from dotenv import load_dotenv
load_dotenv() # take environment variables from .env.
import traceback
from platformdirs import user_data_dir
import ast
import json
import queue
import os
import datetime
from .utils.bytes_to_wav import bytes_to_wav
import re
from fastapi import FastAPI, Request
from fastapi.responses import PlainTextResponse
from starlette.websockets import WebSocket, WebSocketDisconnect
from pathlib import Path
import asyncio
import urllib.parse
from .utils.kernel import put_kernel_messages_into_queue
from .i import configure_interpreter
from interpreter import interpreter
from ..utils.accumulator import Accumulator
from .utils.logs import setup_logging
from .utils.logs import logger
from ..utils.print_markdown import print_markdown
os.environ["STT_RUNNER"] = "server"
os.environ["TTS_RUNNER"] = "server"
markdown = """
○
*Starting...*
"""
print("")
print_markdown(markdown)
print("")
setup_logging()
accumulator = Accumulator()
app = FastAPI()
app_dir = user_data_dir('01')
conversation_history_path = os.path.join(app_dir, 'conversations', 'user.json')
SERVER_LOCAL_PORT = int(os.getenv('SERVER_LOCAL_PORT', 10001))
# This is so we only say() full sentences
def is_full_sentence(text):
return text.endswith(('.', '!', '?'))
def split_into_sentences(text):
return re.split(r'(?<=[.!?])\s+', text)
# Queues
from_computer = queue.Queue() # Just for computer messages from the device. Sync queue because interpreter.run is synchronous
from_user = asyncio.Queue() # Just for user messages from the device.
to_device = asyncio.Queue() # For messages we send.
# Switch code executor to device if that's set
if os.getenv('CODE_RUNNER') == "device":
# (This should probably just loop through all languages and apply these changes instead)
class Python:
# This is the name that will appear to the LLM.
name = "python"
def __init__(self):
self.halt = False
def run(self, code):
"""Generator that yields a dictionary in LMC Format."""
# Prepare the data
message = {"role": "assistant", "type": "code", "format": "python", "content": code}
# Unless it was just sent to the device, send it wrapped in flags
if not (interpreter.messages and interpreter.messages[-1] == message):
to_device.put({"role": "assistant", "type": "code", "format": "python", "start": True})
to_device.put(message)
to_device.put({"role": "assistant", "type": "code", "format": "python", "end": True})
# Stream the response
logger.info("Waiting for the device to respond...")
while True:
chunk = from_computer.get()
logger.info(f"Server received from device: {chunk}")
if "end" in chunk:
break
yield chunk
def stop(self):
self.halt = True
def terminate(self):
"""Terminates the entire process."""
# dramatic!! do nothing
pass
interpreter.computer.languages = [Python]
# Configure interpreter
interpreter = configure_interpreter(interpreter)
@app.get("/ping")
async def ping():
return PlainTextResponse("pong")
@app.websocket("/")
async def websocket_endpoint(websocket: WebSocket):
await websocket.accept()
receive_task = asyncio.create_task(receive_messages(websocket))
send_task = asyncio.create_task(send_messages(websocket))
try:
await asyncio.gather(receive_task, send_task)
except Exception as e:
logger.debug(traceback.format_exc())
logger.info(f"Connection lost. Error: {e}")
@app.post("/")
async def add_computer_message(request: Request):
body = await request.json()
text = body.get("text")
if not text:
return {"error": "Missing 'text' in request body"}, 422
message = {"role": "user", "type": "message", "content": text}
await from_user.put({"role": "user", "type": "message", "start": True})
await from_user.put(message)
await from_user.put({"role": "user", "type": "message", "end": True})
async def receive_messages(websocket: WebSocket):
while True:
try:
try:
data = await websocket.receive()
except Exception as e:
print(str(e))
return
if 'text' in data:
try:
data = json.loads(data['text'])
if data["role"] == "computer":
from_computer.put(data) # To be handled by interpreter.computer.run
elif data["role"] == "user":
await from_user.put(data)
else:
raise("Unknown role:", data)
except json.JSONDecodeError:
pass # data is not JSON, leave it as is
elif 'bytes' in data:
data = data['bytes'] # binary data
await from_user.put(data)
except WebSocketDisconnect as e:
if e.code == 1000:
logger.info("Websocket connection closed normally.")
return
else:
raise
async def send_messages(websocket: WebSocket):
while True:
message = await to_device.get()
#print(f"Sending to the device: {type(message)} {str(message)[:100]}")
try:
if isinstance(message, dict):
await websocket.send_json(message)
elif isinstance(message, bytes):
await websocket.send_bytes(message)
else:
raise TypeError("Message must be a dict or bytes")
except:
# Make sure to put the message back in the queue if you failed to send it
await to_device.put(message)
raise
async def listener():
while True:
try:
while True:
if not from_user.empty():
chunk = await from_user.get()
break
elif not from_computer.empty():
chunk = from_computer.get()
break
await asyncio.sleep(1)
message = accumulator.accumulate(chunk)
if message == None:
# Will be None until we have a full message ready
continue
# print(str(message)[:1000])
# At this point, we have our message
if message["type"] == "audio" and message["format"].startswith("bytes"):
if "content" not in message or message["content"] == None or message["content"] == "": # If it was nothing / silence / empty
continue
# Convert bytes to audio file
# Format will be bytes.wav or bytes.opus
mime_type = "audio/" + message["format"].split(".")[1]
audio_file_path = bytes_to_wav(message["content"], mime_type)
# For microphone debugging:
if False:
os.system(f"open {audio_file_path}")
import time
time.sleep(15)
text = stt(audio_file_path)
print("> ", text)
message = {"role": "user", "type": "message", "content": text}
# At this point, we have only text messages
if type(message["content"]) != str:
print("This should be a string, but it's not:", message["content"])
message["content"] = message["content"].decode()
# Custom stop message will halt us
if message["content"].lower().strip(".,! ") == "stop":
continue
# Load, append, and save conversation history
with open(conversation_history_path, 'r') as file:
messages = json.load(file)
messages.append(message)
with open(conversation_history_path, 'w') as file:
json.dump(messages, file, indent=4)
accumulated_text = ""
if any([m["type"] == "image" for m in messages]) and interpreter.llm.model.startswith("gpt-"):
interpreter.llm.model = "gpt-4-vision-preview"
interpreter.llm.supports_vision = True
for chunk in interpreter.chat(messages, stream=True, display=True):
if any([m["type"] == "image" for m in interpreter.messages]):
interpreter.llm.model = "gpt-4-vision-preview"
logger.debug("Got chunk:", chunk)
# Send it to the user
await to_device.put(chunk)
# Yield to the event loop, so you actually send it out
await asyncio.sleep(0.01)
if os.getenv('TTS_RUNNER') == "server":
# Speak full sentences out loud
if chunk["role"] == "assistant" and "content" in chunk and chunk["type"] == "message":
accumulated_text += chunk["content"]
sentences = split_into_sentences(accumulated_text)
# If we're going to speak, say we're going to stop sending text.
# This should be fixed probably, we should be able to do both in parallel, or only one.
if any(is_full_sentence(sentence) for sentence in sentences):
await to_device.put({"role": "assistant", "type": "message", "end": True})
if is_full_sentence(sentences[-1]):
for sentence in sentences:
await stream_tts_to_device(sentence)
accumulated_text = ""
else:
for sentence in sentences[:-1]:
await stream_tts_to_device(sentence)
accumulated_text = sentences[-1]
# If we're going to speak, say we're going to stop sending text.
# This should be fixed probably, we should be able to do both in parallel, or only one.
if any(is_full_sentence(sentence) for sentence in sentences):
await to_device.put({"role": "assistant", "type": "message", "start": True})
# If we have a new message, save our progress and go back to the top
if not from_user.empty():
# Check if it's just an end flag. We ignore those.
temp_message = await from_user.get()
if type(temp_message) is dict and temp_message.get("role") == "user" and temp_message.get("end"):
# Yup. False alarm.
continue
else:
# Whoops! Put that back
await from_user.put(temp_message)
with open(conversation_history_path, 'w') as file:
json.dump(interpreter.messages, file, indent=4)
# TODO: is triggering seemingly randomly
#logger.info("New user message recieved. Breaking.")
#break
# Also check if there's any new computer messages
if not from_computer.empty():
with open(conversation_history_path, 'w') as file:
json.dump(interpreter.messages, file, indent=4)
logger.info("New computer message recieved. Breaking.")
break
except:
traceback.print_exc()
async def stream_tts_to_device(sentence):
force_task_completion_responses = [
"the task is done",
"the task is impossible",
"let me know what you'd like to do next",
]
if sentence.lower().strip().strip(".!?").strip() in force_task_completion_responses:
return
for chunk in stream_tts(sentence):
await to_device.put(chunk)
def stream_tts(sentence):
audio_file = tts(sentence)
with open(audio_file, "rb") as f:
audio_bytes = f.read()
os.remove(audio_file)
file_type = "bytes.raw"
chunk_size = 1024
# Stream the audio
yield {"role": "assistant", "type": "audio", "format": file_type, "start": True}
for i in range(0, len(audio_bytes), chunk_size):
chunk = audio_bytes[i:i+chunk_size]
yield chunk
yield {"role": "assistant", "type": "audio", "format": file_type, "end": True}
from uvicorn import Config, Server
import os
import platform
from importlib import import_module
# these will be overwritten
HOST = ''
PORT = 0
@app.on_event("startup")
async def startup_event():
server_url = f"{HOST}:{PORT}"
print("")
print_markdown(f"\n*Ready.*\n")
print("")
@app.on_event("shutdown")
async def shutdown_event():
print_markdown("*Server is shutting down*")
async def main(server_host, server_port, llm_service, model, llm_supports_vision, llm_supports_functions, context_window, max_tokens, temperature, tts_service, stt_service):
global HOST
global PORT
PORT = server_port
HOST = server_host
# Setup services
application_directory = user_data_dir('01')
services_directory = os.path.join(application_directory, 'services')
service_dict = {'llm': llm_service, 'tts': tts_service, 'stt': stt_service}
# Create a temp file with the session number
session_file_path = os.path.join(user_data_dir('01'), '01-session.txt')
with open(session_file_path, 'w') as session_file:
session_id = int(datetime.datetime.now().timestamp() * 1000)
session_file.write(str(session_id))
for service in service_dict:
service_directory = os.path.join(services_directory, service, service_dict[service])
# This is the folder they can mess around in
config = {"service_directory": service_directory}
if service == "llm":
config.update({
"interpreter": interpreter,
"model": model,
"llm_supports_vision": llm_supports_vision,
"llm_supports_functions": llm_supports_functions,
"context_window": context_window,
"max_tokens": max_tokens,
"temperature": temperature
})
module = import_module(f'.server.services.{service}.{service_dict[service]}.{service}', package='source')
ServiceClass = getattr(module, service.capitalize())
service_instance = ServiceClass(config)
globals()[service] = getattr(service_instance, service)
interpreter.llm.completions = llm
# Start listening
asyncio.create_task(listener())
# Start watching the kernel if it's your job to do that
if True: # in the future, code can run on device. for now, just server.
asyncio.create_task(put_kernel_messages_into_queue(from_computer))
config = Config(app, host=server_host, port=int(server_port), lifespan='on')
server = Server(config)
await server.serve()
# Run the FastAPI app
if __name__ == "__main__":
asyncio.run(main())