feat(rag): implement real-time token streaming via SSE and FastAPI background tasks

[YTxFSGAMERz]

Summary

This PR resolves a significant UX bottleneck in the /rag/query/stream endpoint by implementing true auto-regressive Server-Sent Events (SSE).

Previously, the endpoint awaited the complete text generation from the LLM before yielding chunks for cosmetic UI streaming. By integrating sse-starlette and converting the Ollama client into a real-time generator, tokens are now streamed immediately as they arrive over an EventSourceResponse, drastically reducing perceived latency.

Architectural Changes

• Dependencies: Added sse-starlette to manage the EventSource streaming protocol standard.
• LLM Service (app/rag/llm.py): Implemented generate_answer_stream using requests.post(stream=True) to yield JSON-decoded tokens asynchronously.
• FastAPI Routes (app/api/routes/rag.py): Offloaded the blocking generator to a background thread (asyncio.to_thread) and wrapped the output in an EventSourceResponse. Background metadata tracking via _dispatch_background remains intact.

Note: Submitted as part of GirlScript Summer of Code (GSSoC) 2026.

[vercel]

@YTxFSGAMERz is attempting to deploy a commit to the devloperdevesh's projects Team on Vercel.

A member of the Team first needs to authorize it.