCompletionResponse"

             x-oaiMeta:
                 name: Create chat completion

Once (1) is addressed, you immediately run into (2), where it will happily decode all the events but will get stuck on the last event, [DONE]:

Swift.DecodingError.dataCorrupted(Swift.DecodingError.Context(codingPath: [], debugDescription: "The given data was not valid JSON.", underlyingError: Optional(Error Domain=NSCocoaErrorDomain Code=3840 "Unexpected character 'D' around line 1, column 2." UserInfo={NSDebugDescription=Unexpected character 'D' around line 1, column 2., NSJSONSerializationErrorIndex=1})))

I have reproduced this in a mock client now and the workaround is to decode it using asDecodedServerSentEvents, filter out the [DONE] event, re-encode, and then decode again as the expected type:

// Replace this...
let responses = try response.ok.body.text_event_hyphen_stream
    .asDecodedServerSentEventsWithJSONData(of: Components.Schemas.CreateChatCompletionStreamResponse.self)

```swift
// ...with this...
let responses = try response.ok.body.text_event_hyphen_stream
    .asDecodedServerSentEvents()
    .filter { $0.data != "[DONE]" }
    .asEncodedServerSentEvents()
    .asDecodedServerSentEventsWithJSONData(of:
        Components.Schemas.CreateChatCompletionStreamResponse.self)

The re-encode/decode step will be unnecessary cost, and you could just encode with a map but the above will mean you don't have to bother with providing your own appropriately configured JSONDecoder.

I'll take a note to see if this [DONE] should work out-of-the box.

[paulhdk]

OK, so I think there's potentially a combination of issues here:

1. The OpenAPI document for this API specifies application/json as the content-type but will actually send text/event-stream based on a boolean parameter.

@czechboy0 and I had actually worked that one out already in #614. I submitted a PR to the OpenAI OpenAPI repo, but haven't heard anything yet. Should've mentioned that in my initial issue description ...

2. The server is sending a [DONE] at the end of the SSE stream, which doesn't decode using the asDecodedServerSentEventsWithJSONData helper.

Yes, that's exactly the problem I'm facing.

(1) will need to be addressed in the upstream OpenAPI doc, by just adding another content-type to the response, which I have done locally. In the meantime, you might be able to swizzle the content-type in a client middleware.

--- a/openapi.yaml
+++ b/openapi.yaml
@@ -61,6 +61,9 @@ paths:
                         application/json:
                             schema:
                                 $ref: "#/components/schemas/CreateChatCompletionResponse"
+                        text/event-stream:
+                            schema:
+                                $ref: "#/components/schemas/CreateChatCompletionResponse"

             x-oaiMeta:
                 name: Create chat completion

Once (1) is addressed, you immediately run into (2), where it will happily decode all the events but will get stuck on the last event, [DONE]:

Swift.DecodingError.dataCorrupted(Swift.DecodingError.Context(codingPath: [], debugDescription: "The given data was not valid JSON.", underlyingError: Optional(Error Domain=NSCocoaErrorDomain Code=3840 "Unexpected character 'D' around line 1, column 2." UserInfo={NSDebugDescription=Unexpected character 'D' around line 1, column 2., NSJSONSerializationErrorIndex=1})))

I have reproduced this in a mock client now and the workaround is to decode it using asDecodedServerSentEvents, filter out the [DONE] event, re-encode, and then decode again as the expected type:

// Replace this...
let responses = try response.ok.body.text_event_hyphen_stream
    .asDecodedServerSentEventsWithJSONData(of: Components.Schemas.CreateChatCompletionStreamResponse.self)

```swift
// ...with this...
let responses = try response.ok.body.text_event_hyphen_stream
    .asDecodedServerSentEvents()
    .filter { $0.data != "[DONE]" }
    .asEncodedServerSentEvents()
    .asDecodedServerSentEventsWithJSONData(of:
        Components.Schemas.CreateChatCompletionStreamResponse.self)

Brilliant! That works.
Again, thank you both for your quick and helpful responses.

The re-encode/decode step will be unnecessary cost, and you could just encode with a map but the above will mean you don't have to bother with providing your own appropriately configured JSONDecoder.

I'll take a note to see if this [DONE] should work out-of-the box.

I'll go ahead and close the issue though? Or would you prefer to keep it open?

[simonjbeaumont]

@czechboy0 and I had actually worked that one out already ... Should've mentioned that in my initial issue description

Ah, well. Now I've been on the voyage of discovery too 😅

Brilliant! That works. Again, thank you both for your quick and helpful responses.

You're very welcome—glad to be of help 👍

The re-encode/decode step will be unnecessary cost, and you could just encode with a map but the above will mean you don't have to bother with providing your own appropriately configured JSONDecoder.
I'll take a note to see if this [DONE] should work out-of-the box.

I'll go ahead and close the issue though? Or would you prefer to keep it open?

Leave it open for now as a reminder for me. Tomorrow I'll close it and open one with a focussed description to investigate what, if anything, we need to do about the [DONE], in general.

[simonjbeaumont]

So I thought it would be fun to have a chat with ChatGPT itself about the API choice to terminate an SSE stream where all the payloads are JSON with a non-JSON value. It acknowledges that this is both unconventional and not very helpful for processing pipelines that are decoding the event stream :)

This was after I did my own research and came to the following conclusions:

1. WHATWG HTML Living Standard, 9.2 Server-sent events doesn't specify anything about how to terminate the stream.
2. Many APIs that offer SSE terminate the stream by closing the connection.
3. Looks like this [DONE] thing has also been adopted by Anthropic, but outside of the OpenAI and Anthropic APIs, I can't see any other use of this.

If there were wider precedent for it, I think we could adapt the helpers in some way.

The only thing I can think of currently is to add an optional parameter to asDecodedServerSentEventsWithJSONData(of:), to specify a known terminal message, which would allow you to call it like this:

.asDecodedServerSentEventsWithJSONData(of: ExpectedType.self, streamTerminalData: "[DONE]")

Better yet, maybe a boolean closure parameter.

I'll park this here and see if more data points come up.

[paulhdk]

So I thought it would be fun to have a chat with ChatGPT itself about the API choice to terminate an SSE stream where all the payloads are JSON with a non-JSON value. It acknowledges that this is both unconventional and not very helpful for processing pipelines that are decoding the event stream :)

Well, at least we’re not the only ones who think so then … :-)

This was after I did my own research and came to the following conclusions:

1. WHATWG HTML Living Standard, 9.2 Server-sent events doesn't specify anything about how to terminate the stream.
2. Many APIs that offer SSE terminate the stream by closing the connection.
3. Looks like this [DONE] thing has also been adopted by Anthropic, but outside of the OpenAI and Anthropic APIs, I can't see any other use of this.

Very interesting!

If there were wider precedent for it, I think we could adapt the helpers in some way.

The only thing I can think of currently is to add an optional parameter to asDecodedServerSentEventsWithJSONData(of:), to specify a known terminal message, which would allow you to call it like this:

.asDecodedServerSentEventsWithJSONData(of: ExpectedType.self, streamTerminalData: "[DONE]")

Better yet, maybe a boolean closure parameter.

Would you mind me giving a PR a go? Or wouldn’t you merge it anyway until there’s more precedent.

I'll park this here and see if more data points come up.

Thank you for doing some more digging on this!

[simonjbeaumont]

If there were wider precedent for it, I think we could adapt the helpers in some way.
The only thing I can think of currently is to add an optional parameter to asDecodedServerSentEventsWithJSONData(of:), to specify a known terminal message, which would allow you to call it like this:

.asDecodedServerSentEventsWithJSONData(of: ExpectedType.self, streamTerminalData: "[DONE]")

Better yet, maybe a boolean closure parameter.

Would you mind me giving a PR a go? Or wouldn’t you merge it anyway until there’s more precedent?

I'm pretty open to that change, so long as it's API-backwards compatible and opt-in, and we're always grateful for folks willing to contribute! 🙏

What do you think @czechboy0?

We should probably discuss the shape of the parameter before you put any time into the coding though. There are at least a few options for it, e.g.:

1. Is the argument used as filter or a terminator?
2. Is the argument provided as a value to match against, or a predicate closure?