Astronomy Catalog Hybrid Search

A real, published star catalog: the HYG Database v4.1 (a combined Hipparcos / Yale Bright Star / Gliese catalog). 22,493 real stars within ~100pc of the Sun, loaded as a hybrid table with a real photometric Vector(3) embedding (z-scored absolute magnitude + B-V color index + a bias term), searched for real solar-analog candidates.

Filtered/hybrid vector search

CREATE VECTOR INDEX ON stars(photo_vec)
SEARCH stars ON photo_vec QUERY [0.0, 0.0, 1.0] LIMIT 20 FILTER spect_class = "G"

FILTER post-filters the index-accelerated top-k results by an ordinary relational predicate: here, restricting an intrinsic-brightness/color similarity search to G-type stars, the Sun’s own spectral class.

The same query, expressed as a hybrid WHERE

SELECT proper_name, dist, absmag, ci
FROM stars
WHERE COSINE_SIM(photo_vec, [0.0, 0.0, 1.0]) > 0.999 AND spect_class = "G"

Both forms return the same 18 real solar-analog candidates: real stars in constellations like Tucana, Cetus, and Hydra whose absmag/ci sit within ~0.02 of the Sun’s own 4.85 / 0.656. EXPLAIN on the second form confirms CosineFilterExec fired: the planner decomposed the AND, index-accelerated the cosine conjunct, and applied spect_class = "G" as a cheap post-filter, not a full scan.

Every result cross-checked exactly (20/20) against a brute-force numpy cosine-similarity pass over the full real embedding array.

An honest methodology finding, not an engine bug

Raw [absmag, ci] makes COSINE_SIM misleading here: cosine similarity is angle/scale-invariant, so a star at exactly double the Sun’s absmag/ci scores a near-perfect match despite being a physically different star. Caught by inspecting the raw top-20 before trusting it; fixed by z-score-standardizing both features (verified against a Euclidean-distance brute force in numpy first, confirming 20/20 agreement, before ever touching the engine). Real data surfaces real modeling lessons, not just engine bugs.

Try it yourself

• View the full notebook
• Try it in the Playground