Bug: /v1/conversations returns 500 — response exceeds Cloud Run 32MB size limit

[beastoin]

GET /v1/conversations returns 500 with "Response size was too large" for users with photo-heavy or long conversations. 20 failures in 24h, concentrated on 2 users. Even with limit=50, fully hydrated responses exceed Cloud Run's 32MB response size limit.

Current Behavior

• The endpoint returns List[Conversation] with full model including all fields
• @with_photos decorator (database/helpers.py:199-256) loads ALL photos with full base64 content for EACH conversation in the list — 50 separate Firestore sub-collection reads
• @prepare_for_read adds copy.deepcopy overhead
• Estimated response size: 50 conversations × ~1.2MB each = ~60MB (1.9x over the 32MB limit)
• Cloud Run returns "Response size was too large" → 500

Expected Behavior

List endpoint should return lightweight conversation objects. Full details (photos, full transcripts) should be loaded on demand per conversation.

Affected Areas

File	Line	Description
backend/routers/conversations.py	119	Returns full List[Conversation] with no field filtering
backend/database/conversations.py	177-178	@with_photos + @prepare_for_read decorators on get_conversations()
backend/database/helpers.py	199-256	@with_photos loads all photos with full base64 per conversation
backend/database/conversations.py	141-147	get_conversation_photos() streams full photo docs
backend/models/conversation.py	66-72, 290-328	ConversationPhoto includes base64, full Conversation model

Size Breakdown

Component	Per Conversation	50 Conversations
Metadata	2-5 KB	100-250 KB
Transcript segments	~95 KB	~4.7 MB
Photos (base64)	~1 MB (10 photos)	~50 MB
Other fields	~40 KB	~2 MB
Total	~1.2 MB	~57 MB

Solution

Recommended: Create a ConversationLite response model for list endpoints that excludes:

• Photos base64 content (keep photo IDs/metadata only)
• Full transcript segment text (keep segment count/duration only)
• Large plugins_results content

Note: test commits for get_conversations_lite() exist (be1489a23, c391d21d8, a80fbdebf) but the function is not in the current codebase — appears incomplete or reverted.

Files to Modify

• backend/models/conversation.py — add ConversationLite model
• backend/routers/conversations.py — use lite model for list endpoint
• backend/database/conversations.py — add get_conversations_lite() without @with_photos

Impact

Users with many photo-heavy conversations are locked out of their conversation list. Workaround: reduce limit parameter to 10-15.

Related

• GET /v1/conversations extreme latency (30-90s) from N+1 queries and full hydration #4904 — N+1 query extreme latency (same root cause: full hydration), PR fix(conversations): eliminate N+1 photo queries from list endpoints #5271 addresses with get_conversations_lite
• Optimize conversation prefetch in transcribe.py (memory pressure from 0.6s re-fetch cycle) #5373 — conversation prefetch memory pressure (related hydration issue)
• fix(backend): reduce Firestore hot-path costs ~15-20% (#5377) #5378 — Firestore hot-path cost reduction PR

---

by AI for @beastoin