Self-Hosting Sentiment and Emotion for DiscourseAI

[Falco]

The Discourse AI plugin has support for requesting emotion/sentiment classification of new posts, which is stored in the database and can be used in reports and admin dashboards.

Discourse AI supports two types of classification, each requiring its own model:

• Sentiment — classifies posts as positive, negative, or neutral (using cardiffnlp/twitter-roberta-base-sentiment-latest)
• Emotion — classifies posts across 28 emotion labels like joy, anger, surprise, etc. (using SamLowe/roberta-base-go_emotions)

To get both sentiment and emotion data in your dashboards, you need to run both models.

Running with HuggingFace TEI

HuggingFace provides an awesome container image that can get you running quickly.

Sentiment model

mkdir -p /opt/tei-sentiment-cache
docker run --rm --gpus all --shm-size 1g -p 8081:80 \
  -v /opt/tei-sentiment-cache:/data \
  ghcr.io/huggingface/text-embeddings-inference:latest \
  --model-id cardiffnlp/twitter-roberta-base-sentiment-latest

This should get you up and running with a local instance of cardiffnlp/twitter-roberta-base-sentiment-latest, an open model that can classify posts into positive/negative/neutral.

You can check if it’s working with

curl http://localhost:8081/ \
    -X POST \
    -H 'Content-Type: application/json' \
    -d "{ \"inputs\": \"I am happy\" }"

Which should return an array of confidence for each label under normal operation.

Emotion model

To also get emotion classification, run a second container with the emotion model:

mkdir -p /opt/tei-emotion-cache
docker run --rm --gpus all --shm-size 1g -p 8082:80 \
  -v /opt/tei-emotion-cache:/data \
  ghcr.io/huggingface/text-embeddings-inference:latest \
  --model-id SamLowe/roberta-base-go_emotions

Supported models

• cardiffnlp/twitter-roberta-base-sentiment-latest · Hugging Face — sentiment (positive/negative/neutral)
• SamLowe/roberta-base-go_emotions · Hugging Face — emotion (28 emotion labels)

Making it available for your Discourse instance

Most of the time, you will be running this on a dedicated server because of the GPU speed-up. When doing so, I recommend running a reverse proxy, doing TLS termination, and securing the endpoint so it can only be connected by your Discourse instance.

Configuring Discourse AI

Discourse AI includes site settings to configure the inference server for open-source models. You should point it to your server using the ai_sentiment_model_configs setting.

This setting accepts a JSON array of model configurations. Each entry requires:

Field	Description
model_name	The HuggingFace model ID (e.g. cardiffnlp/twitter-roberta-base-sentiment-latest)
endpoint	The URL of your TEI instance (e.g. https://your-server:8081)
api_key	API key for the endpoint (can be left blank if not required)

To get both sentiment and emotion dashboards, add an entry for each model you are running. For example, if you’re running both models locally:

• Entry 1: model_name cardiffnlp/twitter-roberta-base-sentiment-latest, endpoint https://your-server:8081
• Entry 2: model_name SamLowe/roberta-base-go_emotions, endpoint https://your-server:8082

After that, enable the classification by toggling ai_sentiment_enabled.

[haozi (耗子)]

Are there plans to support other models in languages ​​other than English?

[RBoy]

@Falco

Most of the time, you will be running this on a dedicated server because of the GPU speed-up

@Falco if one decides to run this on the same server running discourse (e.g. we have a very small deployment with a few thousand posts), could you update the instructions to outline

1. How can discourse integrate with a local instance of HuggingFace TEI container image
2. Suggestions on how much additional RAM/disk is required to run the above (e.g. if the base Discourse is running on 2GB RAM with 20GB disk)

[agent_kith]

So I have setup a new self hosted Discourse instance and trying to setup sentiments. This is my ai_sentiment_model_configs -

Key	Value
model name	cardiffnlp/twitter-roberta-base-sentiment-latest
endpoint	https://my_own_instance
api_key	[blank]

And it works, sort of, I get the sentiment bar graph.

However, the Emotion table is empty. This doc looks incomplete, or poorly worded for me to grasp what needs to be done.

Do I run another Docker container, with a different model ID (roberta-base-go_emotions?), or something else? What do I need to do get that emotion table filled up?

Would prefer to self host these services if possible. TIA if anybody can point me in the right direction.

[Falco]

@agent_kith

However, the Emotion table is empty. This doc looks incomplete, or poorly worded for me to grasp what needs to be done.

Do I run another Docker container, with a different model ID (roberta-base-go_emotions?), or something else? What do I need to do get that emotion table filled up?

For emotions you need to run the

@Falco

SamLowe/roberta-base-go_emotions · Hugging Face

too.

[agent_kith]

Thank you. So I just run a second docker container with some tweaks, like so:

mkdir -p /opt/tei-cache2
docker run --detach --rm --gpus all --shm-size 1g -p 8082:80 \
  -v /opt/tei-cache2:/data \
  ghcr.io/huggingface/text-embeddings-inference:latest \
  --model-id SamLowe/roberta-base-go_emotions

and add a new entry in ai_sentiment_model_configs and it’s all working now. Thank you.

[Lee_Ars]

Ran face-first into a brick wall of not knowing what the hell I’m doing in trying to get this working on an EC2 instance without a GPU. At least to the best of my extremely limited abilities and understanding here, trying to do this with a CPU-only instance means a lot more configuration work than I can understand. The cpu versions of ghcr.io/huggingface/text-embeddings-inference refuse to load either of the two listed models; Claude and GPT5 both tell me I need to convert them to ONNX models with a bunch of python tools, and that’s where I noped out.

(It’s also very possible I’m just dumb and missing an obvious step!)

[sam (Sam Saffron)]

One thing we are planning is allowing LLMs to act as sentiment analysis

That way you can plug in a cheap model like Gemini Flash 2.5 (or even nano) and use it for sentiment analysis.

[Neil_Evans2 (Neil Evans)]

@Falco

cardiffnlp/twitter-roberta-base-sentiment-latest

We are trying to use this feature with Azure AI Language (from our self-hosted Discourse instance) - as we are already using our Azure subscription to integrate GPT-4.5 with Discourse (for summarization and chat-bot functionality):

What is Azure AI Language - Azure AI services

• but we are getting no data in the the sentiment dashboard, and can see these errors in the logs:

Discourse AI: Errors during bulk classification: Failed to classify 208 posts (example ids: 2256, 909, 2270, 2260, 2797) : JSON::ParserError : An empty string is not a valid JSON string.

The backtrace shows that Discourse might be trying to use HuggingFace - are these the only models supported at the moment?

Thanks,

N

[sam (Sam Saffron)]

Yeah we have a very specific implementation, we are thinking about how to make it more generic

I am bullish on just letting LLMs return structured output for sentiment, cause it opens the doors to many options there and I feel that as a trend “hyper specific” APIs are being replaced by “hyper general” APIs LLMs provide.

[dylanb (Dylan Burkhardt)]

Is there any straightforward way to get this up and running on self-hosted WP without setting up your own instance for the sentiment classification?

[Neil_Evans2 (Neil Evans)]

Thanks Sam - that would be great. In the meantime, we will look into running one of the supported HuggingFace models on an Azure VM…