Skip to content

Home

Jump to bottom
Daniel Kornev edited this page Apr 21, 2024 · 30 revisions

DeepPavlov Dream Wiki

Welcome to the DeepPavlov Dream wiki. This is the DeepPavlov Dream user manual and documentation. It is a wiki (like Wikipedia), which means you can edit it, add new material, and so on. With your help we can make DeepPavlov Dream a very well documented system. Please help.

Documentation

Development Version (Main Branch)

Core Technologies

  • Agent orchestrates all components. It passes user input through the given pipeline.
  • Dialog State is a dictionary with the full information about the dialogue, user and bot (bot persona corresponding to the conducted dialogues with the particular user) profiles.
  • Connectors is a mechanism that allows integrating custom components into the Agent's pipeline.

Customizable Components

  • Formatters are the functions that allow converting the input and output API of services into Agent's API. These functions are written in Python.
  • Skills are the components providing response candidates (one, zero or several) for the given context.
  • Annotators are the components fopr Natural Language Understanding (NLU), information extraction, requests for remote APIs.
  • Skill Selectors are the components that choose a set of skills to be called for the current step to generate response candidates for the given context. Skill selection is used to decrease resource consumption and manage the dialogu development triggering not all the available skills. A pipeline uses one Skill Selector.
  • Candidate Annotators are the components for Natural Language Understanding (NLU) of response candidates.
  • Response Selectors are the components that compose a final system response out of available response candidates based on the dialogue context and candidates annotations. A pipeline uses one Response Selector.
  • Postprocessors are the components for Natural Language Understanding (NLU) of the final system response. Postprocessors are needed because the final response can be composed from several candidate responses, and therefore, the final response should be annotated again because we do not have a mechanism for merging annotations of the candidates responses.

Resources

Project Information

Community

Resources

Clone this wiki locally