Files
avc-phone-ai/bot.py
tocmo0nlord a521dc168e Fix GPU OOM: share one Whisper model across calls (was leaking per call)
Calls were dropping right after answer with "CUDA failed with error out of
memory". Cause: each call constructed a new HintedWhisperSTTService -> new
ctranslate2 WhisperModel on the GPU, and that VRAM was never released when the
call ended. Over ~13 calls the python process grew to 9.7GB; with the pinned LLM
(6GB) the 16GB GPU filled (14 MiB free) and Whisper load failed on every call.

Fix: cache one WhisperModel per (model,device,compute) in _WHISPER_MODEL_CACHE
and reuse it across all calls; bake the fixed hotwords into the shared model's
transcribe() once (drops the racy per-call monkey-patch). VRAM now constant
(~6GB LLM + ~1.5GB Whisper). Verified: two instances share one model object;
GPU back to 6.0/16GB used after restart. Documented the VRAM budget.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-27 22:07:59 +00:00

743 lines
38 KiB
Python

#!/usr/bin/env python3
"""AVC optometry phone agent — the Pipecat pipeline for a single inbound call.
Same VAD -> STT -> LLM -> TTS loop as pipecat-run/bot.py, but the ends are swapped
for telephony: audio arrives/leaves as 8 kHz mu-law over a Twilio Media Stream
(WebSocket), decoded by TwilioFrameSerializer. STT runs on the GPU; the LLM is the
local `activeblue-avc` fine-tune via Ollama; TTS is local Kokoro.
This module just builds + runs the pipeline for one connected call. server.py owns
the FastAPI/TwiML/WebSocket side and calls run_call() once per call.
"""
import asyncio
import os
import re
import time
from loguru import logger
from pipecat.adapters.schemas.function_schema import FunctionSchema
from pipecat.adapters.schemas.tools_schema import ToolsSchema
from pipecat.audio.vad.silero import SileroVADAnalyzer
from pipecat.audio.vad.vad_analyzer import VADParams
from pipecat.frames.frames import (
BotStartedSpeakingFrame,
BotStoppedSpeakingFrame,
EndFrame,
EndTaskFrame,
Frame,
InputAudioRawFrame,
LLMFullResponseEndFrame,
LLMTextFrame,
TTSSpeakFrame,
UserStartedSpeakingFrame,
)
from pipecat.pipeline.pipeline import Pipeline
from pipecat.pipeline.runner import PipelineRunner
from pipecat.pipeline.task import PipelineParams, PipelineTask
from pipecat.processors.aggregators.llm_context import LLMContext
from pipecat.processors.aggregators.llm_response_universal import LLMContextAggregatorPair
from pipecat.processors.audio.audio_buffer_processor import AudioBufferProcessor
from pipecat.processors.audio.vad_processor import VADProcessor
from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
from pipecat.serializers.twilio import TwilioFrameSerializer
from pipecat.services.anthropic.llm import AnthropicLLMService
from pipecat.services.kokoro.tts import KokoroTTSService
from pipecat.services.ollama.llm import OLLamaLLMService
from pipecat.services.whisper.stt import WhisperSTTService
from pipecat.transports.websocket.fastapi import (
FastAPIWebsocketParams,
FastAPIWebsocketTransport,
)
from practice import practice_summary
# ── Config (env-overridable) ─────────────────────────────────────────────────
HERE = os.path.dirname(os.path.abspath(__file__))
# Reuse the Kokoro model files already downloaded by the pipecat-run project.
MODEL_DIR = os.environ.get("KOKORO_MODEL_DIR", "/home/tocmo0nlord/pipecat-run/models")
OLLAMA_MODEL = os.environ.get("OLLAMA_MODEL", "activeblue-avc:latest")
OLLAMA_URL = os.environ.get("OLLAMA_URL", "http://127.0.0.1:11434/v1")
# Swappable LLM provider: "ollama" (local) or "anthropic" (Claude API). Same universal
# LLMContext drives both — only the service construction differs (see build_llm_service).
LLM_PROVIDER = os.environ.get("LLM_PROVIDER", "ollama").lower()
ANTHROPIC_API_KEY = os.environ.get("ANTHROPIC_API_KEY", "")
# Defaults to the most capable model. For low-latency PHONE voice, set ANTHROPIC_MODEL to
# claude-haiku-4-5 (fastest) or claude-sonnet-4-6 (balance) — see notes in build_llm_service.
ANTHROPIC_MODEL = os.environ.get("ANTHROPIC_MODEL", "claude-opus-4-8")
# In-call function-calling: AUTO by provider — ON for Claude (reliable tool calls → real-time
# Odoo booking), OFF for local Ollama (llama3.1:8b over-calls / leaks JSON). An explicit
# ENABLE_TOOLS env overrides the auto choice either way.
_enable_tools_env = os.environ.get("ENABLE_TOOLS")
ENABLE_TOOLS = (
_enable_tools_env.lower() in ("1", "true", "yes")
if _enable_tools_env is not None
else (LLM_PROVIDER == "anthropic")
)
LLM_TEMPERATURE = float(os.environ.get("LLM_TEMPERATURE", "0.3"))
LLM_MAX_TOKENS = int(os.environ.get("LLM_MAX_TOKENS", "160"))
KOKORO_VOICE = os.environ.get("KOKORO_VOICE", "af_heart")
WHISPER_MODEL = os.environ.get("WHISPER_MODEL", "medium") # tiny|base|small|medium
WHISPER_DEVICE = os.environ.get("WHISPER_DEVICE", "cuda") # cuda for the 5080
WHISPER_COMPUTE = os.environ.get("WHISPER_COMPUTE", "float16")
# Bias transcription toward our domain vocabulary (office cities + optometry terms) so
# 8 kHz telephony audio doesn't turn "Hialeah" into "high allele" or "eye exam" into "hire".
WHISPER_HOTWORDS = os.environ.get(
"WHISPER_HOTWORDS",
"Advanced Vision Care, eye exam, annual exam, appointment, optometry, contact lens, "
"Hialeah, Kendall, Tamarac, Pembroke Pines, Lauderdale Lakes, Miami Gardens, Boca Raton",
)
# Twilio sends 8 kHz mu-law on the wire, but faster-whisper assumes any numpy array is
# 16 kHz — so we run the PIPELINE at 16 kHz and let TwilioFrameSerializer resample to/from
# the 8 kHz wire. Running the pipeline at 8 kHz makes Whisper hear 2x-speed audio and
# transcribe nothing. (Silero VAD + Kokoro are happy at 16 kHz too.)
WIRE_SAMPLE_RATE = 8000 # Twilio mu-law on the wire (serializer handles this)
PIPELINE_SAMPLE_RATE = 16000 # internal rate Whisper/VAD actually need
# VAD tuning. Defaults (confidence 0.7 / min_volume 0.6) are desktop-mic values that can
# miss short/quiet 8 kHz telephony utterances like "yes" — loosen them for the phone.
# VAD is kept sensitive so a quick/quiet "yes" isn't missed (a caller had to repeat it). This
# is safe because HalfDuplexGate gates out the agent's echo while it speaks, so sensitive VAD
# doesn't cause echo false-triggers — it only listens hard during the caller's own turn.
VAD_CONFIDENCE = float(os.environ.get("VAD_CONFIDENCE", "0.5"))
VAD_MIN_VOLUME = float(os.environ.get("VAD_MIN_VOLUME", "0.15"))
VAD_START_SECS = float(os.environ.get("VAD_START_SECS", "0.1"))
VAD_STOP_SECS = float(os.environ.get("VAD_STOP_SECS", "0.5"))
# Half-duplex: ignore inbound audio while the agent is speaking (+ this tail in seconds)
# so the agent's own voice echoing back the phone line can't trigger a false barge-in that
# cancels its reply (= caller hears silence). Set HALF_DUPLEX=false to allow barge-in.
HALF_DUPLEX = os.environ.get("HALF_DUPLEX", "true").lower() not in ("false", "0", "no")
ECHO_TAIL_SECS = float(os.environ.get("ECHO_TAIL_SECS", "0.25"))
# Silence watchdog: if the caller goes quiet after the agent speaks, re-prompt instead of
# dead-waiting (a missed/clipped answer otherwise hangs the call). After MAX re-prompts with
# no response, close gracefully. SILENCE_REPROMPT_SECS must be > HANGUP_DELAY_SECS so a real
# goodbye hangs up before the watchdog fires.
SILENCE_WATCHDOG = os.environ.get("SILENCE_WATCHDOG", "true").lower() not in ("false", "0", "no")
SILENCE_REPROMPT_SECS = float(os.environ.get("SILENCE_REPROMPT_SECS", "7.0"))
MAX_REPROMPTS = int(os.environ.get("MAX_REPROMPTS", "2"))
# Record each call to a stereo WAV (caller = left, agent = right) for review/debugging.
RECORD_CALLS = os.environ.get("RECORD_CALLS", "true").lower() not in ("false", "0", "no")
RECORDINGS_DIR = os.environ.get("RECORDINGS_DIR", os.path.join(HERE, "recordings"))
# Agent persona name — purely for warmth; change/remove freely.
AGENT_NAME = os.environ.get("AGENT_NAME", "Sofia")
# How the name should be SPOKEN. Kokoro reads all-caps "AVA" as letters ("A-V-A"), so we
# respell it as a word for TTS only (logs/Odoo keep AGENT_NAME). Override to taste, e.g.
# AGENT_NAME_SPOKEN=Eva for an "EE-vuh" sound.
AGENT_NAME_SPOKEN = os.environ.get("AGENT_NAME_SPOKEN", "Ava")
# Grace period after the agent finishes the goodbye before we drop the carrier leg, so
# the caller isn't cut off mid-word. The hang-up itself (EndTaskFrame -> auto_hang_up)
# is unchanged — this only delays it.
HANGUP_DELAY_SECS = float(os.environ.get("HANGUP_DELAY_SECS", "4.0"))
SYSTEM_PROMPT = (
f"You are {AGENT_NAME}, a warm, friendly receptionist for Advanced Vision Care, an "
"optometry practice with eight offices in South Florida. You are on a real phone call, so "
"talk like a helpful human being: natural, relaxed, and genuinely conversational. Keep every "
"reply to ONE short sentence — two at the very most, never a paragraph. Speak in English. Say "
"numbers, dates, and times as words a person would say.\n\n"
"Your job is to answer questions and take appointment requests. Be warm but DIRECT and "
"efficient: when the caller greets you, get to the point and lead the call by asking "
"questions. Never re-ask for something they already told you, and keep each turn to one "
"short question or statement. Work through the call in THIS order:\n"
" 1. REASON FIRST — find out what they are calling about (the reason for the visit, or "
"their question). If it is only a question, answer it.\n"
" 2. LOCATION — ask which city or area is most convenient, then confirm the matching "
"office (see the office rule below).\n"
" 3. CALLER INFO — get their FULL name (first and last; if they give only a first name, "
"ask their last name). From this point on, address the caller by their name. Then ask their "
"insurance (log only — see below) and their preferred day and time (in their own words — "
"see the date rule below).\n"
" 4. CONFIRM PHONE (no yes needed) — near the end, STATE the callback number back in one "
"line and invite a CORRECTION ONLY, exactly like: 'I have your number as <the number spelled "
"out below>; if that's not the best number, just let me know.' Do NOT ask a yes/no question, "
"do NOT ask permission, and do NOT wait for them to say 'yes' — flow straight into the wrap-up. "
"Only act on the phone number if they give you a different one. Don't bring it up before this.\n"
" 5. WRAP UP — recap the booking as a REQUEST in one warm sentence (for example, 'I've "
"noted your request to come in tomorrow afternoon at our Kendall office'), make clear a "
"staff member will call back to CONFIRM it, then ASK IF THERE IS ANYTHING ELSE you can help "
"them with. Only once they are all set, give the closing below.\n"
"KEEP MOMENTUM: until the booking is complete, ALWAYS end your turn with the next question "
"you still need answered. Never reply with only an acknowledgment and then stop — for "
"example, after noting insurance, in the SAME turn go straight on to ask the preferred day "
"and time. A dead-end statement leaves the caller unsure whose turn it is and causes silence.\n\n"
"Stay truthful and within your limits:\n"
"- Use ONLY the facts below for addresses, phone numbers, insurance, and services. Never "
"make any of these up.\n"
"- OFFICE SELECTION: ask what city or area is most convenient. When the caller names a place "
"that matches one of our offices (for example they say 'Kendall' and we have a Kendall "
"office), simply confirm THAT office and move on — do NOT offer, compare, or mention any "
"other office or city, and do NOT ask them to choose between offices. Only if their area "
"matches no office should you name the single nearest one. List offices only if asked.\n"
"- INSURANCE — log only, never promise, never guess: ask open-endedly what insurance they "
"have (for example, 'What insurance do you have?'). Do NOT read out or suggest plan names "
"from the list — let the caller tell you. Capture ONLY what the caller actually says; NEVER "
"fill in, complete, or guess their plan, and never put words in their mouth. If you don't "
"clearly hear the plan, ask them to repeat it. Do not promise, confirm, or deny coverage or "
"any treatment based on their insurance, even if the plan is one we list, and NEVER say 'we "
"accept' or 'we take' a plan — just note it and say our staff will verify their coverage when "
"they call back.\n"
"- DATES — just take down the day and time the caller asks for in their OWN words (e.g. "
"'next Monday', 'the fifth'). Do NOT work out, state, or correct the calendar date, and NEVER "
"argue about what today's date is. Tell them staff will confirm the exact date and time on the "
"callback.\n"
"- You CANNOT see availability or book/confirm anything. Never say a slot is open or "
"available, never offer to 'check availability', and NEVER tell the caller their appointment "
"is 'booked', 'scheduled', 'set', or 'confirmed' — not even in the recap. It is always a "
"REQUEST: say you've NOTED it and a staff member will call back to confirm the date and time.\n"
"- Hours are not published — say they vary by office and staff will confirm; never give "
"specific hours.\n"
"- You don't give medical advice and can't transfer calls. If the caller mentions an eye "
"problem, just note it as the reason and say a staff member or doctor will follow up.\n"
"- If you're not sure you heard something, simply ask them to repeat it.\n"
"- CLOSING — the word 'Goodbye' ENDS the call, so guard it carefully. You MUST first ask "
"'Is there anything else I can help you with?' and hear the caller say they need nothing "
"more. NEVER say 'Goodbye' in the same turn as confirming details or the phone number, and "
"never before that anything-else question. Once they confirm they're all set, give a brief "
"warm closing ending with 'Goodbye'.\n\n"
"PRACTICE FACTS:\n" + practice_summary()
)
def _build_tools() -> ToolsSchema:
# Only the booking action is a tool. Practice facts already live in the system prompt,
# so no get_practice_info tool (avoids needless calls/latency). callback_number is NOT
# required — we have the caller-ID and inject it in the handler.
return ToolsSchema(
standard_tools=[
FunctionSchema(
name="record_appointment_request",
description=(
"Record the caller's appointment request once you have their name and at "
"least the office/city and reason. Call this when the caller wants to book "
"a visit; staff will call back to confirm the exact time."
),
properties={
"patient_name": {"type": "string", "description": "Caller's full name"},
"location": {"type": "string", "description": "Which office/city the caller wants, e.g. Hialeah, Kendall, Tamarac"},
"reason": {"type": "string", "description": "Reason for the visit, e.g. annual exam, broken glasses, eye pain"},
"preferred_time": {"type": "string", "description": "Preferred day/time in the caller's words, if given"},
},
required=["patient_name"],
),
]
)
class EndCallProcessor(FrameProcessor):
"""Lets the agent hang up AND guarantees the callback number is confirmed once.
Sits between the LLM and the TTS: it sees reply text (LLMTextFrame, downstream) and the
upstream BotStoppedSpeakingFrame. On a closing ('goodbye'/'adiós') it waits for TTS to
finish, pauses HANGUP_DELAY_SECS so the caller isn't clipped, then pushes EndTaskFrame
(TwilioFrameSerializer auto_hang_up drops the call).
Deterministic phone confirmation: the prompt asks the agent to read the callback number
back, but the 8B skips it ~half the time. So if a closing is reached and the agent never
spoke the number this call (`phone_marker` not seen in its replies), we suppress the
hang-up and inject a scripted confirmation turn first — guaranteeing it happens exactly
once (the agent's own readback satisfies the gate, so no double-ask in the common case)."""
_CLOSINGS = ("goodbye", "good-bye", "good bye", "adiós", "adios", "hasta luego")
# Only force phone confirmation when a booking was actually underway (not info-only calls).
_BOOKING_KWS = ("appointment", "schedule", "book", "insurance", "what day", "what time",
"come in", "preferred")
def __init__(self, phone_confirm_line: str | None = None, phone_marker: str | None = None):
super().__init__()
self._buf = ""
self._should_end = False
self._end_task = None
self._phone_confirm_line = phone_confirm_line
self._phone_marker = (phone_marker or "").lower()
# Nothing to confirm (no caller ID) → treat as already handled.
self._phone_confirmed = not phone_confirm_line
self._assistant_seen = ""
self._pending_phone_inject = False
@classmethod
def _is_closing(cls, text: str) -> bool:
t = (text or "").lower()
return any(c in t for c in cls._CLOSINGS)
async def _hang_up_after_delay(self):
await asyncio.sleep(HANGUP_DELAY_SECS)
logger.info(f"{AGENT_NAME} ending task / hanging up")
try:
await self.push_frame(EndTaskFrame(), FrameDirection.UPSTREAM)
except Exception:
logger.exception("EndTaskFrame push failed (pipeline already ending?)")
async def process_frame(self, frame: Frame, direction: FrameDirection):
await super().process_frame(frame, direction)
if isinstance(frame, LLMTextFrame):
self._buf += frame.text
self._assistant_seen += frame.text.lower()
if self._phone_marker and self._phone_marker in self._assistant_seen:
self._phone_confirmed = True # the agent read the number back itself
elif isinstance(frame, LLMFullResponseEndFrame):
if self._is_closing(self._buf):
booking = any(k in self._assistant_seen for k in self._BOOKING_KWS)
if self._phone_confirmed or not booking:
self._should_end = True
logger.info(f"{AGENT_NAME} signalled closing -- will hang up "
f"{HANGUP_DELAY_SECS:.0f}s after she finishes speaking")
else:
# Booking call closing without the number confirmed — do it deterministically.
self._pending_phone_inject = True
logger.info(f"{AGENT_NAME} reached closing w/o phone confirmation -- injecting it")
self._buf = ""
elif isinstance(frame, BotStoppedSpeakingFrame):
if self._pending_phone_inject:
self._pending_phone_inject = False
self._phone_confirmed = True
await self.push_frame(TTSSpeakFrame(self._phone_confirm_line), FrameDirection.DOWNSTREAM)
elif self._should_end:
self._should_end = False
# Schedule the teardown so we don't block the pipeline during the grace pause.
if self._end_task is None:
self._end_task = asyncio.create_task(self._hang_up_after_delay())
await self.push_frame(frame, direction)
class AudioHeartbeat(FrameProcessor):
"""Diagnostic: logs how many inbound audio frames arrive every ~5s. If this keeps
ticking but VAD never fires, the issue is VAD/threshold; if it drops to 0 after a
turn, inbound audio stalled at the transport. Cheap, leave it on while stabilizing."""
def __init__(self):
super().__init__()
self._n = 0
self._t = time.time()
async def process_frame(self, frame: Frame, direction: FrameDirection):
await super().process_frame(frame, direction)
if isinstance(frame, InputAudioRawFrame):
self._n += 1
now = time.time()
if now - self._t >= 5:
logger.info(f"[audio-in] {self._n} frames in last {now - self._t:.0f}s")
self._n = 0
self._t = now
await self.push_frame(frame, direction)
class HalfDuplexGate(FrameProcessor):
"""Drops inbound audio while the agent is speaking (plus ECHO_TAIL_SECS after it stops).
In this pipecat build interruptions are VAD-driven and always on (PipelineParams has no
allow_interruptions). On a phone line the agent's own TTS echoes back and the VAD reads it
as the caller speaking → it broadcasts an interruption that cancels the agent mid-reply, so
the caller hears silence. Sitting BEFORE the VAD, this gate withholds inbound audio frames
while the bot is speaking, so its echo never reaches the VAD. Trade-off: the caller can't
barge in mid-utterance (fine for short receptionist replies). Bypass with HALF_DUPLEX=false."""
def __init__(self, tail_secs: float = 0.5):
super().__init__()
self._bot_speaking = False
self._reopen_at = 0.0
self._tail = tail_secs
async def process_frame(self, frame: Frame, direction: FrameDirection):
await super().process_frame(frame, direction)
if isinstance(frame, BotStartedSpeakingFrame):
self._bot_speaking = True
elif isinstance(frame, BotStoppedSpeakingFrame):
self._bot_speaking = False
self._reopen_at = time.time() + self._tail
# Withhold caller audio while the bot speaks (+ echo tail) so echo can't barge in.
if isinstance(frame, InputAudioRawFrame) and (self._bot_speaking or time.time() < self._reopen_at):
return
await self.push_frame(frame, direction)
class SilenceWatchdog(FrameProcessor):
"""Re-prompts on caller silence instead of dead-waiting. After the agent stops speaking it
arms a timer; if the caller hasn't started speaking within `silence_secs`, it injects a
re-prompt ("are you still there?"). After `max_prompts` unanswered re-prompts it speaks a
graceful closing and ends the call. Any caller speech (UserStartedSpeakingFrame) resets it;
the agent speaking cancels it. Place AFTER EndCallProcessor so its injected lines go to TTS
and `silence_secs` > HANGUP_DELAY_SECS so a real goodbye hangs up before it fires."""
def __init__(self, silence_secs: float, max_prompts: int, reprompt_line: str, closing_line: str):
super().__init__()
self._silence_secs = silence_secs
self._max_prompts = max_prompts
self._reprompt_line = reprompt_line
self._closing_line = closing_line
self._timer = None
self._prompts = 0
self._bot_speaking = False
self._ending = False
def _cancel(self):
if self._timer and not self._timer.done():
self._timer.cancel()
self._timer = None
def _arm(self):
self._cancel()
self._timer = asyncio.create_task(self._fire())
async def _fire(self):
try:
await asyncio.sleep(self._silence_secs)
except asyncio.CancelledError:
return
if self._bot_speaking or self._ending:
return
if self._prompts >= self._max_prompts:
self._ending = True
logger.info("SilenceWatchdog: still silent after re-prompts -- closing the call")
await self.push_frame(TTSSpeakFrame(self._closing_line), FrameDirection.DOWNSTREAM)
else:
self._prompts += 1
logger.info(f"SilenceWatchdog: caller silent -- re-prompt #{self._prompts}")
await self.push_frame(TTSSpeakFrame(self._reprompt_line), FrameDirection.DOWNSTREAM)
async def _end_soon(self):
await asyncio.sleep(HANGUP_DELAY_SECS)
logger.info("SilenceWatchdog: ending task after silent close")
try:
await self.push_frame(EndTaskFrame(), FrameDirection.UPSTREAM)
except Exception:
logger.exception("watchdog EndTaskFrame push failed")
async def process_frame(self, frame: Frame, direction: FrameDirection):
await super().process_frame(frame, direction)
if isinstance(frame, UserStartedSpeakingFrame):
self._prompts = 0 # caller engaged again — reset
self._cancel()
elif isinstance(frame, BotStartedSpeakingFrame):
self._bot_speaking = True
self._cancel()
elif isinstance(frame, BotStoppedSpeakingFrame):
self._bot_speaking = False
if self._ending:
asyncio.create_task(self._end_soon())
else:
self._arm() # start counting silence once the agent finishes
await self.push_frame(frame, direction)
# One shared WhisperModel per (model, device, compute) for the whole process. Loading a new
# model per call leaks GPU memory — ctranslate2 doesn't release VRAM when the call's service is
# dropped, so models accumulate and the GPU OOMs after a handful of calls. Sharing one keeps
# VRAM constant.
_WHISPER_MODEL_CACHE = {}
class HintedWhisperSTTService(WhisperSTTService):
"""WhisperSTTService that shares ONE WhisperModel across all calls (avoids the per-call
GPU-memory leak/OOM) and biases transcription toward domain vocabulary via faster-whisper
`hotwords`. Hotwords are a fixed domain list, so they're baked into the shared model's
transcribe() once at load — concurrency-safe (no per-call monkey-patch)."""
def __init__(self, *args, hotwords: str | None = None, **kwargs):
self._hotwords = hotwords # set BEFORE super().__init__ (it calls _load)
super().__init__(*args, **kwargs)
def _load(self):
key = (self._settings.model, self._device, self._compute_type)
model = _WHISPER_MODEL_CACHE.get(key)
if model is None:
super()._load() # base sets self._model
model = self._model
if self._hotwords: # bake hotwords in once (value, not self)
_real = model.transcribe
_hw = self._hotwords
def _patched(audio_arg, **kw):
kw.setdefault("hotwords", _hw)
return _real(audio_arg, **kw)
model.transcribe = _patched
_WHISPER_MODEL_CACHE[key] = model
logger.info(f"Loaded + cached shared Whisper model {key}")
else:
logger.info(f"Reusing shared Whisper model {key}")
self._model = model
# ── TTS number normalization ──────────────────────────────────────────────────
# Kokoro reads digit strings as cardinals with symbols spoken aloud, e.g. "983-4969"
# becomes "nine hundred eighty-three dash forty-nine sixty-nine". For a phone agent that
# reads back phone numbers, street numbers, and zips, that's unusable. We normalize the
# text right before synthesis (run_tts receives the full sentence) so phone numbers and
# long digit runs are spoken one digit at a time, regardless of what the model emitted.
_DIGIT_WORDS = {
"0": "zero", "1": "one", "2": "two", "3": "three", "4": "four",
"5": "five", "6": "six", "7": "seven", "8": "eight", "9": "nine",
}
_PHONE_RE = re.compile(r"(?:\+?1[\s.\-]?)?\(?\d{3}\)?[\s.\-]?\d{3}[\s.\-]?\d{4}")
_LONGNUM_RE = re.compile(r"\b\d{4,5}\b") # street numbers, zip codes
def _say_digits(s: str) -> str:
return " ".join(_DIGIT_WORDS[c] for c in s if c in _DIGIT_WORDS)
def _spoken_phone(number: str) -> str:
"""Phone number as grouped, digit-by-digit words: '+19735731671' ->
'nine seven three, five seven three, one six seven one'."""
d = re.sub(r"\D", "", number or "")
if len(d) == 11 and d[0] == "1": # drop US country code
d = d[1:]
if len(d) == 10: # group as area / prefix / line for natural cadence
return f"{_say_digits(d[:3])}, {_say_digits(d[3:6])}, {_say_digits(d[6:])}"
return _say_digits(d)
def _phone_to_words(m: re.Match) -> str:
return _spoken_phone(m.group(0))
def tts_normalize(text: str) -> str:
"""Make phone numbers, street numbers, and zips speak naturally (digit by digit), and
respell the all-caps agent name so it's said as a word, not letter-by-letter."""
if AGENT_NAME != AGENT_NAME_SPOKEN:
text = re.sub(rf"\b{re.escape(AGENT_NAME)}\b", AGENT_NAME_SPOKEN, text)
text = _PHONE_RE.sub(_phone_to_words, text)
text = _LONGNUM_RE.sub(lambda m: _say_digits(m.group(0)), text)
return text
class SpokenKokoroTTSService(KokoroTTSService):
"""KokoroTTSService that normalizes numbers to digit-by-digit speech before synthesis,
so phone numbers/addresses/zips are read naturally instead of as cardinals + 'dash'."""
async def run_tts(self, text: str, context_id: str):
async for frame in super().run_tts(tts_normalize(text), context_id):
yield frame
def build_llm_service():
"""Build the LLM service for the selected provider. The universal LLMContext +
aggregators work with either, so only this construction differs (true A/B swap)."""
if LLM_PROVIDER == "anthropic":
if not ANTHROPIC_API_KEY:
raise RuntimeError("LLM_PROVIDER=anthropic but ANTHROPIC_API_KEY is not set")
logger.info(f"LLM provider: anthropic ({ANTHROPIC_MODEL})")
# NOTE: Opus 4.8/4.7 reject temperature/top_p/top_k (HTTP 400), so we omit them —
# this keeps the default Opus model working. For low-latency phone voice, prefer
# claude-haiku-4-5 (fastest) or claude-sonnet-4-6 over Opus. enable_prompt_caching
# caches the system prompt + growing conversation prefix (helps multi-turn cost/latency).
return AnthropicLLMService(
api_key=ANTHROPIC_API_KEY,
settings=AnthropicLLMService.Settings(
model=ANTHROPIC_MODEL,
enable_prompt_caching=True,
max_tokens=LLM_MAX_TOKENS,
),
)
logger.info(f"LLM provider: ollama ({OLLAMA_MODEL})")
return OLLamaLLMService(
settings=OLLamaLLMService.Settings(
model=OLLAMA_MODEL,
temperature=LLM_TEMPERATURE,
max_tokens=LLM_MAX_TOKENS,
),
base_url=OLLAMA_URL,
)
async def run_agent(transport, caller_number=None, call_sid=None, do_capture=True):
"""Build + run the AVC voice agent on a given transport. Shared by the phone path
(Twilio Media Stream) and the browser path (WebRTC) — same prompt, model, voice, and
booking/hang-up logic; only the transport differs. do_capture writes the post-call
appointment to Odoo (on for phone; off for browser testing so it doesn't make cards)."""
stt = HintedWhisperSTTService(
settings=WhisperSTTService.Settings(model=WHISPER_MODEL),
device=WHISPER_DEVICE,
compute_type=WHISPER_COMPUTE,
hotwords=WHISPER_HOTWORDS,
)
llm = build_llm_service()
# In-call booking tool — only registered when ENABLE_TOOLS is on (auto: Claude yes,
# local Ollama no, since llama3.1:8b over-calls/leaks). The handler is a closure so it
# can stamp the verified caller-ID + call_sid onto the lead (the model never supplies a
# phone number — we don't ask for one). With tools on, this writes the Odoo lead IN-CALL,
# so the post-call extraction is skipped below to avoid a duplicate.
if ENABLE_TOOLS:
async def _record_appointment(params):
args = params.arguments or {}
if do_capture:
from practice import persist_appointment
persist_appointment({
"call_sid": call_sid,
"patient_name": args.get("patient_name"),
"callback_number": caller_number, # verified caller-ID, not model-supplied
"location": args.get("location"),
"reason": args.get("reason"),
"preferred_time": args.get("preferred_time"),
"source": "in_call_tool",
})
else:
logger.info(f"[capture off] would record appointment: {args.get('patient_name')} / {args.get('location')}")
await params.result_callback(
{"status": "recorded", "message": "Recorded — staff will call to confirm the time."}
)
llm.register_function("record_appointment_request", _record_appointment)
tts = SpokenKokoroTTSService(
model_path=os.path.join(MODEL_DIR, "kokoro-v1.0.onnx"),
voices_path=os.path.join(MODEL_DIR, "voices-v1.0.bin"),
settings=KokoroTTSService.Settings(voice=KOKORO_VOICE),
)
vad = VADProcessor(vad_analyzer=SileroVADAnalyzer(params=VADParams(
confidence=VAD_CONFIDENCE,
start_secs=VAD_START_SECS,
stop_secs=VAD_STOP_SECS,
min_volume=VAD_MIN_VOLUME,
)))
heartbeat = AudioHeartbeat()
gate = HalfDuplexGate(tail_secs=ECHO_TAIL_SECS) if HALF_DUPLEX else None
# Per-call system message = static prompt + the caller-ID number to confirm. Inject it
# ALREADY spelled out digit-by-digit so the model repeats clean words instead of mangling
# the raw digits (e.g. reading 197 as "one hundred ninety-seven").
if caller_number:
caller_line = (
f"\n\nCALLER ID: the caller's number on file, written so you read it digit by digit, "
f"is: {_spoken_phone(caller_number)}. Near the end, state it back and invite a "
"correction only ('...; if that's not the best number, just let me know.') — do NOT "
"ask a yes/no question or wait for a 'yes'. Only change it if they give a different "
"number. Do not say it any earlier in the call."
)
else:
caller_line = (
"\n\nCALLER ID: no number is available — near the end, ask the caller for the best "
"phone number to reach them."
)
system_content = SYSTEM_PROMPT + caller_line
context_kwargs = {"messages": [{"role": "system", "content": system_content}]}
if ENABLE_TOOLS:
context_kwargs["tools"] = _build_tools()
context = LLMContext(**context_kwargs)
agg = LLMContextAggregatorPair(context)
# Deterministic phone-confirmation safety net: if the agent reaches a closing without
# having read the caller-ID back, EndCallProcessor speaks this scripted line first.
if caller_number:
_spoken = _spoken_phone(caller_number)
phone_confirm_line = (
f"Also, I have your number as {_spoken}; if that's not the best number, just let me know."
)
phone_marker = _spoken.split(",")[0].strip() # e.g. "nine seven three"
else:
phone_confirm_line = phone_marker = None
endcall = EndCallProcessor(phone_confirm_line=phone_confirm_line, phone_marker=phone_marker)
watchdog = SilenceWatchdog(
silence_secs=SILENCE_REPROMPT_SECS,
max_prompts=MAX_REPROMPTS,
reprompt_line="I'm sorry, I didn't catch that — are you still there?",
closing_line="I'll let you go for now — please call us back anytime. Goodbye.",
) if SILENCE_WATCHDOG else None
# Stereo recorder (caller left / agent right) at the end so it captures what the system
# actually received + sent — for review and to debug silence with evidence, not guesses.
audiobuffer = AudioBufferProcessor(num_channels=2) if RECORD_CALLS else None
pipeline = Pipeline(
[
transport.input(),
heartbeat,
*( [gate] if gate else [] ), # half-duplex echo gate, before the VAD
vad,
stt,
agg.user(),
llm,
endcall,
*( [watchdog] if watchdog else [] ), # re-prompt on caller silence
tts,
transport.output(),
agg.assistant(),
*( [audiobuffer] if audiobuffer else [] ), # record both directions
]
)
task = PipelineTask(
pipeline,
params=PipelineParams(
audio_in_sample_rate=PIPELINE_SAMPLE_RATE,
audio_out_sample_rate=PIPELINE_SAMPLE_RATE,
allow_interruptions=True,
),
)
if audiobuffer:
os.makedirs(RECORDINGS_DIR, exist_ok=True)
@audiobuffer.event_handler("on_audio_data")
async def _on_audio_data(buf, audio, sample_rate, num_channels):
import wave
from datetime import datetime
if not audio:
return
fname = f"{datetime.now():%Y%m%d-%H%M%S}_{call_sid or 'web'}.wav"
path = os.path.join(RECORDINGS_DIR, fname)
with wave.open(path, "wb") as wf:
wf.setnchannels(num_channels)
wf.setsampwidth(2)
wf.setframerate(sample_rate)
wf.writeframes(audio)
logger.info(f"Saved call recording: {path} "
f"({len(audio)} bytes, {num_channels}ch @ {sample_rate}Hz)")
@transport.event_handler("on_client_connected")
async def on_client_connected(transport, client):
logger.info("Client connected -- greeting")
if audiobuffer:
await audiobuffer.start_recording()
await task.queue_frames(
[TTSSpeakFrame(
f"Thank you for calling Advanced Vision Care, this is {AGENT_NAME}. "
"How can I help you today?"
)]
)
@transport.event_handler("on_client_disconnected")
async def on_client_disconnected(transport, client):
logger.info("Client disconnected -- ending task")
if audiobuffer:
await audiobuffer.stop_recording()
await task.queue_frame(EndFrame())
runner = PipelineRunner(handle_sigint=False)
await runner.run(task)
# Call is over. Post-call extraction is the capture path ONLY when in-call tools are
# off (local Ollama). With tools on (Claude), the booking was already written in-call,
# so skip extraction to avoid a duplicate lead.
if do_capture and not ENABLE_TOOLS:
try:
from extract import extract_and_record
await extract_and_record(
context.messages, OLLAMA_URL, OLLAMA_MODEL,
call_sid=call_sid, caller_number=caller_number,
)
except Exception:
logger.exception("Post-call appointment extraction failed")
async def run_call(websocket, serializer: TwilioFrameSerializer, caller_number=None, call_sid=None):
"""Phone entrypoint: wrap the Twilio Media Stream in a transport, run the shared agent."""
transport = FastAPIWebsocketTransport(
websocket=websocket,
params=FastAPIWebsocketParams(
audio_in_enabled=True,
audio_out_enabled=True,
audio_in_sample_rate=PIPELINE_SAMPLE_RATE,
audio_out_sample_rate=PIPELINE_SAMPLE_RATE,
add_wav_header=False,
serializer=serializer,
),
)
await run_agent(transport, caller_number=caller_number, call_sid=call_sid, do_capture=True)