[Bug]: Chat History and Dirty-Mark Written to Wrong Session in Multi-Session Stream Requests

[Srejoye]

Summary

Description

POST /ask/stream supports multiple session IDs via data.session_ids. The handler iterates over session_id_list and retrieves candidates from each session (lines 3862–3964).

Affected Files

• rag-service/main.py lines 3862–4148 (ask_question_stream, post-loop session_id usage)

Steps to reproduce

POST /ask/stream
{
  "question": "summarize",
  "session_ids": ["uuid-A", "uuid-B"],
  "session_secrets": ["secret-A", "secret-B"]
}

After the call, session A's chat history is empty; session B contains the message twice (if retrieved from both sessions).

Expected behavior

After the loop, the code uses the bare Python variable session_id — which holds the last value assigned during the loop — for all subsequent operations including:

• append_chat_exchange(current_session, ...) (evidence-gate refusal path)
• _mark_session_dirty(session_id)
• sessions.get(session_id) for persisting the final answer

Actual behavior

When a request spans two sessions (e.g. session_ids: ["A", "B"]), every chat message and dirty mark is written to session "B" regardless of which session actually provided the answer context. Session "A"'s history is never updated.

Additional context

No response

[Srejoye]

Hi — requesting GSSoC'26 assignment.

Technical understanding: The for-loop for idx, session_id in enumerate(session_id_list): leaves session_id set to the last element after completion. All post-loop sessions.get(session_id) lookups (refusal path ~line 3972, and _generate_and_stream closure ~lines 4088–4109) therefore reference only the last session. In a single-session call this is benign; in multi-session calls it produces state desynchronisation.

Affected modules/files:

• rag-service/main.py — ask_question_stream(), all sessions.get(session_id) / _mark_session_dirty(session_id) usages after the retrieval loop

Proposed fix approach: For multi-session streaming, the chat exchange should be appended to the session that yielded the highest-scoring chunk (best contributing session), not whichever session happened to be last in the list. Capture primary_session_id as the session associated with scored_candidates[0] (track it during the loop) and use that variable for all persistence after the loop ends.

Regression/unit test considerations:

• Add a test that issues /ask/stream with two valid session IDs, then calls /sessions/lookup on each and asserts both sessions contain the question/answer pair in their chat lists.
• Add a test confirming _mark_session_dirty is called with the session that contributed the top-ranked chunk.

[garima-agarwall]

Hii @FireFistisDead
I would like to work on this issue.
Please assign this issue to me if possible.