omna 0.2.0

Semantic search, PII masking, and schema understanding for Polars DataFrames

Omna

Hybrid search (semantic + keyword), enterprise-grade PII detection & masking, and schema understanding — directly on your Polars DataFrames. No vector database. No API key. Data never leaves your machine.

---

The problem

# Finding every insurance claim denial — painful
keywords = ["claim denied", "coverage rejected", "policy voided", ...]
pattern  = re.compile("|".join(keywords), re.IGNORECASE)
results  = df[df["text"].str.contains(pattern, na=False)]
# Still misses: "insurer refused to honour the policy"
# Still misses: "claim outcome: not payable"
# Still misses: medical claim rejections using clinical terminology
# ...50+ lines per task. Grows with every edge case. Still wrong.

# With Omna
results = df.omna.search("insurance claim denied", on="text", k=5)
# Finds ALL of them — including docs that never say "denied" literally.
# 9ms. 50,000 documents. Zero cloud.

filtered = df.omna.filter("insurance claim denied", on="text", threshold=0.73)
# Every semantically matching document above the threshold.
# No keyword lists. No guesswork. Pure meaning.

answer = results.omna.ask("What personal data do these documents expose?")
# → "These insurance documents expose SSNs, medical record numbers,
#    dates of birth, health plan numbers, and claimant identifiers."
# Instant. One line.

---

# Auditing for PII before the data ships — painful
for col in df.columns:
    for i, val in enumerate(df[col].to_list()):
        if re.search(r'\b\d{3}-\d{2}-\d{4}\b', str(val)):   # SSNs only
            print(f"row {i}, {col}: {str(val)[:60]}")
# Catches one pattern. Misses emails, phones, names, IBANs.
# No confidence score. No audit trail. No redaction.

# With Omna
df.omna.pii_report()          # audit — find every leak, every column
df.omna.mask_pii()            # redact — one line, full audit log
df.omna.mask_pii(model=True)  # + on-device AI model for contextual PII
# A six-layer engine: regex + checksum-validated IDs + 220+ secret rules +
# an on-device AI model — the SAME Rust engine as the Omna Mac app & extension.
# Reversible [PERSON_1] tokens; secrets always irreversibly redacted. Local.
# Benchmarked openly on Gretel (core-PII recall 0.84 with the model, up from
# 0.69 on the prior engine): see docs/benchmark.md.

---

Demo

The Sword — hybrid search (semantic + keyword), filter, and ask across 50,000 documents:

The Shield — PII audit and redaction in one line:

Dataset: Gretel PII Benchmark (acquired by NVIDIA) — 50,000 synthetic documents built to test data privacy tools.

---

Enterprise-grade PII masking

Most "PII for DataFrames" tools are a regex or a Presidio wrapper. Omna's masking is a six-layer detection engine (pure Rust, fully on-device) — the same engine that powers the Omna Mac app and browser extension:

• L1 — patterns + checksum validators: emails, SSNs, cards (Luhn), IBANs, and 30+ international IDs verified, not just pattern-matched
• L2 — secrets: 220+ rules (AWS keys, GitHub tokens, JWTs, private keys) with entropy checks
• L3 — on-device AI model (mask_pii(model=True)): catches contextual PII no regex can — bare names, addresses, medical context
• L4–L6: entity resolution, reversible [PERSON_1] tokens (secrets always irreversibly redacted), and a full audit trail

No cloud, no API key, no data leaves your machine.

Same Gretel benchmark, same scoring — only the engine changed:

Gretel benchmark	Before (Presidio)	After (Omna engine)
Core-PII recall	0.69	0.84
All-types recall	0.35	0.79
All-types F1	0.50	0.82
Types · secret rules · validated IDs	~17 · 0 · none	30+ · 220+ · 30+

df.omna.pii_report()          # audit — every PII column, with confidence
df.omna.mask_pii()            # redact (L1+L2) — instant, full audit log
df.omna.mask_pii(model=True)  # full L1–L6, model-grade

Full methodology + numbers: docs/benchmark.md.

---

Install

pip install "omna[all]"

Supported platforms: Python 3.10–3.12 on macOS or Linux. Windows is not supported. The PII engine (omna[pii], and mask_pii(model=True)) runs on Apple Silicon macOS and Linux — it is not available on Intel Macs, because its on-device model runtime (ONNX Runtime) ships no Intel-Mac build; Intel-Mac users can still use search, filter, and schema understanding.

Extras: omna[embed] (search/filter), omna[pii] (PII detection & masking), omna[ask] (LLM queries). A bare pip install omna (just polars + numpy + rich) gives schema understanding (understand_df()); the heavier features are opt-in. PII masking is powered by the compiled omna-pii-mask engine — a self-contained wheel with no heavy Python ML dependencies, pulled in automatically by omna[pii], omna[ask] (which masks rows before the API call), or omna[all]. No API key needed for search, filter, embed, pii_report, mask_pii, or understand. Only ask() requires ANTHROPIC_API_KEY.

---

Quick start

import polars as pl
import omna

df = pl.read_csv("documents.csv")

# 1 — explore the schema
omna.understand_df(df)

# 2 — audit for PII before anything touches the data
df.omna.pii_report()

# 3 — redact
clean = df.omna.mask_pii()

# 4 — build a search index once
clean.omna.embed("text")

# 5 — search by meaning
results = clean.omna.search("insurance claim denied", on="text", k=5)

# 6 — filter everything above a threshold
flagged = clean.omna.filter("insurance claim denied", on="text", threshold=0.73)

# 7 — ask a question in plain English
results.omna.ask("What personal data do these documents expose?")

---

What Omna does

Method	What it does
omna.understand_df(df)	Schema inference — labels, null rates, samples. No LLM.
df.omna.embed(column)	Vectorize a text column once; reuse across sessions
df.omna.search(query, on, k)	Top-k results by hybrid relevance (semantic + keyword)
df.omna.filter(query, on, threshold)	Every row above a similarity threshold
df.omna.pii_report()	Audit every string column for PII
df.omna.mask_pii()	Redact PII, auto-save audit log
df.omna.ask(question)	Natural language queries over your DataFrame

---

API reference

omna.understand_df(df) — explore before you do anything

No LLM. No API call. Analyzes column names, dtypes, null rates, and sample values.

omna.understand_df(df)

 column                dtype    null_pct   label     sample
 uid                   String     0.0%     category  24bb757...
 domain                String     0.0%     category  insurance, healthcare...
 document_type         String     0.0%     category  Invoice, ClaimForm...
 document_description  String     0.0%     text      An insurance claim...
 text                  String     0.0%     text      **Claim ID: 285-14...

Labels: email phone name id date text numeric boolean category unknown

df.omna.embed(column) — vectorize once, search forever

Converts text to 768-dimensional vectors using FastEmbed (local ONNX, no API key). Saves to .omna/{column}.parquet. Run once — search() and filter() load it automatically on every subsequent call.

df.omna.embed("text")
# → .omna/text.parquet

Model: nomic-ai/nomic-embed-text-v1.5 (768-dim, downloaded once on first use) — the same embedding model the Omna Mac app uses. Embed is a one-time cost.

Hardware	50k rows
MacBook Air M5	~45 min
MacBook Pro M4 Max	~15 min
AWS GPU instance	~2 min

df.omna.search(query, on, k) — hybrid search (semantic + keyword)

Requires df.omna.embed("column") first.

results = df.omna.search("insurance claim denied", on="text", k=5)

Search is hybrid by default: semantic (embedding) similarity and BM25 keyword matching run together, fused with Reciprocal Rank Fusion. Semantics catch meaning; BM25 catches rare exact tokens the embeddings blur — part codes, IDs, surnames, acronyms. Results are ordered by fused relevance. The BM25 index is built from the saved column on first use (no re-embedding, no change to your index file). Pass hybrid=False for pure-semantic search.

results = df.omna.search("XJ9000", on="parts", k=5)              # exact code → BM25 nails it
results = df.omna.search("claim denied", on="text", hybrid=False) # meaning only

 uid            document_type         domain      text                               _score
 67fccc1e207…   ClaimSummary          insurance   **Claim ID: 285-14-1755, Policy…   0.762
 b8ae088cd21…   ClaimSummary          insurance   **Claim Summary**…                 0.749
 de5bba0a2cc…   Insurance Claim Form  healthcare  **Insurance Claim Form**…          0.748
 ebccdde3b42…   Insurance Claim       healthcare  Insurance Claim for MED74974358…   0.747
 aebb0eb55fb…   ClaimForm             healthcare  **Claim Form** - Patient ID…       0.747

_score is cosine similarity (0–1). None of these documents contain the phrase "insurance claim denied" — Omna finds them by meaning.

df.omna.filter(query, on, threshold) — semantic filter

Requires df.omna.embed("column") first.

filtered = df.omna.filter("insurance claim denied", on="text", threshold=0.73)
# → N documents matched — all semantically related to claim denials

Returns every row above the threshold. Default: 0.3. Raise for precision, lower for recall.

Use search() for the top k. Use filter() for everything above a threshold.

df.omna.pii_report() — audit before you redact

df.omna.pii_report()

 column    detected types                                    hit rate   flagged
 entities  CREDIT_CARD, EMAIL_ADDRESS, PERSON, PHONE_NUMBER   85.4%    ✓ YES
 text      CREDIT_CARD, EMAIL_ADDRESS, PERSON, PHONE_NUMBER   78.1%    ✓ YES

Scans every string column. Returns hit rates, PII types, and confidence scores. Nothing is modified.

df.omna.mask_pii() — redact in one line

clean = df.omna.mask_pii()
# → Omna's own six-layer Rust engine (same kernel as the Mac app + browser
#   extension): reversible [PERSON_1]-style tokens; credentials/secrets
#   always irreversibly [REDACTED:KIND]; checksum-validated IDs; 220+ secret
#   rules. No heavy Python ML dependencies.
# → audit log saved to .omna/pii_audit.parquet automatically

clean = df.omna.mask_pii(model=True)
# → adds L3, the on-device AI model, for contextual PII regex can't catch
#   (bare prose names, addresses). Downloads the model (~809 MB) once.
# Requires the omna_pii_mask wheel (built from omna-workspace); see CHANGELOG.md.

Detects: PERSON EMAIL PHONE CREDIT_CARD US_SSN IP_ADDRESS IBAN MEDICAL_RECORD_NUMBER BANK_ACCOUNT, 220+ secret types, and 30+ international IDs — checksum-validated where applicable.

# Add the on-device AI layer (L3) for contextual PII regex can't catch —
# bare names, addresses, medical context. Downloads the model once.
clean = df.omna.mask_pii()              # L1+L2 (instant, deterministic)

df.omna.ask(question) — natural language queries

Privacy (since 2026-06-10): the sampled rows in the prompt are masked before they leave your machine (PII replaced with tokens; secrets redacted). Pass mask_rows=False to send raw rows for synthetic/public data.

Sends schema + up to 20 sample rows to Claude. Requires ANTHROPIC_API_KEY.

export ANTHROPIC_API_KEY=sk-ant-...

results.omna.ask("What personal data do these documents expose?")
# → "These insurance documents expose SSNs, medical record numbers,
#    dates of birth, health plan numbers, and claimant identifiers."

# Override model
results.omna.ask("Summarise the key themes", model="claude-sonnet-4-6")

Default model: claude-haiku-4-5-20251001.

---

How it works

df.omna.search("insurance claim denied", on="text", k=5)
         │
         ▼
   embedder.py       FastEmbed — nomic-embed-text-v1.5, local ONNX
                     query → [0.12, -0.34, 0.87, ...]  768-dim vector
         │
         ▼
   index.py          loads .omna/text.parquet → Arrow memory, zero-copy
                     50,000 stored vectors in Polars' own allocation
         │
         ▼
   similarity.rs     Rust kernel — cosine similarity over all vectors
                     returns top-k sorted descending, no Python loop
         │
         ▼
   frame.py          slices result rows, attaches _score → pl.DataFrame

The Rust kernel is under 70 lines. Dot products and norms in machine code, no intermediate allocations. 500,000 × 768-dim vectors scored in milliseconds on a single core.

---

Performance

	50k rows	500k rows
Omna search	9ms	27ms
Omna filter	9ms	27ms
Pandas + FAISS	~25ms + index build	~25ms + index build
Polars keyword regex	1ms — exact match only	1ms — exact match only

Benchmarked on MacBook Air M5 with the prior 384-dim model, 10-query median, warm index. The current model (nomic-embed-text-v1.5, 768-dim) roughly doubles the per-query cosine cost — still single-digit-to-low-tens of milliseconds at these sizes.

Omna inherits Polars' Arrow columnar memory. The Rust similarity kernel operates on the same memory — no copy into NumPy, no copy into a C buffer.

---

FAQ

Does Omna send my data to the cloud?

No. Embedding, search, filter, PII detection, and masking all run locally. The only method that makes a network call is ask(), which sends schema metadata and sample rows to Claude via the Anthropic API — and only when you explicitly call it.

Do I need a GPU?

No. FastEmbed uses ONNX and runs on CPU. On Apple Silicon, it uses CoreML automatically. Embedding 50,000 documents takes ~45 minutes on a MacBook Air M5 — a one-time cost. After that, search() and filter() run in milliseconds from the saved index.

Why not FAISS / ChromaDB / Pinecone?

Those are vector databases. Omna is a Polars plugin. If your data already lives in a DataFrame, Omna adds hybrid search (semantic + keyword) with zero infrastructure — no separate process, no index server, no network hop. It's the difference between df.omna.search(...) and spinning up a separate service just to query your own data.

What PII types does Omna detect?

PERSON, EMAIL, PHONE, CREDIT_CARD, US_SSN, IP_ADDRESS, IBAN, MEDICAL_RECORD_NUMBER, BANK_ACCOUNT, 220+ secret types (API keys, tokens), and 30+ international IDs, and more. Detection runs on Omna's own six-layer Rust engine — fully local, no heavy Python ML dependencies.

Which Polars versions are supported?

Omna is tested on Polars 1.0+. It installs as a namespace plugin via df.omna.* — no import needed after import omna.

The embed step took 45 minutes. Do I have to redo it every time?

No. embed() saves the index to .omna/{column}.parquet. Every subsequent search() or filter() call loads it in ~300ms. You only re-run embed() if your data changes.

---

Roadmap

# Coming in v0.2
matched = transactions.omna.join(regulatory_categories, on="description")
# Match rows between two DataFrames by meaning, not exact key.

Star the repo to follow progress.

---

What's new

The PII engine was rebuilt from a Presidio + spaCy wrapper into Omna's own six-layer Rust engine (recall up ~2×, see Enterprise-grade PII masking), and the embedding model was upgraded to nomic-embed-text-v1.5. Full history in CHANGELOG.md.

---

License

Layer	License
Python package (omna/)	MIT
Rust engine (src/)	Proprietary — ships as a compiled binary in the pip wheel

---

omna.dev · PyPI · GitHub