IMDB review filter

In this example, we filter the full Stanford IMDB dataset (100,000 reviews) with two questions. We keep only reviews where the model answers TRUE to both: does the review discuss the ending, and does the reviewer recommend the movie?

The full example is in demos/imdb_ending_filter.py.

The query

SQL = """
SELECT r.review_id
FROM reviews AS r
WHERE AI.IF(
    PROMPT(
        'Does this review discuss the ending of the movie?\\n\\n{0}',
        r.review
    ),
    {'selectivity': 0.25}
)
AND AI.IF(
    PROMPT(
        'Does the reviewer recommend watching the movie?\\n\\n{0}',
        r.review
    ),
    {'selectivity': 0.5}
)
"""

We chain two AI.IF filters with AND. Quail evaluates the cheaper filter first (lower selectivity means fewer reviews reach the second filter). Each review is read by the model once; the two questions are short suffixes on the review's KV.

Run the example

On a machine with an H100 GPU:

uv run python demos/imdb_ending_filter.py

You can change the device and GPU count:

uv run python demos/imdb_ending_filter.py --device h100-sxm --gpus 4

The script prints how many reviews matched, the selectivity at each stage, query time, throughput, and GPU cost.