Vapi

A phone assistant that remembers the caller. Recall loads their facts before the greeting, saves what they ask it to remember, and carries corrections into the next call. “Actually, call me Sam” replaces the old name and keeps the history.

recall-vapi connects Recall to Vapi through authenticated webhooks. It uses the same starter predicates as recall-livekit. Vapi handles the voice conversation; your service owns the memory. No separate extraction model runs.

Start with an inbound call

Use a Vapi phone number that you can call from your own phone. You need Python 3.10 or newer, Vapi credits and provider access, and a public HTTPS address for the webhook. The local test helper can create a temporary address using an installed cloudflared.

Install the adapter with its FastAPI webhook router from PyPI:

pip install "recall-vapi[fastapi]"

For the complete phone assistant and live test helper, use the runnable example from the repository. Its requirements install the adapter from the checkout:

git clone https://github.com/Polign/polign.git
cd polign/examples/vapi
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python live_test.py init --phone-number-id YOUR_VAPI_PHONE_NUMBER_ID
python live_test.py serve --tunnel

Use the phone number’s resource ID from Vapi, not the dialable number. The helper creates a private .env with two independent secrets, starts the webhook server, and prints its public URL. Keep this terminal running.

Connect your Vapi number

In a second terminal, activate the same virtual environment and run these commands from examples/vapi:

python live_test.py check
python live_test.py configure

check verifies HTTPS, authentication, and synthetic memory calls. configure prints the dashboard steps; it does not change your Vapi account.

1. In Vapi’s Integrations → Server Configuration, create a Bearer Token credential using VAPI_WEBHOOK_TOKEN from your local .env. Use the Authorization header with the Bearer prefix.
2. Set your phone number’s Server URL to the printed HTTPS origin followed by /vapi/webhook, and select that credential.
3. Leave the inbound assistant unassigned. Vapi will request an assistant from your service, which returns one with the caller’s memory and tools.

The webhook service needs no Vapi private API key. See Vapi’s documentation for server URLs, server authentication, and assistant-request events.

Verify memory across calls

1. Call your Vapi number. Say “My name is Alex,” wait for a reply, then say “Actually, call me Sam.” Wait for confirmation and hang up.
2. Call again from the same phone. Listen for Sam in the greeting or ask “What’s my name?” Then hang up.

python live_test.py report
python live_test.py report --first-call CALL_1_ID --second-call CALL_2_ID

The report checks the saved correction, the next call’s loaded memory, matching caller identities, completed calls, and tool errors. Confirm the spoken reply by listening. Synthetic checks are separate from live phone evidence.

Press Ctrl-C in the first terminal when finished. The temporary URL stops working; each tunnel launch gets a new one, so update Vapi before testing again. Saved memory and test evidence remain in the local data/ directory.

What happens during a call

• assistant-request resolves a stable caller identity and loads up to 20 current facts before the first reply.
• remember validates facts against your predicate registry and returns saved corrections to the model. recall retrieves more facts on demand.
• The example enables forget to withdraw a current fact. History remains; this is not permanent deletion.
• A persistent retry ledger prevents duplicate delivery of the same tool call from applying an old correction again. End reports close the call’s identity binding.

Initial memory is loaded once. New tool results guide the ongoing conversation; the adapter does not rewrite the live system prompt or extract facts from transcripts.

Identity and deployment

The example derives a caller key from your tenant and the inbound phone number using CALLER_IDENTITY_KEY. Keep that key stable. Caller ID is suitable for a preferences demo; use an authenticated customer lookup for account-specific memory. Unresolved callers receive an assistant without memory tools.

For ongoing use, deploy the included Docker example behind permanent HTTPS and retain the whole data/ directory. It holds both the memory store and the SQLite call and retry ledger. You can point memory at a shared Polign server with POLIGN_URL and POLIGN_API_KEY; multiple webhook hosts also need shared transactional call state.

Keep webhook credentials on the backend. The example’s transient assistant contains them and must not be sent to a browser. The adapter guide covers outbound and web calls, including binding a call to a customer before tools run.

Source and example

recall-vapi is open source under Apache License 2.0. The runnable example includes FastAPI, Docker, local verification, and the live phone test helper. See the adapter README for custom predicates, timeouts, retries, and persistence limits.