Skip to content

Architecture Overview

Patrizio is a lightweight Delta Chat bot written in pure Go (1.25+). The code is split into logical layers, each with a single responsibility. The whole application tries to follow the Hexagonal architecture, having ports and adapters, to keep elements isolated. We're also trying to avoid having functions with side effects, pushing towards pure function calls.

Flow Diagram

The code flow is basically the following:

flowchart TD
  A[User] -->|Delta Chat RPC| B[BotCli]
  B -->|OnNewMsg| C[handler.go]
  C -->|Domain logic| D[domain]
  D -->|Repository| E[sqlite/repository]
  E -->|SQLite DB| F[SQLite]
  C -->|AI completion| H[openai/client]
  H -->|OpenAI-compatible API| I[LLM Provider]
  C -->|Reply| G[Delta Chat RPC]

The deltabot-cli-go framework powers the bot. internal/bot/bot.go registers two callbacks:

  • OnBotInit - called once when the bot starts. It's mainly a hook for future extensions.
  • OnNewMsg - invoked for every message received. This is where the bot does its work.

The handler.go plays a central role in parsing the incoming message and then calling the right handler, that will process the message accordingly. Once the message has been processed, Delta Chat RPC is invoked to reply to the user accordingly.

About Delta Chat RPC

The component should have its own Repository, but for easiness of development we've decided to keep it inside the bot package. This will be addressed in the future.

Message dispatching

The functionalities served between 1-to-1 chats (DMs) and Groups are different.

Group chats

For groups the handler first looks for bot commands (/filter, /stop, /prompt, etc.). If the text is a command, it's parsed by the domain code. If the message is not a command, the handler checks whether it quotes a known conversation message - if so, it's treated as a thread continuation and dispatched to the AI flow. Otherwise, the message is normalised (lower-cased, punctuation removed) and the repository is queried for matching filters. Every matching filter triggers a reply: text, media, or a reaction, with media files fetched from the storage adapter.

Direct chats

In a one-to-one conversation the bot first checks for the /prompt command, then for thread continuations (replies to a previous AI message). If neither applies, it falls back to a short help message listing the available commands.