OpenAI error payloads in the body may not get parsed fully during chat streaming

[atesgoral]

Describe the bug

With OpenAI chat streaming, if OpenAI returns an HTTP error (e.g. 400), the error handler optimistically tries to parse the first chunk it receives as the entire body. However, networks conditions or an intermediary like an OpenAI proxy may cause the response to be fragmented into different chunks.

The symptom is an error being raised with just the first chunk of the body. While this is not a very severe problem, since it happens during what's already an API call failure (that should ideally be rare), not collecting and parsing the full error would make it hard for application developers to figure out exactly what went wrong.

The issue is here, since #344:

ruby-openai/lib/openai/http.rb

Line 64 in 0a897ba

raise_error.on_complete(env.merge(body: try_parse_json(chunk)))

To Reproduce

Hard to reproduce in vivo, unless you're running a special HTTP/TCP proxy that chunks the original TCP packets for you, or have poor network conditions, etc.

Minimal reproducible unit test: #480

-"error" => {"code"=>"test", "message"=>"Test error", "param"=>nil, "type"=>"test_error"},
+"{\"error\":{\"message\":\"Test error\",\"type\":\""

Expected behavior

The full response body is collected (chunks concatenated) and parsed as JSON.