Balance System Only Counting Tokens from Title Generation

[twinity1]

Problem Description

I've set up LibreChat with the balance system enabled, but I've noticed it only counts tokens from conversation titling and not from regular conversation messages. The balance feature seems to be partially working, but not tracking actual conversation token usage.

Configuration

My balance configuration in librechat.yaml:

balance:
  enabled: true
  startBalance: 200000
  autoRefillEnabled: true
  refillIntervalValue: 1
  refillIntervalUnit: "hours"
  refillAmount: 200000

I'm using LiteLLM as a custom provider:

custom:
  - name: "LiteLLM"
    apiKey: "xxxx"
    baseURL: "https://litellm.xxxx.com/v1"
    models:
      default: ["eu.anthropic.claude-3-7-sonnet-20250219-v1:0", "gemini-2.0-flash-lite", "gemini-2.5-pro-think-low", "vertex_ai/gemini-2.5-pro", "vertex_ai/gemini-2.5-flash"]
    fetch: true
    titleConvo: true
    titleModel: "gemini-2.0-flash-lite"
    titleMessageRole: "user"
    forcePrompt: false
    modelDisplayLabel: "LiteLLM"

Transactions

Looking at my transactions collection in MongoDB, I only see entries related to title generation:

_id	context	conversationId	createdAt	model	rate	rawAmount	tokenType	tokenValue
68713f4a5d6f269c6e5483db	title	a7b291f6-3fed-4d32-9af1-a301381ec1ff	2025-07-11T16:43:54.827Z	gemini-2.0-flash-lite	0.3	-25	completion	-7.5
68713f4a5d6f269c6e5483d5	title	a7b291f6-3fed-4d32-9af1-a301381ec1ff	2025-07-11T16:43:54.807Z	gemini-2.0-flash-lite	0.075	-496	prompt	-37.199999999999996

Evidence of Token Usage

However, looking at the messages collection, it's clear that substantial tokens are being used but not tracked:

_id	messageId	model	tokenCount	updatedAt
68713f7c949297e08e4b7b56	ec8184cb-0eef-4b19-a8ba-bc97821ec96d_	agent_KMY-m92d2m6PnyUqdyaWj	5708	2025-07-11T16:44:44.474Z
68713f4a949297e08e4b7b55	ec8184cb-0eef-4b19-a8ba-bc97821ec96d	agent_KMY-m92d2m6PnyUqdyaWj	5759	2025-07-11T16:43:54.010Z
68713f33949297e08e4b7b53	8c14f422-73ba-4d7d-83f8-17ae2517be3f	null	11	2025-07-11T16:44:44.502Z

As you can see, there are messages with substantial token counts (5708 and 5759) that should be deducted from the balance, but they aren't appearing in the transactions collection.

The logs show token tracking for title generation:

2025-07-11T16:43:54.803Z debug: [spendTokens] conversationId: a7b291f6-3fed-4d32-9af1-a301381ec1ff | Context: title | Token usage:
{
  promptTokens: 496,
  completionTokens: 25,
}

But no token tracking appears for the actual conversation messages, even though they're clearly being processed.

Interestingly, I found this in the logs which suggests the balance system is checking tokens but not recording them:

2025-07-11T16:44:22.563Z debug: [Balance.check] Initial state
{
  user: "67d448dd765a6dba0b739302",
  model: "gemini-2.5-pro",
  endpoint: "agents",
  valueKey: undefined,
  tokenType: "prompt",
  amount: 14,
  balance: 198819.275,
  multiplier: 1.25,
  endpointTokenConfig: false,
}
2025-07-11T16:44:22.564Z debug: [Balance.check] Token cost
{
  tokenCost: 17.5,
}

This shows that the balance system is calculating token costs (17.5) for the prompt, but these transactions aren't being recorded in the database. The endpointTokenConfig: false might be the key issue here?

Any help would be greatly appreciated! Thank you!

[danny-avila]

Not seeing this on my end, checking mongodb manually and also using a mongodb mcp server:

[danny-avila]

double checking with agents

[danny-avila]

make sure you are on the latest version as I just confirmed there are no issues with agents

[twinity1]

It appears there's an issue specifically with the LiteLLM. I've tested with OpenRouter and the balance tracking works correctly there, so this seems to be a LiteLLM-specific problem.. :X

[danny-avila]

I don’t have litellm setup to test with but double check if it’s returning any usage metadata. It should still do an estimate when there’s one missing.

[twinity1]

Thank you for pointing me to the right direction. LiteLLM indeed doesn't return the usage metadata by default. To get this working, I needed to enhance every request with:

"stream_options": {
  "include_usage": true
}

Adding parameters like:

addParams:
  stream_options:
    include_usage: true

didn't work, so I created a simple nginx proxy that intercepts the requests.