litellm - 💡(How to fix) Fix store_prompts_in_spend_logs: false still stores full embedding response vectors in SpendLogs [1 participants]

Bug Description

With store_prompts_in_spend_logs: false in general_settings, embedding responses (full float vectors) are still written to the response column in LiteLLM_SpendLogs. This causes significant DB bloat — each embedding request stores the complete vector array (e.g. 1024 floats for bge-m3).

Expected Behavior

With store_prompts_in_spend_logs: false, the response column should remain {} for all call types, including embedding and aembedding.

Actual Behavior

The response column contains the full embedding response including all float vectors, despite the flag being set to false.

Steps to Reproduce

1. Set store_prompts_in_spend_logs: false in general_settings
2. Send an embedding request through the proxy
3. Query the DB:

   SELECT call_type, length(response::text)
   FROM "LiteLLM_SpendLogs"
   WHERE call_type IN ('embedding', 'aembedding')
   ORDER BY "startTime" DESC
   LIMIT 5;

4. Observe that response contains the full vector data instead of {}

Environment

• LiteLLM version: 1.82.3
• Deployment: Docker (LiteLLM Proxy)
• DB: PostgreSQL
• Embedding model: BAAI/bge-m3 (1024 dimensions) via vLLM backend

Config (relevant section)

general_settings:
  store_prompts_in_spend_logs: false
  maximum_spend_logs_retention_period: "7d"

litellm_settings:
  callbacks: custom_callbacks.proxy_handler_instance
  cache: true
  cache_params:
    type: redis
    supported_call_types:
      - completion
      - acompletion

Impact

Each bge-m3 embedding response is ~15-20 KB of JSON. At high request volume this bloats the DB quickly. Chat completions may or may not be affected — not yet verified.

Workaround

Periodic cleanup:

UPDATE "LiteLLM_SpendLogs"
SET response = '{}'
WHERE call_type IN ('embedding', 'aembedding');

Related Issues

• #15641 — inverse bug: store_prompts_in_spend_logs: true does NOT store messages/response
• #23636 — data stored in proxy_server_request instead of messages