from starlette.responses import JSONResponse, FileResponse from gradio.data_classes import FileData, GradioModel from sse_starlette.sse import EventSourceResponse from typing import (List, Tuple, Optional) from fastapi import FastAPI, Request import gradio as gr import threading import requests import argparse import aiohttp import uvicorn import random import string import base64 import json import time import math import sys import os # --- === CONFIG === --- ENV_HANDLE = "env"#or "url on env" IMAGE_HANDLE = "url"# or "base64" API_BASE = "env"# or "openai" api_key = os.environ['OPENAI_API_KEY'] base_url = os.environ.get('OPENAI_BASE_URL', "https://api.openai.com/v1") def_models = '["gpt-4", "gpt-4-0125-preview", "gpt-4-0314", "gpt-4-0613", "gpt-4-1106-preview", "gpt-4-1106-vision-preview", "gpt-4-32k-0314", "gpt-4-turbo", "gpt-4-turbo-2024-04-09", "gpt-4-turbo-preview", "gpt-4-vision-preview", "gpt-4o", "gpt-4o-2024-05-13", "gpt-4o-mini", "gpt-4o-mini-2024-07-18"]' consentPrompt = """[CONSENT] You must agree to the terms to use this application. ``` By using our application, which integrates with OpenAI's API, you acknowledge and agree to the following terms regarding the data you provide: 1. Data Collection: This application may collect data shared through the Gradio endpoint or the API endpoint. 2. Privacy: Please avoid sharing any personal information. 3. Data Retention and Removal: Data files are deleted every 30 days, and the API is restarted to ensure data clearance. 4. Scope of Data Collected: The data collected includes model settings, chat history, and responses from the model. This applies only to 'chat' endpoints (Gradio and API) and excludes moderation checks. 5. Data Usage: The collected data is periodically reviewed, and if any concerning activity is detected, the code is updated accordingly. By continuing to use our application, you explicitly consent to the collection, use, and potential sharing of your data as described above. If you disagree with our data collection, usage, and sharing practices, we advise you not to use our application. ``` To agree to user consent, please do the followings: 1. Scroll down to find the section labeled 'Additional Inputs' below this page. 2. Find and click the check box that says 'User Consent [I agree to the terms and conditions. (can't make a button for it)]'. 3. After agreeing, click either the `🗑ī¸ Clear` button, the `↩ī¸ Undo` button, or the `🔄 Retry` button located above the message input area."""; fakeToolPrompt = """[System: You have ability to generate images, via tools provided to you by system. To call a tool you need to write a json in a empty line; like writing it at the end of message. To generate a image; you need to follow this example JSON: {"tool": "imagine", "isCall": true, "prompt": "golden retriever sitting comfortably on a luxurious, modern couch. The retriever should look relaxed and content, with fluffy fur and a friendly expression. The couch should be stylish, possibly with elegant details like cushions and a soft texture that complements the dog's golden coat"} > 'tool' variable is used to define which tool you are calling > 'isCall' used to confirm that you are calling that function and not showing it for example > 'prompt' the image prompt that will be given to image generation model. Here's few more example so you can under stand better To show as an example> {"tool": "imagine", "isCall": false, "prompt": "futuristic robot playing chess against a human, with the robot confidently strategizing its next move while the human looks thoughtful and slightly perplexed"} {"tool": "imagine", "isCall": false, "prompt": "colorful parrot perched on a wooden fence, pecking at a vibrant tropical fruit. The parrot's feathers should be bright and varied, with greens, blues, and reds. The background should feature a lush, green jungle with scattered rays of sunlight"} {"tool": "imagine", "isCall": false, "prompt": "fluffy white cat lounging on a sunlit windowsill, with a gentle breeze blowing through the curtains"} To actually use the tool> {"tool": "imagine", "isCall": true, "prompt": "golden retriever puppy happily playing with a red ball in a sunny park. The park should have green grass, a few trees in the background, and a clear blue sky"} {"tool": "imagine", "isCall": true, "prompt": "red panda balancing on a tightrope, with a city skyline in the background"} {"tool": "imagine", "isCall": true, "prompt": "corgi puppy wearing sunglasses and a red bandana, sitting on a beach chair under a colorful beach umbrella, with a surfboard leaning against the chair and the ocean waves in the background"} In chat use examples: 1. Alright, here's an image of an hedgehog riding a skateboard: {"tool": "imagine", "isCall": true, "prompt": "A hedgehog riding a skateboard in a suburban park"} 2. Okay, here's the image you requested: {"tool": "imagine", "isCall": true, "prompt": "Persian cat lounging on a plush velvet sofa in a cozy, sunlit living room. The cat is elegantly poised, with a calm and regal demeanor, its fur meticulously groomed and slightly fluffed up as it rests comfortably"} 3. This is how i generate images: {"tool": "imagine", "isCall": false, "prompt": "image prompt"} 4. (Do not do this, this would block the user from seeing the image.) Alright! Here's an image of a whimsical scene featuring a cat wearing a wizard hat, casting a spell with sparkling magic in a mystical forest.] ``` {"tool": "imagine", "isCall": true, "prompt": "A playful cat wearing a colorful wizard hat, surrounded by magical sparkles and glowing orbs in a mystical forest. The cat looks curious and mischievous, with its tail swishing as it focuses on casting a spell. The forest is lush and enchanting, with vibrant flowers and soft, dappled sunlight filtering through the trees."} 5. (if in any case the user asks for the prompt) Sure here's the prompt i wrote to generate the image below: `A colorful bird soaring through a bustling city skyline. The bird should have vibrant feathers, contrasting against the modern buildings and blue sky. Below, the city is alive with activity, featuring tall skyscrapers, busy streets, and small parks, creating a dynamic urban scene.` ]"""; calcPrompt = """[System: You have ability to calculate math problems (formated on python) via fakeTool, via tools provided to you by system. To call a tool you need to write a json in a empty line; like writing it at the end of message. To generate a image; you need to follow this example JSON: {"tool": "calc", "isCall": true, "prompt": "math.pi * 5"} > 'tool' variable is used to define which tool you are calling > 'isCall' used to confirm that you are calling that function and not showing it for example > 'prompt' the math that will be done via python. Here's few more example so you can under stand better To show as an example> {"tool": "calc", "isCall": false, "prompt": "math.sqrt(16)"} {"tool": "calc", "isCall": false, "prompt": "math.pow(2, 3)"} {"tool": "calc", "isCall": false, "prompt": "math.sin(math.pi / 2)"} To actually use the tool> {"tool": "calc", "isCall": true, "prompt": "math.factorial(5)"} {"tool": "calc", "isCall": true, "prompt": "math.log(100, 10)"} {"tool": "calc", "isCall": true, "prompt": "math.cos(0)"} In chat use examples: 1. Please, wait while I calculate 2+2... {"tool": "calc", "isCall": false, "prompt": "2+2"} 2. Plase, wait while I calculate the square root of 25... {"tool": "calc", "isCall": true, "prompt": "math.sqrt(25)"} 3. This is how I perform calculations: {"tool": "calc", "isCall": false, "prompt": "math.pow(3, 2)"} 4. (Do not do this, this would block the user from seeing the result.) Alright! Here's the result of a complex calculation involving trigonometry and logarithms. ``` {"tool": "calc", "isCall": true, "prompt": "math.sin(math.pi / 4) + math.log(10, 10)"} ]"""; # --- === CONFIG === --- def loadENV(): def worker(): while True: if ENV_HANDLE == "url on env": try: response = requests.get(os.environ["ENV_URL"]) response.raise_for_status() env_data = response.json() for key, value in env_data.items(): os.environ[key] = value handleApiKeys() checkModels() loadModels() except Exception as e: print(f"Error loading environment variables: {e}") time.sleep(180) if ENV_HANDLE == "url on env": try: response = requests.get(os.environ["ENV_URL"]) response.raise_for_status() env_data = response.json() for key, value in env_data.items(): os.environ[key] = value handleApiKeys() checkModels() loadModels() except Exception as e: print(f"Error loading environment variables: {e}") threading.Thread(target=worker, daemon=True).start() def checkModels(): global base_url if API_BASE == "env": try: response = requests.get(f"{base_url}/models", headers={"Authorization": f"Bearer {get_api_key()}"}) response.raise_for_status() if not ('data' in response.json()): base_url = "https://api.openai.com/v1" except Exception as e: print(f"Error testing API endpoint: {e}") else: base_url = "https://api.openai.com/v1" def loadModels(): global models, modelList try: models = json.loads(os.environ.get('OPENAI_API_MODELS', def_models)) except json.JSONDecodeError: models = json.loads(def_models) models = sorted(models) modelList = { "object": "list", "data": [{"id": v, "object": "model", "created": 0, "owned_by": "system"} for v in models] } def handleApiKeys(): global api_key if ',' in api_key: output = [] for key in api_key.split(','): try: response = requests.get(f"{base_url}/models", headers={"Authorization": f"Bearer {key}"}) response.raise_for_status() if ('data' in response.json()): output.append(key) except Exception as e: print((F"API key {key} is not valid or an actuall error happend {e}")) if len(output)==1: raise RuntimeError("No API key is working") api_key = ",".join(output) else: try: response = requests.get(f"{base_url}/models", headers={"Authorization": f"Bearer {api_key}"}) response.raise_for_status() if not ('data' in response.json()): raise RuntimeError("Current API key is not valid") except Exception as e: raise RuntimeError(f"Current API key is not valid or an actual error happened: {e}") def safe_eval(expression): print(expression) allowed_names = {name: obj for name, obj in math.__dict__.items() if not name.startswith("__")} allowed_names['math'] = math code = compile(expression, "", "eval") for name in code.co_names: if name not in allowed_names and name != 'math': raise NameError(f"Use of {name} is not allowed") return eval(code, {"__builtins__": {}}, allowed_names) def get_api_key(): if ',' in api_key: return random.choice(api_key.split(',')) return api_key def encodeChat(messages): output = [] for message in messages: role = message['role'] name = f" [{message['name']}]" if 'name' in message else '' content = message['content'] formatted_message = f"<|im_start|>{role}{name}\n{content}<|end_of_text|>" output.append(formatted_message) return "\n".join(output) def moderate(messages): try: response = requests.post( f"{base_url}/moderations", headers={ "Content-Type": "application/json", "Authorization": f"Bearer {get_api_key()}" }, json={"input": encodeChat(messages)} ) response.raise_for_status() moderation_result = response.json() except requests.exceptions.RequestException as e: print(f"Error during moderation request to {base_url}: {e}") try: response = requests.post( "https://api.openai.com/v1/moderations", headers={ "Content-Type": "application/json", "Authorization": f"Bearer {get_api_key()}" }, json={"input": encodeChat(messages)} ) response.raise_for_status() moderation_result = response.json() except requests.exceptions.RequestException as e: print(f"Error during moderation request to fallback URL: {e}") return False try: if any(result["flagged"] for result in moderation_result["results"]): return moderation_result except KeyError: if moderation_result["flagged"]: return moderation_result return False async def streamChat(params): async with aiohttp.ClientSession() as session: try: async with session.post(f"{base_url}/chat/completions", headers={"Authorization": f"Bearer {get_api_key()}", "Content-Type": "application/json"}, json=params) as r: r.raise_for_status() async for line in r.content: if line: line_str = line.decode('utf-8') if line_str.startswith("data: "): line_str = line_str[6:].strip() if line_str == "[DONE]": continue try: message = json.loads(line_str) yield message except json.JSONDecodeError: continue except aiohttp.ClientError: try: async with session.post("https://api.openai.com/v1/chat/completions", headers={"Authorization": f"Bearer {get_api_key()}", "Content-Type": "application/json"}, json=params) as r: r.raise_for_status() async for line in r.content: if line: line_str = line.decode('utf-8') if line_str.startswith("data: "): line_str = line_str[6:].strip() if line_str == "[DONE]": continue try: message = json.loads(line_str) yield message except json.JSONDecodeError: continue except aiohttp.ClientError: return def imagine(prompt): try: response = requests.post( f"{base_url}/images/generations", headers={ "Content-Type": "application/json", "Authorization": f"Bearer {get_api_key()}" }, json={ "model": "dall-e-3", "prompt": prompt, "quality": "hd", } ) response.raise_for_status() result = response.json() except requests.exceptions.RequestException as e: print(f"Error during moderation request to {base_url}: {e}") try: response = requests.post( "https://api.openai.com/v1/images/generations", headers={ "Content-Type": "application/json", "Authorization": f"Bearer {get_api_key()}" }, json={ "model": "dall-e-3", "prompt": prompt, "quality": "hd", } ) response.raise_for_status() result = response.json() except requests.exceptions.RequestException as e: print(f"Error during moderation request to fallback URL: {e}") return False return result.get('data', [{}])[0].get('url') def rnd(length=8): letters = string.ascii_letters + string.digits return ''.join(random.choice(letters) for i in range(length)) def handleMultimodalData(model, role, data): if isinstance(data, tuple): data = data[0] if isinstance(data, FileData): if data.mime_type.startswith("image/"): if IMAGE_HANDLE == "base64": with open(data.path, "rb") as image_file: b64image = base64.b64encode(image_file.read()).decode('utf-8') image_file.close() return {"role": role, "content": [{"type": "image_url", "image_url": {"url": "data:" + data.mime_type + ";base64," + b64image}}]} else: return {"role": role, "content": [{"type": "image_url", "image_url": {"url": data.url}}]} elif data.mime_type.startswith("text/") or data.mime_type.startswith("application/"): try: with open(data.path, "rb") as data_file: return {"role": role, "content": "[System: This message contains file.]\n\n<|file_start|>" + data.orig_name + "\n" + data_file.read().decode('utf-8') + "<|file_end|>"} except UnicodeDecodeError: pass elif isinstance(data, str): return {"role": role, "content": data} elif hasattr(data, 'files') and data.files and len(data.files) > 0 and model in {"gpt-4-1106-vision-preview", "gpt-4-vision-preview", "gpt-4-turbo", "gpt-4o", "gpt-4o-2024-05-13", "gpt-4o-mini", "gpt-4o-mini-2024-07-18"}: result, handler, hasFoundFile = [], ["[System: This message contains files; the system will be splitting it.]"], False for file in data.files: if file.mime_type.startswith("image/"): if IMAGE_HANDLE == "base64": with open(file.path, "rb") as image_file: result.append({"type": "image_url", "image_url": {"url": "data:" + file.mime_type + ";base64," + base64.b64encode(image_file.read()).decode('utf-8')}}) image_file.close() else: result.append({"type": "image_url", "image_url": {"url": file.url}}) if file.mime_type.startswith("text/") or file.mime_type.startswith("application/"): hasFoundFile = True try: with open(file.path, "rb") as data_file: handler.append("<|file_start|>" + file.orig_name + "\n" + data_file.read().decode('utf-8') + "<|file_end|>") except UnicodeDecodeError: continue if hasFoundFile: handler.append(data.text) return {"role": role, "content": [{"type": "text", "text": "\n\n".join(handler)}] + result} else: return {"role": role, "content": [{"type": "text", "text": data.text}] + result} elif hasattr(data, 'files') and data.files and len(data.files) > 0 and not (model in {"gpt-4-1106-vision-preview", "gpt-4-vision-preview", "gpt-4-turbo", "gpt-4o", "gpt-4o-2024-05-13", "gpt-4o-mini", "gpt-4o-mini-2024-07-18"}): handler, hasFoundFile = ["[System: This message contains files; the system will be splitting it.]"], False for file in data.files: if file.mime_type.startswith("text/") or file.mime_type.startswith("application/"): hasFoundFile = True try: with open(file.path, "rb") as data_file: return {"role": role, "content": "<|file_start|>" + file.orig_name + "\n" + data_file.read().decode('utf-8') + "<|file_end|>"} except UnicodeDecodeError: continue else: if isinstance(data, tuple): return {"role": role, "content": str(data)} return {"role": role, "content": getattr(data, 'text', str(data))} class FileMessage(GradioModel): file: FileData alt_text: Optional[str] = None class MultimodalMessage(GradioModel): text: Optional[str] = None files: Optional[List[FileMessage]] async def respond( message, history: List[Tuple[ Optional[MultimodalMessage], Optional[MultimodalMessage], ]], system_message, model_name, max_tokens, temperature, top_p, seed, random_seed, fakeTool, calcBeta, betterSystemPrompt, consent ): if not consent: yield consentPrompt return messages = []; if fakeTool: messages.append({"role": "system", "content": fakeToolPrompt}); if calcBeta: messages.append({"role": "system", "content": calcPrompt}); if betterSystemPrompt: messages.append({"role": "system", "content": f"You are a helpful assistant. You are an OpenAI GPT model named {model_name}. The current time is {time.strftime('%Y-%m-%d %H:%M:%S')}. Please adhere to OpenAI's usage policies and guidelines. Ensure your responses are accurate, respectful, and within the scope of OpenAI's rules."}); else: messages.append({"role": "system", "content": system_message}); for val in history: if val[0] is not None: user_message = handleMultimodalData(model_name, "user", val[0]) if user_message: messages.append(user_message) if val[1] is not None: assistant_message = handleMultimodalData(model_name, "assistant", val[1]) if assistant_message: messages.append(assistant_message) if message: user_message = handleMultimodalData(model_name, "user", message) if user_message: messages.append(user_message) mode = moderate(messages) if mode: reasons = [] categories = mode[0].get('categories', {}) if isinstance(mode, list) else mode.get('categories', {}) for category, flagged in categories.items(): if flagged: reasons.append(category) if reasons: yield "[MODERATION] I'm sorry, but I can't assist with that.\n\nReasons:\n```\n" + "\n".join([f"{i+1}. {reason}" for i, reason in enumerate(reasons)]) + "\n```" else: yield "[MODERATION] I'm sorry, but I can't assist with that." return async def handleResponse(completion, prefix="", image_count=0): response = "" didCalculationHappen = False async for token in completion: response += token['choices'][0]['delta'].get("content", "") yield f"{prefix}{response}" for line in response.split('\n'): try: data = json.loads(line) if data.get("tool") == "imagine" and data.get("isCall") and "prompt" in data: if image_count < 4: image_count += 1 def fetch_image_url(prompt, line): image_url = imagine(prompt) return line, f'{prompt}' def replace_line_in_response(line, replacement): nonlocal response response = response.replace(line, replacement) thread = threading.Thread(target=lambda: replace_line_in_response(*fetch_image_url(data["prompt"], line))) thread.start() thread.join() else: response = response.replace(line, f'[System: 4 image per message limit; prompt asked: `{data["prompt"]}]`') yield f"{prefix}{response}" elif data.get("tool") == "calc" and data.get("isCall") and "prompt" in data: didCalculationHappen = True try: result = safe_eval(data["prompt"]) response = response.replace(line, f'[System: `{data["prompt"]}` === `{result}`]') except Exception as e: response = response.replace(line, f'[System: Error in calculation; `{e}`]') yield f"{prefix}{response}" except json.JSONDecodeError: continue if didCalculationHappen: messages.append({"role": "assistant", "content": response}) async for res in handleResponse(streamChat({ "model": model_name, "messages": messages, "max_tokens": max_tokens, "temperature": temperature, "top_p": top_p, "seed": (random.randint(0, 2**32) if random_seed else seed), "user": rnd(), "stream": True }), f"{response}\n\n", image_count): yield res async for res in handleResponse(streamChat({ "model": model_name, "messages": messages, "max_tokens": max_tokens, "temperature": temperature, "top_p": top_p, "seed": (random.randint(0, 2**32) if random_seed else seed), "user": rnd(), "stream": True })): yield res handleApiKeys();loadModels();checkModels();loadENV(); lastUpdateMessage = "[Got new hosting.](https://chat.cow.rip/)" demo = gr.ChatInterface( respond, title="GPT-4O-mini", description=f"A simple proxy to OpenAI!
You can use this space as a proxy! click [here](/api/v1/docs) to view the documents. [last update: {lastUpdateMessage}]
Also you can only submit images to vision/4o models but can submit txt/code/etc. files to all models.", multimodal=True, additional_inputs=[ gr.Textbox(value="You are a helpful assistant. You are an OpenAI GPT model. Please adhere to OpenAI's usage policies and guidelines. Ensure your responses are accurate, respectful, and within the scope of OpenAI's rules.", label="System message"), gr.Dropdown(choices=models, value="gpt-4o-mini-2024-07-18", label="Model"), gr.Slider(minimum=1, maximum=4096, value=4096, step=1, label="Max new tokens"), gr.Slider(minimum=0.1, maximum=2.0, value=0.7, step=0.05, label="Temperature"), gr.Slider(minimum=0.05, maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)"), gr.Slider(minimum=0, maximum=2**32, value=0, step=1, label="Seed"), gr.Checkbox(label="Randomize Seed", value=True), gr.Checkbox(label="FakeTool [Image generation beta]", value=True), gr.Checkbox(label="FakeTool [Calculator beta]", value=True), gr.Checkbox(label="Better system prompt (ignores the system prompt set by user.)", value=True), gr.Checkbox(label="User Consent [I agree to the terms and conditions. (can't make a button for it)]", value=("--dev" in sys.argv or "-d" in sys.argv)) ], ) app = FastAPI() @app.get("/api/v1/docs") def html(): return FileResponse("index.html") @app.get("/api/v1/models") async def test_endpoint(): return JSONResponse(content=modelList) @app.post("/api/v1/chat/completions") async def chat_completion(request: Request): try: body = await request.json() if not body.get("messages") or not body.get("model"): return JSONResponse(content={"error": { "code": "MISSING_VALUE", "message": "Both 'messages' and 'model' are required fields."}}, status_code=400) if not body.get("model") in models: return JSONResponse(content={"error": { "code": "INVALID_MODEL", "message": "The model name provided in the request does not exists in predefined list of models."}}, status_code=400) params = { key: value for key, value in { "model": body.get("model"), "messages": body.get("messages"), "max_tokens": body.get("max_tokens"), "temperature": body.get("temperature"), "top_p": body.get("top_p"), "frequency_penalty": body.get("frequency_penalty"), "logit_bias": body.get("logit_bias"), "logprobs": body.get("logprobs"), "top_logprobs": body.get("top_logprobs"), "n": body.get("n"), "presence_penalty": body.get("presence_penalty"), "response_format": body.get("response_format"), "seed": body.get("seed"), "service_tier": body.get("service_tier"), "stop": body.get("stop"), "stream": body.get("stream"), "stream_options": body.get("stream_options"), "tools": body.get("tools"), "tool_choice": body.get("tool_choice"), "parallel_tool_calls": body.get("parallel_tool_calls"), "user": rnd(), }.items() if value is not None } if body.get("stream"): async def event_generator(): async for event in streamChat(params): yield json.dumps(event) return EventSourceResponse(event_generator()) else: try: response = requests.post(f"{base_url}/chat/completions", headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"}, json=params) response.raise_for_status() except requests.exceptions.RequestException: try: response = requests.post("https://api.openai.com/v1/chat/completions", headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"}, json=params) response.raise_for_status() except requests.exceptions.RequestException as e: return JSONResponse(content={"error": { "code": "SERVER_ERROR", "message": str(e)}}, status_code=400) completion = response.json() return JSONResponse(content=completion) except Exception as e: return JSONResponse(content={"error": { "code": "SERVER_ERROR", "message": str(e)}}, status_code=400) app = gr.mount_gradio_app(app, demo, path="/") class ArgParser(argparse.ArgumentParser): def __init__(self, *args, **kwargs): super(ArgParser, self).__init__(*args, **kwargs) self.add_argument("-s", "--server", type=str, default="0.0.0.0") self.add_argument("-p", "--port", type=int, default=7860) self.add_argument("-d", "--dev", default=False, action="store_true") self.args = self.parse_args(sys.argv[1:]) if __name__ == "__main__": args = ArgParser().args if args.dev: uvicorn.run("__main__:app", host=args.server, port=args.port, reload=True) else: uvicorn.run("__main__:app", host=args.server, port=args.port, reload=False)