What the file holds.
vectors/vectors.parquet in the Hugging Face dataset taktekhq/lebanese-legislation has one row per passage of a law, article or ruling: an id, the document it comes from, the embedding model (gemini-embedding-001), a hash of the text and 768 numbers. It is 2.7 GB. You can query it directly with DuckDB or load it into any vector database.
Why we moved them.
The vectors and their index took about 7 GB of our 10 GB database. Moving them to Cloudflare Vectorize made the search database small enough to ship inside one container image, and Vectorize answers a nearest-passage query in a fraction of a second.
Kept in step.
A GitHub workflow exports the vectors, publishes the Parquet file and updates Vectorize. Later runs send only the passages that changed, so the open file and the live search always come from the same export.
How to use it
DuckDB can read the file straight from Hugging Face without downloading it first: select from read_parquet with the file's address, and filter by document or model. To search by meaning, embed your question with gemini-embedding-001 in 768 dimensions and compare it with the values column by cosine similarity. The file is free to reuse; the source texts keep their own rights and attribution.