DeepPavlov Dream

DeepPavlov Dream is a platform for creating multi-skill generative AI assistants.

To learn more about the platform and how to build AI assistants with it, please visit Dream. If you want to learn more about DeepPavlov Agent that powers Dream visit DeepPavlov Agent documentation.

Star History

Distributions

We've already included six distributions: four of them are based on lightweight Deepy socialbot, one is a full-sized Dream chatbot (based on Alexa Prize Challenge version) in English and a Dream chatbot in Russian.

Deepy Base

Base version of Lunar assistant. Deepy Base contains Spelling Preprocessing annotator, template-based Harvesters Maintenance Skill, and AIML-based open-domain Program-y Skill based on Dialog Flow Framework.

Deepy Advanced

Advanced version of Lunar assistant. Deepy Advanced contains Spelling Preprocessing, Sentence Segmentation, Entity Linking and Intent Catcher annotators, Harvesters Maintenance GoBot Skill for goal-oriented responses, and AIML-based open-domain Program-y Skill based on Dialog Flow Framework.

Deepy FAQ

FAQ version of Lunar assistant. Deepy FAQ contains Spelling Preprocessing annotator, template-based Frequently Asked Questions Skill, and AIML-based open-domain Program-y Skill based on Dialog Flow Framework.

Deepy GoBot

Goal-oriented version of Lunar assistant. Deepy GoBot Base contains Spelling Preprocessing annotator, Harvesters Maintenance GoBot Skill for goal-oriented responses, and AIML-based open-domain Program-y Skill based on Dialog Flow Framework.

Dream

Full version of DeepPavlov Dream Socialbot. This is almost the same version of the DREAM socialbot as at the end of Alexa Prize Challenge 4. Some API services are replaced with trainable models. Some services (e.g., News Annotator, Game Skill, Weather Skill) require private keys for underlying APIs, most of them can be obtained for free. If you want to use these services in local deployments, add your keys to the environmental variables (e.g., ./.env, ./.env_ru). This version of Dream Socialbot consumes a lot of resources because of its modular architecture and original goals (participation in Alexa Prize Challenge). We provide a demo of Dream Socialbot on our website.

Dream Mini

Mini version of DeepPavlov Dream Socialbot. This is a generative-based socialbot that uses English DialoGPT model to generate most of the responses. It also contains intent catcher and responder components to cover special user requests. Link to the distribution.

Dream Russian

Russian version of DeepPavlov Dream Socialbot. This is a generative-based socialbot that uses Russian DialoGPT by DeepPavlov to generate most of the responses. It also contains intent catcher and responder components to cover special user requests. Link to the distribution.

Prompted Dream Distributions

Mini version of DeepPavlov Dream Socialbot with the use of prompt-based generative models. This is a generative-based socialbot that uses large language models to generate most of the responses. You can upload your own prompts (json files) to common/prompts, add prompt names to PROMPTS_TO_CONSIDER (comma-separated), and the provided information will be used in LLM-powered reply generation as a prompt. Link to the distribution.

Quick Start

System Requirements

• Operating System: Ubuntu 18.04+, Windows 10+ (через WSL & WSL2), MacOS Big Sur;
• Version of docker from 20 and above;
• Version of docker-compose v1.29.2;
• Operative Memory from 2 Gb (using proxy), from 4 Gb (LLM-based prompted distributions) and from 20 Gb (old scripted distributions).

Clone the repo

git clone https://github.com/deeppavlov/dream.git

Install docker and docker-compose

If you get a "Permission denied" error running docker-compose, make sure to configure your docker user correctly.

Run one of the Dream distributions

Deepy Base

docker-compose -f docker-compose.yml -f assistant_dists/deepy_base/docker-compose.override.yml up --build

Deepy Advanced

docker-compose -f docker-compose.yml -f assistant_dists/deepy_adv/docker-compose.override.yml up --build

Deepy FAQ

docker-compose -f docker-compose.yml -f assistant_dists/deepy_faq/docker-compose.override.yml up --build

Deepy GoBot

docker-compose -f docker-compose.yml -f assistant_dists/deepy_gobot_base/docker-compose.override.yml up --build

Dream (via proxy)

The easiest way to try out Dream is to deploy it via proxy. All the requests will be redirected to DeepPavlov API, so you don't have to use any local resources. See proxy usage for details.

docker-compose -f docker-compose.yml -f assistant_dists/dream/docker-compose.override.yml -f assistant_dists/dream/dev.yml -f assistant_dists/dream/proxy.yml up --build

Dream (locally)

Please note, that DeepPavlov Dream components require a lot of resources. Refer to the components section to see estimated requirements.

docker-compose -f docker-compose.yml -f assistant_dists/dream/docker-compose.override.yml -f assistant_dists/dream/dev.yml up --build

We've also included a config with GPU allocations for multi-GPU environments:

AGENT_PORT=4242 docker-compose -f docker-compose.yml -f assistant_dists/dream/docker-compose.override.yml -f assistant_dists/dream/dev.yml -f assistant_dists/dream/test.yml up

When you need to restart particular docker container without re-building (make sure mapping in assistant_dists/dream/dev.yml is correct):

AGENT_PORT=4242 docker-compose -f docker-compose.yml -f assistant_dists/dream/docker-compose.override.yml -f assistant_dists/dream/dev.yml restart container-name

Prompted Dream

docker-compose -f docker-compose.yml -f assistant_dists/dream_persona_prompted/docker-compose.override.yml -f assistant_dists/dream_persona_prompted/dev.yml -f assistant_dists/dream_persona_prompted/proxy.yml up --build

We've also included a config with GPU allocations for multi-GPU environments.

Let's chat

DeepPavlov Agent provides several options for interaction: a command line interface, an HTTP API, and a Telegram bot

CLI

In a separate terminal tab run:

docker-compose exec agent python -m deeppavlov_agent.run agent.channel=cmd agent.pipeline_config=assistant_dists/dream/pipeline_conf.json

Enter your username and have a chat with Dream!

HTTP API

Once you've started the bot, DeepPavlov's Agent API will run on http://localhost:4242. You can learn about the API from the DeepPavlov Agent Docs.

A basic chat interface will be available at http://localhost:4242/chat.

Telegram Bot

Currently, Telegram bot is deployed instead of HTTP API. Edit agent command definition inside docker-compose.override.yml config:

agent:
  command: sh -c 'bin/wait && python -m deeppavlov_agent.run agent.channel=telegram agent.telegram_token=<TELEGRAM_BOT_TOKEN> agent.pipeline_config=assistant_dists/dream/pipeline_conf.json'

NOTE: treat your Telegram token as a secret and do not commit it to public repositories!

Configuration and proxy usage

Dream uses several docker-compose configuration files:

./docker-compose.yml is the core config which includes containers for DeepPavlov Agent and mongo database;

./assistant_dists/*/docker-compose.override.yml lists all components for the distribution;

./assistant_dists/dream/dev.yml includes volume bindings for easier Dream debugging;

./assistant_dists/dream/proxy.yml is a list of proxied containers.

If your deployment resources are limited, you can replace containers with their proxied copies hosted by DeepPavlov. To do this, override those container definitions inside proxy.yml, e.g.:

convers-evaluator-annotator:
  command: ["nginx", "-g", "daemon off;"]
  build:
    context: dp/proxy/
    dockerfile: Dockerfile
  environment:
    - PROXY_PASS=proxy.deeppavlov.ai:8004
    - SERVICE_PORT=8004

and include this config in your deployment command:

docker-compose -f docker-compose.yml -f assistant_dists/dream/docker-compose.override.yml -f assistant_dists/dream/dev.yml -f assistant_dists/dream/proxy.yml up --build

By default, proxy.yml contains all available proxy definitions.

Components English Version

Dream Architecture is presented in the following image:

| Name | Requirements | Description | |---------------------|--------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | Rule Based Selector | | Algorithm that selects list of skills to generate candidate responses to t