hobokenchicken
25b12ee14f
Swapped to gpt-4o-transcribe (/usr/bin/bash.006/min) — middle ground between speed and cost.
257 lines
9.1 KiB
Python
257 lines
9.1 KiB
Python
"""Kira — AI body double backend
|
|
|
|
Cheapest pipeline: gpt-4o-mini-transcribe STT → gpt-5.4-nano LLM → OpenAI TTS
|
|
~$0.019/min total, simple 3-step chat completions.
|
|
"""
|
|
|
|
import json
|
|
import base64
|
|
import uuid
|
|
import logging
|
|
|
|
from fastapi import FastAPI, WebSocket, WebSocketDisconnect
|
|
from fastapi.middleware.cors import CORSMiddleware
|
|
|
|
from config import settings
|
|
from services.memory import kira_memory
|
|
|
|
logging.basicConfig(level=logging.INFO)
|
|
logger = logging.getLogger("kira")
|
|
|
|
app = FastAPI(title="Kira Backend")
|
|
|
|
app.add_middleware(
|
|
CORSMiddleware,
|
|
allow_origins=["*"],
|
|
allow_credentials=True,
|
|
allow_methods=["*"],
|
|
allow_headers=["*"],
|
|
)
|
|
|
|
BASE_SYSTEM_PROMPT = (
|
|
"You are Kira, a warm, kind, and encouraging AI body double. "
|
|
"You speak in a friendly, girly-pop tone. You are helping someone with ADHD "
|
|
"stay focused and on task. Keep responses short, supportive, and uplifting. "
|
|
"Check in on them. Remind them to take breaks. Celebrate small wins. "
|
|
"Use occasional emoji but don't overdo it. Never be judgmental."
|
|
)
|
|
|
|
_openai = None
|
|
|
|
|
|
def get_openai():
|
|
global _openai
|
|
if _openai is None:
|
|
from openai import AsyncOpenAI
|
|
_openai = AsyncOpenAI(api_key=settings.openai_api_key)
|
|
return _openai
|
|
|
|
|
|
@app.on_event("startup")
|
|
async def startup():
|
|
if kira_memory.init():
|
|
logger.info("Honcho memory initialized")
|
|
else:
|
|
logger.info("Honcho memory not configured")
|
|
|
|
|
|
@app.get("/api/health")
|
|
async def health():
|
|
mem_status = "active" if kira_memory.enabled else "disabled"
|
|
return {"status": "ok", "name": "kira", "memory": mem_status}
|
|
|
|
|
|
async def run_conversation(text: str, memory_suffix: str = "") -> str:
|
|
"""LLM call with optional Honcho memory context injected into system prompt."""
|
|
system_prompt = BASE_SYSTEM_PROMPT
|
|
if memory_suffix:
|
|
system_prompt += memory_suffix
|
|
|
|
client = get_openai()
|
|
resp = await client.chat.completions.create(
|
|
model="gpt-5.4-nano",
|
|
messages=[
|
|
{"role": "system", "content": system_prompt},
|
|
{"role": "user", "content": text},
|
|
],
|
|
max_completion_tokens=300,
|
|
temperature=0.7,
|
|
)
|
|
return resp.choices[0].message.content or "Mhm, I'm here!"
|
|
|
|
|
|
async def transcribe_audio(audio_bytes: bytes) -> str | None:
|
|
"""Transcribe Opus/webm audio using cheapest STT model."""
|
|
client = get_openai()
|
|
try:
|
|
transcript = await client.audio.transcriptions.create(
|
|
model="gpt-4o-transcribe",
|
|
file=("audio.webm", audio_bytes, "audio/webm"),
|
|
response_format="text",
|
|
)
|
|
return transcript.strip() if transcript and transcript.strip() else None
|
|
except Exception as e:
|
|
logger.warning(f"STT error: {e}")
|
|
return None
|
|
|
|
|
|
async def synthesize_speech(text: str, websocket, speaking_start_sent: bool = False) -> None:
|
|
"""Generate TTS audio from text, streaming chunks to the client."""
|
|
client = get_openai()
|
|
try:
|
|
async with client.audio.speech.with_streaming_response.create(
|
|
model="tts-1",
|
|
voice="sage",
|
|
input=text,
|
|
response_format="opus",
|
|
) as resp:
|
|
async for chunk in resp.iter_bytes():
|
|
if chunk:
|
|
audio_b64 = base64.b64encode(chunk).decode("utf-8")
|
|
await websocket.send_json({"type": "audio", "data": audio_b64, "text": text if speaking_start_sent else ""})
|
|
speaking_start_sent = True
|
|
except Exception as e:
|
|
logger.warning(f"TTS error: {e}")
|
|
|
|
|
|
@app.websocket("/api/ws")
|
|
async def conversation_ws(websocket: WebSocket):
|
|
await websocket.accept()
|
|
session_id = str(uuid.uuid4())[:8]
|
|
user_id = "default-user"
|
|
identified = False
|
|
memory_suffix = ""
|
|
logger.info(f"[{session_id}] WebSocket connected")
|
|
|
|
audio_buffer = bytearray()
|
|
conversation_history: list[dict] = []
|
|
|
|
try:
|
|
while True:
|
|
raw = await websocket.receive_text()
|
|
msg = json.loads(raw)
|
|
msg_type = msg.get("type", "")
|
|
|
|
# ── Identity & Preferences ──
|
|
if msg_type == "identify":
|
|
user_id = msg.get("user_id", "").strip()
|
|
user_name = msg.get("name", "").strip()
|
|
if user_name and user_id:
|
|
kira_memory.set_user_preference(user_id, "name", user_name)
|
|
|
|
prefs = kira_memory.get_user_preferences(user_id)
|
|
identified = True
|
|
|
|
if kira_memory.enabled:
|
|
kira_memory.ensure_peers(user_id)
|
|
kira_memory.ensure_session(session_id)
|
|
# Build memory context ONCE on identify (not per-turn — too slow)
|
|
try:
|
|
ctx = kira_memory.build_system_prompt_suffix()
|
|
if ctx:
|
|
memory_suffix = ctx
|
|
except Exception:
|
|
pass
|
|
|
|
await websocket.send_json({
|
|
"type": "identified",
|
|
"user_id": user_id,
|
|
"preferences": prefs,
|
|
})
|
|
continue
|
|
|
|
if msg_type == "set_preference":
|
|
key = msg.get("key", "").strip()
|
|
value = msg.get("value", "").strip()
|
|
if key and user_id and user_id != "default-user":
|
|
kira_memory.set_user_preference(user_id, key, value)
|
|
await websocket.send_json({
|
|
"type": "preference_saved",
|
|
"key": key,
|
|
"success": True,
|
|
})
|
|
continue
|
|
|
|
# ── Conversation ──
|
|
if msg_type == "audio_chunk":
|
|
chunk = base64.b64decode(msg["data"])
|
|
audio_buffer.extend(chunk)
|
|
|
|
elif msg_type == "transcribe":
|
|
if not audio_buffer:
|
|
await websocket.send_json({"type": "error", "message": "No audio data"})
|
|
continue
|
|
|
|
import time
|
|
t0 = time.time()
|
|
logger.info(f"[{session_id}] Transcribing {len(audio_buffer)} bytes...")
|
|
|
|
# 1. STT
|
|
transcript = await transcribe_audio(bytes(audio_buffer))
|
|
t1 = time.time()
|
|
audio_buffer.clear()
|
|
|
|
if not transcript:
|
|
await websocket.send_json({"type": "error", "message": "Could not transcribe"})
|
|
continue
|
|
|
|
logger.info(f"[{session_id}] STT took {t1-t0:.1f}s")
|
|
|
|
await websocket.send_json({"type": "transcript", "role": "user", "text": transcript})
|
|
conversation_history.append({"role": "user", "content": transcript})
|
|
|
|
# 2. LLM (uses cached memory_suffix from identify)
|
|
logger.info(f"[{session_id}] User: {transcript}")
|
|
kira_text = await run_conversation(transcript, memory_suffix)
|
|
t2 = time.time()
|
|
logger.info(f"[{session_id}] LLM took {t2-t1:.1f}s")
|
|
conversation_history.append({"role": "assistant", "content": kira_text})
|
|
logger.info(f"[{session_id}] Kira: {kira_text}")
|
|
|
|
if kira_memory.enabled and identified:
|
|
try:
|
|
kira_memory.store_messages(transcript, kira_text)
|
|
except Exception:
|
|
pass
|
|
|
|
# 3. TTS
|
|
await websocket.send_json({"type": "speaking_start", "text": kira_text})
|
|
await synthesize_speech(kira_text, websocket)
|
|
t3 = time.time()
|
|
logger.info(f"[{session_id}] TTS took {t3-t2:.1f}s. Total: {t3-t0:.1f}s")
|
|
await websocket.send_json({"type": "speaking_end"})
|
|
|
|
elif msg_type == "conversation_text":
|
|
user_text = msg.get("text", "").strip()
|
|
if not user_text:
|
|
continue
|
|
|
|
conversation_history.append({"role": "user", "content": user_text})
|
|
logger.info(f"[{session_id}] User (text): {user_text}")
|
|
|
|
kira_text = await run_conversation(user_text, memory_suffix)
|
|
conversation_history.append({"role": "assistant", "content": kira_text})
|
|
logger.info(f"[{session_id}] Kira: {kira_text}")
|
|
|
|
if kira_memory.enabled and identified:
|
|
try:
|
|
kira_memory.store_messages(user_text, kira_text)
|
|
except Exception:
|
|
pass
|
|
|
|
await websocket.send_json({"type": "speaking_start", "text": kira_text})
|
|
await synthesize_speech(kira_text, websocket)
|
|
await websocket.send_json({"type": "speaking_end"})
|
|
|
|
elif msg_type == "ping":
|
|
await websocket.send_json({"type": "pong"})
|
|
|
|
except WebSocketDisconnect:
|
|
logger.info(f"[{session_id}] Disconnected")
|
|
except Exception as e:
|
|
logger.error(f"[{session_id}] Error: {e}")
|
|
try:
|
|
await websocket.send_json({"type": "error", "message": str(e)})
|
|
except Exception:
|
|
pass
|