Skip to content

Latest commit

 

History

History
163 lines (129 loc) · 8.23 KB

README.md

File metadata and controls

163 lines (129 loc) · 8.23 KB
Raw

LLM Slack Bot Channels

LLM-SlackBot-Channels is a Slack bot developed using the Slack Bolt framework. It allows users to interact with the bot through Slack channels by employing various commands. The bot leverages a Large Language Model (LLM) to generate responses based on both user input and channel-specific configurations. Unique to each channel, the bot can adopt different personalities and follow a customized set of instructions. This includes the ability to use it as an agent, integrating tools and documents.

This repository mainly uses langchain. It supports the usage of open LLMs and embeddings, and OpenAI's models as well.

See a video example here:

example_usage.mp4

What's new? v0.2

  • Improve the way a file is uploaded to a QA thread
  • If using OpenAI models, then you can customize which model you want to use in each channel
  • Use the LLM model as an Agent with your own tools! see more in v0.2 release
    • You can add files to the channel, which are used by the agent with a doc retriever tool

Commands

  • /modify_bot This command allows you to customize the bot's personality, instructions, and temperature within the channel it's operating in. If !no-notify is included, then no notification is sent to the channel.

  • /bot_info This command presents the initial prompt used by the bot, as well as the default 'temperature' for generating responses.

  • /ask Use this command to ask questions or make requests. The bot employs the LLM to generate appropriate responses.

    Command syntax: /ask (<!all>) (<!temp=temp>) <question or request>

    Here, if !all is included, the bot sends its response to the entire channel. If !temp is included, the response's "temperature" (randomness of the bot's output) is adjusted.

  • /permissions (optional) This command modify which users can interact with the bot. It requires a password defined in the environment variables

    Command syntax: /permissions <PERMISSIONS_PASSWORD>

    If no password was defined inside .env, then this command do nothing.

  • /edit_docs This command allows the user to edit the descriptions of the documents that have been uploaded to the channel. These edited descriptions are used in the doc_retriever tool (See Mentions).

Mentions

  • When the bot is mentioned in a thread, it can respond based on the context. The context limit is handled using a max token limit in a similar way as ConversationTokenBufferMemory from langchain.

  • If the bot is mentioned in channel along with uploaded file, then it ask if you want to start a QA thread or upload the file directly to the channel: The user has the possibility to add some context and new separators to chunk the file(s). The files are downloaded in data/tmp to define a persistent VectorStore in data/db, after the generation of the VectorStore all files are deleted.

    • QA Thread: The bot responds to the user's message that contains the uploaded file(s), stating that a QA thread has been created with the uploaded file(s) and the context provided by the user.
      • If the user wants to remove the QA thread, use the flag !delete-qa while mentioning the bot.

    • Upload to channel: The file is upload to the channel and the tool doc_retriever appears in the list of tools once at least one file has been added to the channel. This tool take as context all the files uploaded by the users using this method. If the channel is used as a simple LLM chain, then a ConversationRetrievalChain, otherwise a tool to retrieve the important information from the documents is created and passed to the Agent.

How are documents handled?

The documents are handled using ChromaDB, saving the database to data/db/{channel_id}/{timestamp} for each QA thread, where channel_id refers to the channel which contains the thread and timestamp to the time when the QA thread was initiated. It is important to mention that typically embedding models are not compatible, so if you change the embedding model after creating the database for a given QA thread, then that thread will not work.

Usage

Requirements

To install the necessary requirements, use:

pip install -r requirements.txt

For CTransformers or OpenAI functionalities, you will need to install these packages separately:

pip install ctransformers
pip install openai

For Open-source embeddings, you will need to install sentence-transformers:

pip install sentence-transformers

Environment variables

Duplicate example.env to .env and adjust as necessary:

OPENAI_API_KEY= Your OpenAI key
CTRANSFORMERS_MODEL= model name from HuggingFace or model path in your computer
EMB_MODEL=all-MiniLM-L6-v2 # Embedding model
SLACK_BOT_TOKEN=xoxb-... Slack API Bot Token 
SLACK_APP_TOKEN=xapp-... Slack API App Token
PERMISSIONS_PASSWORD=CHANGEME # Password to activate /permissions command

Starting the Bot

To start the bot, simply run:

python main.py

This file contains the basic configuration to run the bot:

from src.slackbot import SlackBot
from src.handlers import create_handlers

# Set model_type
# OpenAI
# Llama (CTransformers)
# FakeLLM (just for testing) 
model_type='OpenAI'
bot = SlackBot(name='SlackBot', model_type=model_type)

# Set configuration
config = dict(model_name="gpt-3.5-turbo", temperature=0.8, max_tokens=500)

# Initialize LLM and embeddings
# max_tokens_threads refers to the max tokens to consider in a thread message history
bot.initialize_llm(max_tokens_threads=2000, config=config)

# If you don't want to use OpenAI Embeddings, you can add model_type parameter  to use a model from EMB_MODEL env variable
bot.initialize_embeddings() # model_type='llama' for HugginFaceEmbeddings

# Create handlers for commands /ask, /modify_bot, /bot_info  and bot mentions
create_handlers(bot)

### You can create new handlers for other commands as follow
# @bot.app.command("/foo")
# async def handle_foo(say, respond, ack, command):
#     await ack()
#     # do something..

Slack API configuration

The bot requires the following permissions:

  1. Enable Socket Mode

  2. Activate Incoming Webhooks

  3. Create Slash Commands

    • /ask Ask a question or make a request
    • /modify_bot Modify bot's configuration for the current channel
    • /bot_info Get prompt and temperature of the bot in the current channel
    • /permissions (optional) Modify which users can interact with the bot
    • /edit_docs Modify documents uploaded to the bot in the current channel

  4. Enable Events

    • Subscribe to app_mention

  5. Set Scopes

    • app_mention:read
    • channels:history
    • channels:join
    • channels:read
    • chat:write
    • files:read
    • im:write <- To notify users about change in permissions
    • users:read <- To get list of users

    Note that for groups you will require also groups:history, groups:join and groups:read

ToDo / Ideas

  • Add a command to modify which users can interact with the bot. The command should be initialized using a password, example /permissions <PASSWORD>
  • A ingest method to create a vector database and use a QA retriever
  • add a custom CallbackHandler to update the messages on the go
  • A modal view to modify files description
  • a method to remove files from the vectorstore
  • A way to delete unused QA threads
  • Create a doc retriever for each document, currently is using the same approach from privateGPT
  • Create tests