👾 Discord moderation bot with d1-ultra

Build a Discord bot that asks one question: is this message unsolicited solicitation?

It reads new messages in selected text channels of one server and detects sales, recruitment, or signup pitches. It saves a report in a private moderator channel, then deletes eligible text messages with a solicitation score of at least 0.80. Below 0.80, it takes no action.

Message → d1-ultra → below 0.80: no action
                  → 0.80 or above: archive first, then delete eligible text messages

Just want to try d1-ultra? Open the notebook. It walks through the API independently of Discord.

1. Create a virtual environment

You need Python 3.11+, a Liquid API key, and a Discord server where you have "Manage Server" permission.

Get the code and enter its folder:

git clone https://github.com/Liquid4All/d1-discord-mod-tutorial.git
cd d1-discord-mod-tutorial

Create a local .venv once, then activate it. This keeps the project's Python packages separate from your other projects.

macOS / Linux:

python3 -m venv .venv
source .venv/bin/activate

Windows PowerShell:

py -3 -m venv .venv
.venv\Scripts\Activate.ps1

Install the dependencies into the active environment:

python -m pip install -r requirements.txt

When you open a new terminal, return to this folder and run the activation command again. You only need to create the environment once. Run deactivate when finished.

Copy .env.example to a new file named .env for your keys and settings. .venv holds installed packages; .env holds configuration. Both are ignored by Git. Leave the example file empty.

2. Get your API key and bot token

Liquid: open Liquid Console, select your organization, then go to Dashboard → API Keys and create a key.

Discord: open the Developer Portal:

1. Create a New Application named d1-mod-tutorial.
2. Open Bot → Reset Token and copy the token.
3. Enable Message Content Intent on the Bot page and save.
4. Under Installation, enable Guild Install, select Discord Provided Link, and add the bot scope to the Guild Install defaults.
5. Select View Channels, Read Message History, Manage Messages, Send Messages, Embed Links, and Attach Files permissions.
6. Open the install link, choose Add to server, and select your test server.

3. Choose the channels to monitor

Choose one or more text channels to monitor, such as #general. Create a private #mod-review text channel in the same server. Give the bot View Channel, Read Message History, and Manage Messages in each monitored channel. In #mod-review, give it View Channel, Send Messages, Embed Links, and Attach Files. This channel holds archives for post-mortem analysis; keep it visible only to moderators and the bot.

In Discord, enable User Settings → Developer → Developer Mode. Right-click each channel and choose Copy Channel ID. Fill in your local .env:

LIQUID_API_KEY=your_liquid_key_here
DISCORD_BOT_TOKEN=your_discord_bot_token_here
DISCORD_CHANNEL_IDS=111111111111111111,222222222222222222
DISCORD_REVIEW_CHANNEL_ID=your_private_mod_review_channel_id

Replace the example numbers with your monitored channels' IDs, separated by commas. Use one numeric ID for the private review channel and keep it out of that list. All channels must be in the same server; the bot checks this at startup.

4. Understand the decision

classifier.py uses TypeSafeClient and Noul, following Liquid's Python SDK example:

from typesafe_sdk import Noul, TypeSafeClient

with TypeSafeClient(
    api_key=api_key,
    base_url="https://api.liquid.ai/decisions",
    compression=False,
) as client:
    result = client.system_one(
        model="d1-ultra",
        state=message,
        questions={"solicitation": Noul(instructions=INSTRUCTIONS)},
    )

probability = result.answers["solicitation"].noul

state is the message text. INSTRUCTIONS explains what counts as solicitation. Noul asks a yes/no question; .noul reads the probability of yes. The pinned SDK appends /v1/systemone, so the base URL includes /decisions to reach Liquid's documented endpoint. compression=False sends ordinary JSON. The bot also sets a timeout and disables retries to make one attempt per message.

bot.py uses one threshold from classifier.py:

DELETE_THRESHOLD = 0.80  # Archive in the private channel, then delete.

This threshold is a starting point. Feel free to adjust it based on your own data and moderation preferences. Test your chosen value on representative messages before enabling automatic deletion.

5. Run it

With your virtual environment active:

python bot.py

Wait for Ready: watching ...; reports go to #mod-review, listing your selected channels. Then post these invented examples in a monitored test channel from your own account, one at a time:

Message	Desired result
Buy my AI consulting package. DM me for pricing.	Archive and delete if score ≥ 0.80
Join my business opportunity. DM to earn money.	Archive and delete if score ≥ 0.80
I built an open-source chatbot. Feedback is welcome.	No action
How do I run this model locally?	No action
Watch out for “DM me to earn money” messages — it's a scam.	No action

Keep the program running while testing. Press Ctrl+C to stop. Restart after changing settings or code.

6. Scope

This repo is a minimal starting point for building your own Discord moderation bot with d1-ultra. It demonstrates one policy: detecting unsolicited solicitation, then archiving and deleting eligible text messages that meet your chosen threshold.

Possible next steps include:

• Continuous monitoring: deploy the bot on a cloud server and configure it to restart automatically after a crash or reboot.
• Observability and evaluation: store decisions, scores, errors, and moderator feedback in a database. Use those records to build evaluation datasets and refine the policy instructions and deletion threshold.
• Image moderation: extend the bot to check attached images with d1-ultra, which supports image inputs.
• Reliability and customization: add a message queue and retries for temporary API failures, or expand the bot with moderation policies tailored to your server.

Support

• 📖 Liquid AI Documentation
• 💬 Join our community on Discord