digdir-assistants

A mono-repo for generative AI assistants, based on RAG architecture.

Currently deployed as a Slack application, we also plan to integrate with web applications, in particular Altinn Studio.

Agent flow

This diagram shows the main functional blocks and how data flows between them:

See it in action

#altinn-assistant on Altinn Devops Slack

Local development quickstart

1. Install dependencies:

$ yarn install

2. Build all packages and apps

$ yarn build

3. Run slack-app and admin ui

$ yarn run:slack-app

Note: in order for your local bot endpoint to receive traffic from Slack, you need to configure a proxy service such as ngrok, and configure a Slack app to use the URL allocated by ngrok.

4. Configure a new Slack app in your test workspace

It is recommended to create a few Slack workspace specifically for development and testing of Slack apps

Deploy to Azure Container app

1. Create a resouce group
2. Create a Container App Environment
3. Create a container repository
4. Create a Container App

For step by step instructions see: Create a Container App

Manual job execution

1. Update crawler index

2. Update search phrase index

$ cd cli $ yarn run:generateSearchPhrases <typesense collection name>

Example collection name: TEST_altinn-studio-docs-search-phrases`

System architecture

The Altinn Assistants architecture consists of the following technology components:

1. Front-end (Slack app, Admin web app)
2. Database-as-a-service (Supabase)
3. Serverless function runtime (Supabase Functions)
4. GPU-accelerated (Python app running in a GPU-enabled VM)
5. Text search index (Typesense Cloud)

Sequence diagram showing execution of a typical end user query:

sequenceDiagram
  participant slack.com
  participant slack-app
  participant llm-api
  participant supabase-db
  participant supabase-functions
  participant semantic-search
  participant gpu-services
  slack.com->>slack-app: new text message
  slack-app->>supabase-db: GET /config (cached)
  supabase-db->>slack-app: workspace + channel config
  slack-app->>llm-api: stage 1: analyze user query, translate if necessary
  llm-api->>slack-app: query analysis, translation
  slack-app->>llm-api: stage 2: query relaxation
  llm-api->>slack-app: generated search phrases
  slack-app->>semantic-search: hybrid search (cosine similarity + keyword)
  semantic-search->>slack-app: sorted doc list
  slack-app->>gpu-services: POST /colbert/rerank w/user query
  gpu-services->>slack-app: Re-ranked doc list
  slack-app->>llm-api: Prompt + retrieved context
  llm-api->>slack-app: Helpful response (streamed)
  slack-app->>slack.com: Helpful response (streamed response)
  slack-app->>slack.com: Finalized response w/metadata
  slack-app->>supabase-functions: Log query, response and metadata
  supabase-functions->>supabase-db: upsert team, channel, <br>user and message data
  slack-app->>llm-api: Translate to user's language
  llm-api->>slack-app: (streamed response)
  slack-app->>slack.com: (streamed response)
  slack-app->>slack.com: Finalized response w/metadata
  slack-app->>supabase-db: Log query, response and metadata

Loading