Token
I would like to understand the reason behind the excessive token consumption. The Voiceflow dashboard doesn’t show what caused this, and I believe a large number of tokens were used inappropriately. The day started with 7,000 tokens consumed, and within just 3 hours, the number jumped to 21,000, without any significant usage on my part. I’d appreciate it if you could investigate this issue and restore the tokens that were consumed unfairly.
1 Reply
Hi there — I know sudden jumps in usage can be alarming. Here’s what you can do to pinpoint (and reduce) the unexpected token / credit spend:
1. Open your project → Analytics → Credits usage chart
• Switch the date filter to the 3-hour window you noticed.
• The chart breaks costs down by feature (chat, voice, etc.) and by the LLM model that ran. This will tell you whether the spike came from extra chat turns, a phone call, or a large LLM prompt/response.
2. Compare with the Transcripts tab
• Each conversation is time-stamped, so you can match any large bursts of messages to the same period. If you see the agent looping, or an unusually long answer, that typically explains big token counts.
3. Common causes we see
• A logic loop (e.g. a prompt that keeps re-calling the model).
• Uploading a large Knowledge Base source and immediately querying it many times.
• Running bulk tests or hitting the Dialog Manager API from an external script.
4. Optimisation tips
• Use cheaper models (e.g. GPT-4o-Mini) where possible.
• Add guard-rail conditions (e.g. max turns) to stop accidental loops.
• Keep prompts concise and strip unnecessary variables before sending them to the LLM.
If the Analytics breakdown still doesn’t explain the spike, grab the exact project & time range and we can look deeper on our side. Hope this helps you track down the out-of-scope usage!
📚 Docs for reference: https://docs.voiceflow.com/docs/monitoring-your-credit-usage
(see “The credits usage chart” section)
• Switch the date filter to the 3-hour window you noticed.
• The chart breaks costs down by feature (chat, voice, etc.) and by the LLM model that ran. This will tell you whether the spike came from extra chat turns, a phone call, or a large LLM prompt/response.
2. Compare with the Transcripts tab
• Each conversation is time-stamped, so you can match any large bursts of messages to the same period. If you see the agent looping, or an unusually long answer, that typically explains big token counts.
3. Common causes we see
• A logic loop (e.g. a prompt that keeps re-calling the model).
• Uploading a large Knowledge Base source and immediately querying it many times.
• Running bulk tests or hitting the Dialog Manager API from an external script.
4. Optimisation tips
• Use cheaper models (e.g. GPT-4o-Mini) where possible.
• Add guard-rail conditions (e.g. max turns) to stop accidental loops.
• Keep prompts concise and strip unnecessary variables before sending them to the LLM.
If the Analytics breakdown still doesn’t explain the spike, grab the exact project & time range and we can look deeper on our side. Hope this helps you track down the out-of-scope usage!
📚 Docs for reference: https://docs.voiceflow.com/docs/monitoring-your-credit-usage
(see “The credits usage chart” section)