ValGuard.ai logoValGuard.ai

ValGuard + Vapi: policy on assistant LLM calls

Point the Vapi assistant custom LLM URL at ValGuard so spoken-agent prompts and completions pass the same rule packs as chat.

Last verified:

Vapi assistants call a custom LLM URL for reasoning while Vapi handles telephony and speech. Point that URL at ValGuard so voice workflows reuse the same validators as chat — with realistic latency expectations.

When you need this

A phone agent confirms a medical appointment the clinic never approved. The transcript sounds polite. ValGuard blocks approval language until structured status allows it — but block rules buffer the LLM reply, which adds dead air. Budget UX accordingly.

Setup

  1. Create a ValGuard agent for the voice assistant role.
  2. In Vapi, set the custom LLM endpoint to https://api.valguard.ai/v1/chat/completions (verify exact Vapi field names in their docs).
  3. Send ValGuard API key and X-VG-Agent per assistant.
  4. Start in shadow mode on test numbers.

Code

Treat Vapi as an OpenAI-compatible client:

# Equivalent to what Vapi sends when configured with ValGuard as custom LLM
curl -s https://api.valguard.ai/v1/chat/completions \
  -H "Authorization: Bearer $VG_API_KEY" \
  -H "X-VG-Agent: vapi-support" \
  -H "Content-Type: application/json" \
  -d '{"model":"openai/gpt-4o-mini","messages":[{"role":"user","content":"Caller asks to cancel policy 8821"}]}'

Honest limits

  • Speech-to-text and text-to-speech quality are outside ValGuard.
  • Real-time voice needs block rules only where buffering is acceptable, or use warn/log plus human review queues.
  • Telephony compliance (recording consent, PCI over voice) is still your program.

Related

Next step

Quickstart and shadow mode.