SQL AI functions

PostHog SQL includes AI functions that classify query rows and generate embeddings for semantic search.

Classifying rows with AI

The jev() and decide() SQL functions use a small AI model to answer a question about each row of a query. You can use them to tag support messages, sort feedback into categories, or flag events that match a description, without writing the rules yourself.

Pass the text to evaluate and a question. By default, the function returns the probability (from 0 to 1) that the answer is yes:

SELECT
    properties.$message AS message,
    jev(toString(properties.$message), 'Is this person asking for a refund?') AS refund_probability
FROM events
WHERE event = 'support message sent'
LIMIT 100

To pick one label from a list, pass choice. The result has the chosen label, the probability of each label, and a confidence score:

SELECT result.choice AS topic, count() AS feedback
FROM (
    SELECT jev(
        toString(properties.$message),
        'What is this feedback about?',
        choice := ['pricing', 'bugs', 'feature request', 'other']
    ) AS result
    FROM events
    WHERE event = 'feedback submitted'
    LIMIT 200
)
GROUP BY topic
ORDER BY feedback DESC

decide() works the same way and also takes a model argument, either 'jeeves' (the default, and what jev() uses) or 'jevk5'. For example, decide(message, 'Is this spam?', model := 'jevk5').

Where your data goes

PostHog builds and hosts both models. They run on PostHog's own servers in the same region as your project, so your data is not sent to a third-party AI provider such as OpenAI or Anthropic. If your project is in the EU, the evaluation also stays in the EU.

Jeeves, the default model, is open source. Its code, training data, and benchmark results are in the PostHog/jeeves repository.

Limits

• Use the function as a named column in a SELECT. To filter or sort on its result, wrap it in a subquery and filter in the outer query.
• Each query has two limits, and the error tells you the current value of whichever one you hit:
  • Rows: the number of rows a SELECT can send to an AI function. Add a LIMIT to stay under it.
  • Evaluations: the total across the whole query. Each AI column on each row counts as one, so two AI columns over 500 rows use 1,000 evaluations. A SELECT without a LIMIT reserves the full row limit for each AI column, so add a LIMIT when you use more than one.
• Each input can be up to 8 KiB of text. Convert other types with toString(). Empty (NULL) inputs are skipped.
• Identical inputs with the same question are only evaluated once per query.

Pricing

Each evaluation uses PostHog AI credits and appears under PostHog AI on your billing page. If your organization runs out of credits, the query stops with an error that says so.

Find similar text with embeddings

Use embedText() to convert text into an embedding, which is an array of numbers that represents its meaning. Pair it with cosineDistance() to find stored text with a similar meaning. A lower distance means a closer match.

The function has two arguments:

• text - The text to embed.
• model_name - The embedding model. Use text-embedding-3-small-1536 for 1,536 values or text-embedding-3-large-3072 for 3,072 values.

Both arguments must be string literals. You can't pass a column or another expression because PostHog generates the embedding when it compiles the query.

For example, this query finds documents that are similar to a natural-language description:

SELECT
    content,
    cosineDistance(
        embedding,
        embedText('users seeing errors during checkout', 'text-embedding-3-small-1536')
    ) AS distance
FROM document_embeddings
WHERE model_name = 'text-embedding-3-small-1536'
ORDER BY distance ASC
LIMIT 10

The model in embedText() must match the model for the stored embeddings. You can't compare embeddings from models that return different numbers of values.