Bug (?): False matches in full-text search

[do-me]

Searching for "aldi" it returns "ALDEA HOMES" for the first two results which seems wrong. All following results are correctly identifying actual "ALDI" matches though.

Running latest version with this:

embedding-atlas places_full_with_embeddings_1M.parquet --text primary_name --x embedding_x --y embedding_y

File is too large to share here so I put it in my cloud if you might need it: https://nextcloud.geo.rocks/s/LQyf6PCdinEpYQW

[donghaoren]

We use flexsearch with these options:

const options: IndexOptions = {
  tokenize: "forward",
  encoder: Charset.LatinBalance,
};

Perhaps the particular words are being encoded the same?

[do-me]

Seems like it, thanks for the pointer.
From a usability point of view, I'd expect "aldi" in quotes to search for the exact string while I'm ok with any other advanced token-based functions.
Do you think that exposing these options as user settings or via the cli would make sense?