👋 Welcome to the LLMChat repository, a full-stack implementation of an API server built with Python FastAPI, and a beautiful frontend powered by Flutter. 💬 This project is designed to deliver a seamless chat experience with the advanced ChatGPT and other LLM models. 🔝 Offering a modern infrastructure that can be easily extended when GPT-4's Multimodal and Plugin features become available. 🚀 Enjoy your stay!
- It supports both
mobileandPCenvironments. Markdownis also supported, so you can use it to format your messages.
-
You can use the Duckduckgo search engine to find relevant information on the web. Just activate the 'Browse' toggle button!
Watch the demo video for full-browsing: https://www.youtube.com/watch?v=mj_CVrWrS08
-
With the
/embedcommand, you can store the text indefinitely in your own private vector database and query it later, anytime. If you use the/sharecommand, the text is stored in a public vector database that everyone can share. EnablingQuerytoggle button or/querycommand helps the AI generate contextualized answers by searching for text similarities in the public and private databases. This solves one of the biggest limitations of language models: memory. -
You can embed PDF file by clicking
Embed Documenton the bottom left. In a few seconds, text contents of PDF will be converted to vectors and embedded to Redis cache.
-
You can change your chat model by dropdown menu. You can define whatever model you want to use in
LLMModelswhich is located inapp/models/llms.py.
-
You can change your chat title by clicking the title of the chat. This will be stored until you change or delete it!
For the local Llalam LLMs, it is assumed to work only in the local environment and uses the http://localhost:8002/v1/completions endpoint. It continuously checks the status of the llama API server by connecting to http://localhost:8002/health once a second to see if a 200 OK response is returned, and if not, it automatically runs a separate process to create a the API server.
The main goal of llama.cpp is to run the LLaMA model using GGML 4-bit quantization with plain C/C++ implementation without dependencies. You have to download GGML bin file from huggingface and put it in the llama_models/ggml folder, and define LLMModel in app/models/llms.py. There are few examples, so you can easily define your own model.
Refer to the llama.cpp repository for more information: https://github.com/ggerganov/llama.cpp
A standalone Python/C++/CUDA implementation of Llama for use with 4-bit GPTQ weights, designed to be fast and memory-efficient on modern GPUs. It uses pytorch and sentencepiece to run the model. It is assumed to work only in the local environment and at least one NVIDIA CUDA GPU is required. You have to download tokenizer, config, and GPTQ files from huggingface and put it in the llama_models/gptq/YOUR_MODEL_FOLDER folder, and define LLMModel in app/models/llms.py. There are few examples, so you can easily define your own model. Refer to the exllama repository for more detailed information: https://github.com/turboderp/exllama
- FastAPI - High-performance
web frameworkfor building APIs with Python. - Flutter -
Webappfrontend with beautiful UI and rich set of customizable widgets. - ChatGPT - Seamless integration with the
OpenAI APIfor text generation and message management. - LLAMA - Suporting LocalLLM,
LlamaCppandExllamamodels. - WebSocket Connection -
Real-time, two-way communication with the ChatGPT, and other LLM models, with Flutter frontend webapp. - Vectorstore - Using
RedisandLangchain, store and retrieve vector embeddings for similarity search. It will help AI to generate more relevant responses. - Auto summarization - Using Langchain's summarize chain, summarize the conversation and store it in the database. It will help saving a lot of tokens.
- Web Browsing - Using
Duckduckgosearch engine, browse the web and find relevant information. - Concurrency - Asynchronous programming with
async/awaitsyntax for concurrency and parallelism. - Security - Token validation and authentication to keep API secure.
- Database - Manage database connections and execute
MySQLqueries. Easily perform Create, Read, Update, and Delete actions, withsqlalchemy.asyncio - Cache - Manage cache connections and execute
Redisqueries with aioredis. Easily perform Create, Read, Update, and Delete actions, withaioredis.
To set up the on your local machine, follow these simple steps.
Before you begin, ensure you have docker and docker-compose installed on your machine. If you want to run the server without docker, you have to install Python 3.11 additionally. Even though, you need Docker to run DB servers.
To recursively clone the submodules to use Exllama or llama.cpp models, use the following command:
git clone --recurse-submodules https://github.com/c0sogi/llmchat.gitYou only want to use core features(OpenAI), use the following command:
git clone https://github.com/c0sogi/llmchat.gitcd LLMChatSetup an env file, referring to .env-sample file. Enter database information to create, OpenAI API Key, and other necessary configurations. Optionals are not required, just leave them as they are.
Execute these. It may take a few minutes to start the server for the first time:
docker-compose -f docker-compose-local.yaml updocker-compose -f docker-compose-local.yaml downNow you can access the server at http://localhost:8000/docs and the database at db:3306 or cache:6379. You can also access the app at http://localhost:8000/chat.
-
To run the server without docker If you want to run the server without docker, you have to install
Python 3.11additionally. Even though, you needDockerto run DB servers. Turn off the API server already running withdocker-compose -f docker-compose-local.yaml down api. Don't forget to run other DB servers on Docker! Then, run the following commands:python -m main
Your Server should now be up and running on
http://localhost:8001in this case.
This project is licensed under the MIT License, which allows for free use, modification, and distribution, as long as the original copyright and license notice are included in any copy or substantial portion of the software.
🚀 FastAPI is a modern web framework for building APIs with Python.
💪 It has high performance, easy to learn, fast to code, and ready for production.
👍 One of the main features of FastAPI is that it supports concurrency and async/await syntax.
🤝 This means that you can write code that can handle multiple tasks at the same time without blocking each other, especially when dealing with I/O bound operations, such as network requests, database queries, file operations, etc.
📱 Flutter is an open-source UI toolkit developed by Google for building native user interfaces for mobile, web, and desktop platforms from a single codebase.
👨💻 It uses Dart, a modern object-oriented programming language, and provides a rich set of customizable widgets that can adapt to any design.
You can access ChatGPT or LlamaCpp through WebSocket connection using two modules: app/routers/websocket and app/utils/chat/chat_stream_manager. These modules facilitate the communication between the Flutter client and the Chat model through a WebSocket. With the WebSocket, you can establish a real-time, two-way communication channel to interact with the LLM.
To start a conversation, connect to the WebSocket route /ws/chat/{api_key} with a valid API key registered in the database. Note that this API key is not the same as OpenAI API key, but only available for your server to validate the user. Once connected, you can send messages and commands to interact with the LLM model. The WebSocket will send back chat responses in real-time. This websocket connection is established via Flutter app, which can accessed with endpoint /chat.
websocket.py is responsible for setting up a WebSocket connection and handling user authentication. It defines the WebSocket route /chat/{api_key} that accepts a WebSocket and an API key as parameters.
When a client connects to the WebSocket, it first checks the API key to authenticate the user. If the API key is valid, the begin_chat() function is called from the stream_manager.py module to start the conversation.
In case of an unregistered API key or an unexpected error, an appropriate message is sent to the client and the connection is closed.
@router.websocket("/chat/{api_key}")
async def ws_chat(websocket: WebSocket, api_key: str):
...stream_manager.py is responsible for managing the conversation and handling user messages. It defines the begin_chat() function, which takes a WebSocket, a user ID as parameters.
The function first initializes the user's chat context from the cache manager. Then, it sends the initial message history to the client through the WebSocket.
The conversation continues in a loop until the connection is closed. During the conversation, the user's messages are processed and GPT's responses are generated accordingly.
class ChatStreamManager:
@classmethod
async def begin_chat(cls, websocket: WebSocket, user: Users) -> None:
...The SendToWebsocket class is used for sending messages and streams to the WebSocket. It has two methods: message() and stream(). The message() method sends a complete message to the WebSocket, while the stream() method sends a stream to the WebSocket.
class SendToWebsocket:
@staticmethod
async def message(...):
...
@staticmethod
async def stream(...):
...The MessageHandler class also handles AI responses. The ai() method sends the AI response to the WebSocket. If translation is enabled, the response is translated using the Google Translate API before sending it to the client.
class MessageHandler:
...
@staticmethod
async def ai(...):
...User messages are processed using the HandleMessage class. If the message starts with /, such as /YOUR_CALLBACK_NAME. it is treated as a command and the appropriate command response is generated. Otherwise, the user's message is processed and sent to the LLM model for generating a response.
Commands are handled using the ChatCommands class. It executes the corresponding callback function depending on the command. You can add new commands by simply adding callback in ChatCommands class from app.utils.chat.chat_commands.
Using Redis for storing vector embeddings of conversations 🗨️ can aid the ChatGPT model 🤖 in several ways, such as efficient and fast retrieval of conversation context 🕵️♀️, handling large amounts of data 📊, and providing more relevant responses through vector similarity search 🔎.
Some fun examples of how this could work in practice:
- Imagine a user is chatting with ChatGPT about their favorite TV show 📺 and mentions a specific character 👤. Using Redis, ChatGPT could retrieve previous conversations where that character was mentioned and use that information to provide more detailed insights or trivia about that character 🤔.
- Another scenario could be a user discussing their travel plans
✈️ with ChatGPT. If they mention a particular city 🌆 or landmark 🏰, ChatGPT could use vector similarity search to retrieve previous conversations that discussed the same location and provide recommendations or tips based on that context 🧳. - If a user mentions a particular cuisine 🍝 or dish 🍱, ChatGPT could retrieve previous conversations that discussed those topics and provide recommendations or suggestions based on that context 🍴.
When a user enters a command in the chat window like /embed <text_to_embed>, the VectorStoreManager.create_documents method is called. This method converts the input text into a vector using OpenAI's text-embedding-ada-002 model and stores it in the Redis vectorstore.
@staticmethod
@command_response.send_message_and_stop
async def embed(text_to_embed: str, /, buffer: BufferedUserContext) -> str:
"""Embed the text and save its vectors in the redis vectorstore.\n
/embed <text_to_embed>"""
...When the user enters the /query <query> command, the asimilarity_search function is used to find up to three results with the highest vector similarity to the embedded data in the Redis vectorstore. These results are temporarily stored in the context of the chat, which helps AI answer the query by referring to these data.
@staticmethod
async def query(query: str, /, buffer: BufferedUserContext, **kwargs) -> Tuple[str | None, ResponseType]:
"""Query from redis vectorstore\n
/query <query>"""
...When running the begin_chat function, if a user uploads a file containing text (e.g., a PDF or TXT file), the text is automatically extracted from the file, and its vector embedding is saved to Redis.
@classmethod
async def embed_file_to_vectorstore(cls, file: bytes, filename: str, collection_name: str) -> str:
# if user uploads file, embed it
...In the commands.py file, there are several important components:
command_response: This class is used to set a decorator on the command method to specify the next action. It helps to define various response types, such as sending a message and stopping, sending a message and continuing, handling user input, handling AI responses, and more.command_handler: This function is responsible for performing a command callback method based on the text entered by the user.arguments_provider: This function automatically supplies the arguments required by the command method based on the annotation type of the command method.
There is a way to save tokens by adding a task to the LLM that summarizes the message. The auto summarization task is a crucial feature that enhances the efficiency of chatbot. Let's break down the functionality of this feature:
-
Task Triggering: This feature is activated whenever a user types a message or the AI responds with a message. At this point, an automatic summarization task is generated to condense the text content.
-
Task Storage: The auto-summarization task is then stored in the
task_listattribute of theBufferUserChatContext. This serves as a queue for managing tasks linked to the user's chat context. -
Task Harvesting: Following the completion of a user-AI question and answer cycle by the
MessageHandler, theharvest_done_tasksfunction is invoked. This function collects the results of the summarization task, making sure nothing is left out. -
Summarization Application: After the harvesting process, the summarized results replace the actual message when our chatbot is requesting answers from language learning models (LLMs), such as OPENAI and LLAMA_CPP. By doing so, we're able to send much more succinct prompts than the initial lengthy message.
-
User Experience: Importantly, from the user's perspective, they only see the original message. The summarized version of the message is not shown to them, maintaining transparency and avoiding potential confusion.
-
Simultaneous Tasks: Another key feature of this auto-summarization task is that it doesn't impede other tasks. In other words, while the chatbot is busy summarizing the text, other tasks can still be carried out, thereby improving the overall efficiency of our chatbot.
By default, summarize chain only works for messages of 512 tokens or more. This can be turned on/off and the threshold set in ChatConfig.
This repository contains different LLM models, defined in llms.py. Each LLM Model class inherit from the base class LLMModel. The LLMModels enum is a collection of these LLMs.
All operations are handled asynchronously without interupting the main thread. However, Local LLMs are not be able to handle multiple requests at the same time, as they are too computationally expensive. Therefore, a Semaphore is used to limit the number of requests to 1.
The default LLM model used by the user via UserChatContext.construct_default is gpt-3.5-turbo. You can change the default for that function.
OpenAIModel generates text asynchronously by requesting chat completion from the OpenAI server. It requires an OpenAI API key.
LlamaCppModel reads a locally stored GGML model. The LLama.cpp GGML model must be put in the llama_models/ggml folder as a .bin file. For example, if you downloaded a q4_0 quantized model from "https://huggingface.co/TheBloke/robin-7B-v2-GGML",
The path of the model has to be "robin-7b.ggmlv3.q4_0.bin".
ExllamaModel read a locally stored GPTQ model. The Exllama GPTQ model must be put in the llama_models/gptq folder as a folder. For example, if you downloaded 3 files from "https://huggingface.co/TheBloke/orca_mini_7B-GPTQ/tree/main":
- orca-mini-7b-GPTQ-4bit-128g.no-act.order.safetensors
- tokenizer.model
- config.json
Then you need to put them in a folder. The path of the model has to be the folder name. Let's say, "orca_mini_7b", which contains the 3 files.
Handle exceptions that may occur during text generation. If a ChatLengthException is thrown, it automatically performs a routine to re-limit the message to within the number of tokens limited by the cutoff_message_histories function, and resend it. This ensures that the user has a smooth chat experience regardless of the token limit.
This project aims to create an API backend to enable the large language model chatbot service. It utilizes a cache manager to store messages and user profiles in Redis, and a message manager to safely cache messages so that the number of tokens does not exceed an acceptable limit.
The Cache Manager (CacheManager) is responsible for handling user context information and message histories. It stores these data in Redis, allowing for easy retrieval and modification. The manager provides several methods to interact with the cache, such as:
read_context_from_profile: Reads the user's chat context from Redis, according to the user's profile.create_context: Creates a new user chat context in Redis.reset_context: Resets the user's chat context to default values.update_message_histories: Updates the message histories for a specific role (user, ai, or system).lpop_message_history/rpop_message_history: Removes and returns the message history from the left or right end of the list.append_message_history: Appends a message history to the end of the list.get_message_history: Retrieves the message history for a specific role.delete_message_history: Deletes the message history for a specific role.set_message_history: Sets a specific message history for a role and index.
The Message Manager (MessageManager) ensures that the number of tokens in message histories does not exceed the specified limit. It safely handles adding, removing, and setting message histories in the user's chat context while maintaining token limits. The manager provides several methods to interact with message histories, such as:
add_message_history_safely: Adds a message history to the user's chat context, ensuring that the token limit is not exceeded.pop_message_history_safely: Removes and returns the message history from the right end of the list while updating the token count.set_message_history_safely: Sets a specific message history in the user's chat context, updating the token count and ensuring that the token limit is not exceeded.
To use the cache manager and message manager in your project, import them as follows:
from app.utils.chat.managers.cache import CacheManager
from app.utils.chat.message_manager import MessageManagerThen, you can use their methods to interact with the Redis cache and manage message histories according to your requirements.
For example, to create a new user chat context:
user_id = "example@user.com" # email format
chat_room_id = "example_chat_room_id" # usually the 32 characters from `uuid.uuid4().hex`
default_context = UserChatContext.construct_default(user_id=user_id, chat_room_id=chat_room_id)
await CacheManager.create_context(user_chat_context=default_context)To safely add a message history to the user's chat context:
user_chat_context = await CacheManager.read_context_from_profile(user_chat_profile=UserChatProfile(user_id=user_id, chat_room_id=chat_room_id))
content = "This is a sample message."
role = ChatRoles.USER # can be enum such as ChatRoles.USER, ChatRoles.AI, ChatRoles.SYSTEM
await MessageManager.add_message_history_safely(user_chat_context, content, role)This project uses token_validator middleware and other middlewares used in the FastAPI application. These middlewares are responsible for controlling access to the API, ensuring only authorized and authenticated requests are processed.
The following middlewares are added to the FastAPI application:
- Access Control Middleware: Ensures that only authorized requests are processed.
- CORS Middleware: Allows requests from specific origins, as defined in the app configuration.
- Trusted Host Middleware: Ensures that requests are coming from trusted hosts, as defined in the app configuration.
The Access Control Middleware is defined in the token_validator.py file. It is responsible for validating API keys and JWT tokens.
The StateManager class is used to initialize request state variables. It sets the request time, start time, IP address, and user token.
The AccessControl class contains two static methods for validating API keys and JWT tokens:
api_service: Validates API keys by checking the existence of required query parameters and headers in the request. It calls theValidator.api_keymethod to verify the API key, secret, and timestamp.non_api_service: Validates JWT tokens by checking the existence of the 'authorization' header or 'Authorization' cookie in the request. It calls theValidator.jwtmethod to decode and verify the JWT token.
The Validator class contains two static methods for validating API keys and JWT tokens:
api_key: Verifies the API access key, hashed secret, and timestamp. Returns aUserTokenobject if the validation is successful.jwt: Decodes and verifies the JWT token. Returns aUserTokenobject if the validation is successful.
The access_control function is an asynchronous function that handles the request and response flow for the middleware. It initializes the request state using the StateManager class, determines the type of authentication required for the requested URL (API key or JWT token), and validates the authentication using the AccessControl class. If an error occurs during the validation process, an appropriate HTTP exception is raised.
Token utilities are defined in the token.py file. It contains two functions:
create_access_token: Creates a JWT token with the given data and expiration time.token_decode: Decodes and verifies a JWT token. Raises an exception if the token is expired or cannot be decoded.
The params_utils.py file contains a utility function for hashing query parameters and secret key using HMAC and SHA256:
hash_params: Takes query parameters and secret key as input and returns a base64 encoded hashed string.
The date_utils.py file contains the UTC class with utility functions for working with dates and timestamps:
now: Returns the current UTC datetime with an optional hour difference.timestamp: Returns the current UTC timestamp with an optional hour difference.timestamp_to_datetime: Converts a timestamp to a datetime object with an optional hour difference.
The logger.py file contains the ApiLogger class, which logs API request and response information, including the request URL, method, status code, client information, processing time, and error details (if applicable). The logger function is called at the end of the access_control function to log the processed request and response.
To use the token_validator middleware in your FastAPI application, simply import the access_control function and add it as a middleware to your FastAPI instance:
from app.middlewares.token_validator import access_control
app = FastAPI()
app.add_middleware(dispatch=access_control, middleware_class=BaseHTTPMiddleware)Make sure to also add the CORS and Trusted Host middlewares for complete access control:
app.add_middleware(
CORSMiddleware,
allow_origins=config.allowed_sites,
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
app.add_middleware(
TrustedHostMiddleware,
allowed_hosts=config.trusted_hosts,
except_path=["/health"],
)Now, any incoming requests to your FastAPI application will be processed by the token_validator middleware and other middlewares, ensuring that only authorized and authenticated requests are processed.
This module app.database.connection provides an easy-to-use interface for managing database connections and executing SQL queries using SQLAlchemy and Redis. It supports MySQL, and can be easily integrated with this project.
- Create and drop databases
- Create and manage users
- Grant privileges to users
- Execute raw SQL queries
- Manage database sessions with async support
- Redis caching support for faster data access
First, import the required classes from the module:
from app.database.connection import MySQL, SQLAlchemy, CacheFactoryNext, create an instance of the SQLAlchemy class and configure it with your database settings:
from app.common.config import Config
config: Config = Config.get()
db = SQLAlchemy()
db.start(config)Now you can use the db instance to execute SQL queries and manage sessions:
# Execute a raw SQL query
result = await db.execute("SELECT * FROM users")
# Use the run_in_session decorator to manage sessions
@db.run_in_session
async def create_user(session, username, password):
await session.execute("INSERT INTO users (username, password) VALUES (:username, :password)", {"username": username, "password": password})
await create_user("JohnDoe", "password123")To use Redis caching, create an instance of the CacheFactory class and configure it with your Redis settings:
cache = CacheFactory()
cache.start(config)You can now use the cache instance to interact with Redis:
# Set a key in Redis
await cache.redis.set("my_key", "my_value")
# Get a key from Redis
value = await cache.redis.get("my_key")In fact, in this project, the MySQL class does the initial setup at app startup, and all database connections are made with only the db and cache variables present at the end of the module. 😅
All db settings will be done in create_app() in app.common.app_settings.
For example, the create_app() function in app.common.app_settings will look like this:
def create_app(config: Config) -> FastAPI:
# Initialize app & db & js
new_app = FastAPI(
title=config.app_title,
description=config.app_description,
version=config.app_version,
)
db.start(config=config)
cache.start(config=config)
js_url_initializer(js_location="app/web/main.dart.js")
# Register routers
# ...
return new_appThis project uses simple and efficient way to handle database CRUD (Create, Read, Update, Delete) operations using SQLAlchemy and two module and path: app.database.models.schema and app.database.crud.
The schema.py module is responsible for defining database models and their relationships using SQLAlchemy. It includes a set of classes that inherit from Base, an instance of declarative_base(). Each class represents a table in the database, and its attributes represent columns in the table. These classes also inherit from a Mixin class, which provides some common methods and attributes for all the models.
The Mixin class provides some common attributes and methods for all the classes that inherit from it. Some of the attributes include:
id: Integer primary key for the table.created_at: Datetime for when the record was created.updated_at: Datetime for when the record was last updated.ip_address: IP address of the client that created or updated the record.
It also provides several class methods that perform CRUD operations using SQLAlchemy, such as:
add_all(): Adds multiple records to the database.add_one(): Adds a single record to the database.update_where(): Updates records in the database based on a filter.fetchall_filtered_by(): Fetches all records from the database that match the provided filter.one_filtered_by(): Fetches a single record from the database that matches the provided filter.first_filtered_by(): Fetches the first record from the database that matches the provided filter.one_or_none_filtered_by(): Fetches a single record or returnsNoneif no records match the provided filter.
The users.py and api_keys.py module contains a set of functions that perform CRUD operations using the classes defined in schema.py. These functions use the class methods provided by the Mixin class to interact with the database.
Some of the functions in this module include:
create_api_key(): Creates a new API key for a user.get_api_keys(): Retrieves all API keys for a user.get_api_key_owner(): Retrieves the owner of an API key.get_api_key_and_owner(): Retrieves an API key and its owner.update_api_key(): Updates an API key.delete_api_key(): Deletes an API key.is_email_exist(): Checks if an email exists in the database.get_me(): Retrieves user information based on user ID.is_valid_api_key(): Checks if an API key is valid.register_new_user(): Registers a new user in the database.find_matched_user(): Finds a user with a matching email in the database.
To use the provided CRUD operations, import the relevant functions from the crud.py module and call them with the required parameters. For example:
import asyncio
from app.database.crud.users import register_new_user, get_me, is_email_exist
from app.database.crud.api_keys import create_api_key, get_api_keys, update_api_key, delete_api_key
async def main():
# `user_id` is an integer index in the MySQL database, and `email` is user's actual name
# the email will be used as `user_id` in chat. Don't confuse with `user_id` in MySQL
# Register a new user
new_user = await register_new_user(email="test@test.com", hashed_password="...")
# Get user information
user = await get_me(user_id=1)
# Check if an email exists in the database
email_exists = await is_email_exist(email="test@test.com")
# Create a new API key for user with ID 1
new_api_key = await create_api_key(user_id=1, additional_key_info={"user_memo": "Test API Key"})
# Get all API keys for user with ID 1
api_keys = await get_api_keys(user_id=1)
# Update the first API key in the list
updated_api_key = await update_api_key(updated_key_info={"user_memo": "Updated Test API Key"}, access_key_id=api_keys[0].id, user_id=1)
# Delete the first API key in the list
await delete_api_key(access_key_id=api_keys[0].id, access_key=api_keys[0].access_key, user_id=1)
if __name__ == "__main__":
asyncio.run(main())