Background: User can upload an image to the bot. The bot should be able to reply with text from Gemini Pro model.

• Indicate that typing is in progress (bot status area)
• If user provides prompt, ask model with that prompt
• If user doesn't provide prompt, implement prompt generation as decided in #2

Background: User can record the message and send to the bot. The bot should be able to reply with the audio recording.

• Indicate that voice recording is in progress (bot status area)
• Transcribe audio from the user with Speech-To-Text API
• Send text request to the Gemini Pro model
• Generate audio based on the model response with Text-To-Speech API and send back to user

*Optional. Discuss with the team and decide if we want to moderate the input? I.e. if it's not related to city exploration, then bot can reply with "I apologize, but that's beyond my capabilities at the moment" (Check with tech writer for the best possible response)

• add tests for images with a few categories of artwork, e.g. paintings, sculptures (what else can be included?)
• add tests for images with city objects, e.g. the Colosseum (Rome), the Forbidden City (Beijing), etc
• discuss with the team and determine additional scenarios that can be included
• cover 5 different options per each category
• parametrise tests within each category

Think of adopting BDD test framework pytest-bdd and design tests as features, e.g.:

Feature: Vision Pro can detect artwork
    A model can correctly identify the artwork on the image

    Scenario: Can detect paintings
        Given The painting "nightwatch.jpg"
        When Gemini Pro Vision triggered
        Then The model response should contain painting name "The Night Watch"
        And The model response should contain author "Rembrandt"

    Scenario: Can detect sculptures
        Given The sculpture "kiss.jpg"
        When Gemini Pro Vision triggered
        Then The model response should contain sculpture name "The Kiss"
        And The model response should contain author "Rodin"

• Why did we choose Telegram (great platform, a lot of users, save efforts on UI development, etc)
• Why Google Cloud? (leaders in AI dev, excellent integration across the services in ecosystem, etc)
• Why these use cases?
• Target audience
• Future vision
• What else?

Background: Users are not required to provide any prompt when uploading an image. This raises an issue about how to best ask the model about the image without any context.

• Option 1: Use a generic prompt, e.g. "Tell me something about this thing"
• Option 2*: Use a more specific prompt that has knowledge of what is in the image, i.e. painting, sculpture or any other artwork

AC: Run some experiments and decide which option to choose.

• Implementation will be defined in Dev story

Background: The current prompt doesn't give different answers. Let's pick a city from the list of cities using python's random.choice and send in prompt to the model

• List can be hardcoded (use some 30 different famous cities)
• Or can be retrieved with some public API (check for possible options)

Background: CD is cool and we are looking to set it up with our Cloud Run service

• Follow GitHub Action template from test_idp
  • Update SERVICE env variable to city-bot
  • Specify as sources of interest: bot_service/*, Dockerfile

The message text should be less than 4096 characters.
Cut the model reply into several pieces and send as multiple messages.
More info: tdlib/telegram-bot-api#374

Background: For our new bot MyCityGuideBot telegram-specific metadata should be provided.

Following properties should be filled:

• About - will be shown on bot settings page
• Description -
• Description picture -
• Botpic - circle bot icon shown in telegram interface

Acceptance criterias:

• Choose 3-5 logo options (use existing pictures or generate something unique with the model, e.g. Kandinsky, Midjourney)
• Create Poll in the team chat to vote
• Create new static folder in the repo and add image file logo.jpg with the most votes
• Add metadata.txt file formatted in the following way:

About: <text>
Description: <text>
Description picture: <text>

Check telegram docs for more details on metadata

Implement new command /story that will give some fun fact about a random city:

• Support English and Russian
• The prompt language should be defined based on telegram profile settings

• Create new folder service_bot and put minimal code to reply with welcome phrase
• Register bot with God Father
• Add telegram secret to Secret Manager
• Deploy to Cloud Run
• Add to Team 25 group chat

Background: The model response generation can be customised with a set of various parameters

Explore them and decide on the best combination:

• temperature
• top-k
• top-p