WikiA is Answered

An engaging virtual assistant service for answering (almost) any question about some event or character from the Fandom/WikiA knowledge base

Architecture

• designed for the Google Cloud Platform
• Python server, running on App Engine Flexible Environment
• conversational agent running on Dialogflow
• Redis database running on Cloud Memorystore
• Custom Search engine for retrieving articles from Fandom

Dialogflow agent

• welcome and fallback intents (as per default)
• identity intent (e.g. 'To whom am I speaking?'')
• description intent (e.g. 'Please describe yourself')
• workflow intent (e.g. 'What may I ask about?')
• dynamic knowledge base (new documents added as necessary)

Workflow

• the user is greeted with an introductory message from the Dialogflow agent
• the user may submit a message (usually a question) using the input area
• the server sends the question to the Dialogflow agent and obtains the initial list answers (excluding the knowledge base)
• if there are no relevant answers in the list, the query is searched in the database; if found, its associated answer is shown to the user
• otherwise the query is sent to the Custom Search engine; a list of articles URLs is thus obtained
• each article URL is checked for existence in the database and, if necessary, scraped and stored in plain text in the agent's knowledge base
• the input query is then sent once again to the agent and the knowledge base answers are retrieved
• if available, the most relevant answer is shown; otherwise the fallback message is displayed
• this process can be repeated indefinitely

Setup

Requirements

• Dialogflow agent restored from knowledge-agent.zip (see releases):
  • the Fandom knowledge base ('https://{0}.fandom.com/wiki/{1}') - enabled, identified by Fandom KB ID (the part after '.../editKnowledgeBase/');
• <project root>/config.yaml with the following:

google_api:
  key: <API key>   # for Custom Search

custom_search:
  cx: <Custom Search ID>

dialogflow:
  fandom: <Fandom KB ID>

redis:
  - host: <host>
    port: <port>      # optional (default: 6379)
    auth: <password>  # optional
  - ...

• (optional) <project root>/account.json with valid GCP service account data.

The Redis credentials are tried sequentially until the first successful database connection.

Notes

There's a significant delay during answer retrieval - caused by latencies that occur during scraping and uploading new documents into the knowledge base. There's room for improvement in that area.

MIT License