How do I do semantic search on a Polars DataFrame?

Install Omna with pip install omna, then call df.omna.search('your query', top_k=3) on any Polars DataFrame. Omna embeds the rows locally with a Rust kernel and returns the top-k most semantically similar rows in milliseconds — no network calls, no API keys.

Is Omna faster than Pinecone or Weaviate for Polars?

For DataFrame-scale workloads (up to ~10M rows on a laptop), yes. Omna runs in-process with a Rust kernel — there is no network round-trip, no container to spin up, and no index to manage. Pinecone and Weaviate are designed for distributed, multi-tenant production search and are overkill for analytical Polars pipelines.

Does my data leave my machine?

No. Omna is local-first: embeddings, indexing, and search all happen in your Python process. There are zero network calls. This makes it HIPAA-compatible by design — no BAA needed, because no records ever leave the host.

Polars · Semantic Search

Polars Semantic Search in One Line of Python

Omna gives Polars DataFrames native semantic (vector) search. Local-first, Rust-powered, zero network calls. Search by meaning, not by string match.

The one-line API

import polars as pl
import omna  # registers the .omna namespace

df = pl.read_parquet("clinical_notes.parquet")

# Semantic search — returns the top-3 rows most similar
# to the query, with a relevance score column appended.
results = df.omna.search("chest pain radiating to left arm", top_k=3)

Why semantic search on Polars?

Polars is the fastest DataFrame library in Python. But until now, searching text columns meant brittle regex, str.contains, or shipping your data to Pinecone, Weaviate, or pgvector. Omna closes that gap: vector search becomes a method on the DataFrame itself, indexed and queried in-process with a Rust kernel.

Benchmarks

· ~12ms p50 latency on 1M rows (M2 MacBook, MiniLM embeddings)
· 0 network calls — runs in your Python process
· HIPAA-compatible by design — no egress, no BAA required

vs Pinecone, Weaviate, pgvector

Pinecone, Weaviate, Qdrant, and pgvector are excellent for production, multi-tenant, low-latency search-as-a-service. For analytical Polars pipelines — notebooks, ETL jobs, ad-hoc exploration — they are operationally heavy: containers, indexes, network hops, vendor approvals. Omna trades that infrastructure for an in-process Rust kernel that runs wherever Polars runs.

Try it now

The live playground on omna.dev runs a real Polars semantic search on a sample healthcare dataset, in your browser.

pip install omna