Use JSONL for streaming rather then SSE

[pmeier]

Looking at the builtin assistants, there are currently three ways of streaming:

1. Using SSE
   • Anthropic
   • OpenAI
2. Using streaming responses
   1. with one object per line (JSONL):
      • Cohere
   2. with streaming JSON:
      • Google

Out of those, I think 2.ii. is by far the worst as it is the hardest to parse. When we implemented streaming for Ragna, I went with 1., since that is what I was familiar with and seen other providers to as well. At that point, I was not aware of streaming JSONL (2.i.) as an alternative.

2.i. has the upside, that one does not need an extra library to handle as it is "required" for the other two cases (of course we could write the logic ourselves as well).

Due to its simplicity I would actually prefer JSONL streaming. We should make a decision here before we release streaming with 0.2.0.

[pmeier]

I've added a PR for this in #357. We eliminate the dependency on sse-starlette, httpx_sse becomes an optional dependency for Anthropic and OpenAI assistants, and we simplify the streaming code on the user side from

ragna/docs/examples/gallery_streaming.py

Lines 151 to 162 in 4a667f9

	import httpx_sse
	import json
	chunks = []
	with httpx_sse.connect_sse(
	client,
	"POST",
	f"/chats/{chat['id']}/answer",
	json={"prompt": "What is Ragna?", "stream": True},
	) as event_source:
	for sse in event_source.iter_sse():
	chunks.append(json.loads(sse.data))

to

import json

with client.stream(
    "POST",
    f"/chats/{chat['id']}/answer",
    json={"prompt": "What is Ragna?", "stream": True},
) as response:
    chunks = [json.loads(data) for data in response.iter_lines()]

I currently don't see a downside.

[aktech]

That's very nicely explained issue, thanks for doing that!