Using AI Nodes

How to choose between Graze's AI nodes, set their thresholds, and get good results from them.

Keyword nodes are precise, but they only find the words you think of. AI nodes judge what a post is about, how it sounds or what its images show, so they catch posts your keywords would miss and drop ones that only look like a match.

You'll find them all on the AI tab of the Rules menu (the filter icon in the editor's left sidebar). For every option on every node, see the Editor Documentation.

Which AI node should I use?

There are three families.

Natural Language and Ask a Question let you describe what you want in your own words. Use them when your idea is specific ("posts asking for book recommendations", "job postings that aren't internships") and no fixed category fits. They're the most flexible, and the only ones with an hourly budget. They have their own guide: Natural Language and Ask a Question.

Classifier nodes score a post against a fixed list of categories. They're unmetered, so they're the right choice when one of their categories already says what you mean:

Content Moderation: offensive content, in categories such as sexual, hate, violence and harassment, plus OK. English only.
Toxicity Analysis: Toxic, Severe Toxicity, Obscene, Threat, Insult, Identity Hate.
Spam Analysis: Marketing Spam or Organic Content.
Topic Analysis: 19 broad topics, such as Sports, Music and Science & Technology.
Text Categorization: 25 categories, such as Politics, Animals and Programming.
Language Analysis: 20 languages, detected from the text itself.
Sentiment Analysis and Financial Sentiment Analysis: Positive, Negative or Neutral.
Emotion Analysis: 28 emotions, from Joy and Gratitude to Anger and Grief.
Image Moderation: NSFW or SFW.
Image Categorization: the same 25 categories as Text Categorization, applied to images.

Text Similarity compares a post with a sample of text you write, and scores how close they are in meaning. It sits between the other two: you write the input, but it's unmetered.

A good rule of thumb: try a classifier first if one of its categories fits, Text Similarity if you can show an example of what you want, and Natural Language when you need to describe it.

How thresholds work

Each AI node gives every post a score between 0% and 100%: how likely the model thinks it is that the post fits the category you picked. You choose a comparator and a threshold:

at least: keep posts scoring this or higher. Use it to bring posts in, such as Topic Analysis Sports at least 30%.
at most: keep posts scoring this or lower. Use it to keep posts out, such as Toxicity Analysis Toxic at most 50%.

The editor's own guide to thresholds:

30%: somewhat likely
70%: pretty likely
90%: very likely

A lower threshold catches more posts, including more mistakes. A higher one is stricter but misses some real matches. Most nodes start at 50%; Content Moderation and Language Analysis start at 90%.

Getting the right values

Choosing a threshold is a bit of an art. Models make mistakes, and the right number depends on your feed. The best approach:

Start from the editor's guide above.
Run a few posts you definitely want, and a few you definitely don't, through Debug. It shows the score each AI node gave.
Move the threshold so the good posts pass and the bad ones don't.

If moving a threshold seems to change nothing, try moving it a long way rather than a little. Scores often bunch up near 0% and near 100%, so small changes in the middle can select the same posts.

Combining AI with other nodes

AI nodes work best alongside simpler ones. Put a Word List or a Languages filter in the same All of these block to narrow things down, and let the AI node make the harder call. For example, a feed about football might need a Word List for "football" and a Topic Analysis Sports node, so posts about the word in other senses are dropped.

To exclude, use at most. Content Moderation sexual at most 20% keeps out posts that are even somewhat likely to be sexual.

Image Moderation and Image Categorization never match a post without images. If you use one to keep unsafe images out, put it in an Any of these block next to an Embed node set to missing Image, or text-only posts will be dropped too.

Text Categorization and Image Categorization use fixed lists

These two nodes choose from a fixed list of 25 categories: Academic & Intellectual, Adult & Sexual Content, AI & Machine Learning, Animals, Arts & Creative, Aviation & Maritime, Data & Computing, Entertainment & Culture, Film & TV, Food & Beverages, Food & Lifestyle, Game Development, Gaming, Healthcare & Medicine, Medical Education, Medical Specialties, Music, Nature & Outdoors, News & Media, Politics, Programming, Society & Culture, Software Development, Sports, Visual Arts.

Earlier versions let you type any label. That no longer works, and a node with a label that isn't on the list pulses and asks you to "choose a standard category". If none of the 25 fits your idea, use Natural Language instead: it accepts any description you can write.

Using Text Similarity

Text Similarity checks whether a post's text is close in meaning to a piece of text you supply. Write your sample in the box (up to about 250 tokens, roughly a short paragraph), choose the field to compare (usually Text), and set a threshold. New nodes start at at least 75%.

There's room for creativity in what you compare against.

Structure

You can match posts that look like your sample. With "I took my dog to the park", you might match:

I took my dog to the mall
I walked my cat at the park

Both look roughly like the text you provided, so both might score highly.

Content

You can also match posts about a subject, even when they don't use your words. A Word List for "dog" won't find a post that only talks about yorkies, but a Text Similarity node with a sample about dogs may, because the model knows a yorkie is a dog.

Tips

Write your sample the way posts are written. A short, natural sentence usually works better than a list of keywords.
Expect some false matches. A post saying "I got that dog in me" might score well for a dog feed. Add a Word List set to missing to drop the phrases you don't want.
If you need several different samples, use several Text Similarity nodes inside an Any of these block.

Budgets

Classifier nodes and Text Similarity aren't metered. Natural Language and Ask a Question are: each feed has an hourly budget for them. See Rate Limits for how budgets work and what happens when one runs out.