Omna gives Polars semantic search and PII masking — in one line of Python.
Local-first. Rust-powered. Zero data egress.
Search your DataFrames by meaning, not strings. Mask sensitive columns before they ever reach a model.
We're building the universal semantic layer between enterprise data and AI. We started with Polars because that's where the fastest-growing data engineering community is. Our roadmap takes us to every format and every model.
# Before — regex hellimport re, polars as pldf = pl.read_parquet("patients.parquet") PAIN = re.compile(r"chest|cardiac|angina|heart\s+pain", re.I)hits = []for row in df.iter_rows(named=True): if PAIN.search(row["notes"] or ""): hits.append(row)# ...still missing 'shortness of breath',# 'tightness', 'pressure'... etc. forever."They cannot see your data until you show it to them in exactly the right way."
"Omna is to AI models what a good search engine is to websites."
"Google didn't kill websites — it made them more findable, and therefore more valuable. Omna doesn't replace AI. It makes AI usable on data that was previously unreachable — and therefore drives more AI usage."
"AI is smart but blind. Your data is rich but invisible to AI. Omna is the bridge."
Performance-critical similarity engine written in Rust · local-first by design.
No signup. No data leaves your machine. Real Polars syntax, real output, runs entirely in your browser.
1import polars as pl2import omna # registers .omna namespace34df = pl.read_parquet("healthcare.parquet")56(df.omna7 .search("heart pain", top_k=3)8 .collect())Every tile below is live — the same Rust kernel that runs in your notebook is rendering this page.
df.omna.mask_pii() — one line, audit-logged.› df.omna.filter("kids under 12 with chest pain")try it in the playground →› df.omna.ask("flag anyone at risk of CHF")try it in the playground →Omna is built natively on Polars — which means it inherits Arrow's columnar memory format, SIMD vectorization, and zero-copy data structures.
Your embeddings never leave the Arrow memory layout. No serialization. No copying. No overhead.
When you call df.omna.search(), the Rust similarity kernel operates directly on the same memory Polars is already using.
"Polars uses Apache Arrow's columnar memory format with SIMD vectorization — the same memory our Rust similarity kernel operates on directly. On Pandas we'd need to copy data into NumPy arrays first. On Polars it's zero-copy end to end."
Ritchie Vink's original post on the Polars architecture explains it better than we can. Read it →
Anonymized while teams are still in private beta. Names dropped on request.
Replaced 400 lines of regex with one method call. The audit team stopped opening tickets.
The first tool that didn't ask me to ship our patient records to a vendor for an embedding.
`.understand()` caught a PHI column three audits had missed. We bought the cloud tier the same week.
We're building a complete semantic layer for regulated data on Polars. Here's the path.
Local-first vector search across Polars columns. ~12ms on 1M rows.
Built-in detectors for SSN, PHI, payment data, plus custom regex graphs.
Fuzzy joins on meaning, not strings. df_a.omna.join(df_b, on='intent').
Optional cross-encoder pass entirely in Rust. No Python in the hot loop.
Query across machines without moving the data. SOC2-aligned audit log.
Hosted control plane for teams. Local kernel stays the source of truth.
Four years writing brittle regex pipelines to scrub patient records before they could touch a model. Every new column type meant a new edge case, a new on-call page, a new audit finding.
Omna is the tool that should have existed: a Polars-native API where .search() and .mask_pii() are first-class operators, the kernel is Rust, and your data never leaves the machine. If you've felt this pain, we'd like to hear from you.
Help build the kernel that handles a billion sensitive rows without breaking a sweat. Founding team equity, async-first, strong opinions weakly held.
Bug reports, feature requests, or "this would be 10× more useful if…" — we read every one.