API reference
Detectors
Structured identifiers are found by patterns and validated by their own checksums where one exists. Names and places come from an NER model and are best-effort.
Supported types
| Label | What | How |
|---|---|---|
CREDIT_CARD | Payment cards, 13–19 digits | pattern + Luhn checksum |
IBAN | International bank account numbers | pattern + mod-97 checksum |
AADHAAR | Indian Aadhaar numbers | pattern + Verhoeff checksum |
GSTIN_IN | Indian GST identification numbers | pattern + check character |
PAN_IN | Indian PAN | pattern |
SSN_US | US Social Security numbers | pattern |
EMAIL | Email addresses | pattern |
PHONE | Phone numbers, incl. Indian +91 formats | pattern |
IPV4, IPV6 | IP addresses | pattern |
MAC | MAC addresses | pattern |
AWS_ACCESS_KEY | AWS access key IDs | pattern |
PERSON | Names | NER (spaCy) |
LOCATION | Places | NER (spaCy) |
The live list: GET https://api.obscate.com/v1/entities. Restrict a request to some types with "labels": ["EMAIL", "PHONE"]. Company names are not detected: they are generally not personal data.
Accuracy, stated honestly
- Structured IDs are near-perfect. F1 1.00 for every structured and checksum type on our synthetic benchmark. Checksum validation is why precision stays at 1.00: a number that fails its checksum is rejected. Clean synthetic input makes this an upper bound, not a real-world guarantee.
- Names and places are best-effort. PERSON F1 0.80 (recall 0.80), LOCATION 0.77. Single first names with little context are the main miss. You will see misses on real text: test on your data.
- Malformed IDs are missed by design. A mistyped card or Aadhaar that fails its checksum is not flagged: we favour precision over guessing.
- Higher recall for names needs a larger model or an LLM. The optional LLM detection tier exists in the library and is coming to hosted plans; it is not available on the hosted API today.
Methodology and per-type numbers: Benchmark.
Confidence and context
Each match has a confidence between 0 and 1. A match near a trigger word gets a boost: card, visa for cards; phone, call, mobile, whatsapp for phones; aadhaar, pan, ssn, ip. Set min_confidence to drop low-confidence matches. In our sample, “Priya Sharma” scored 0.6 and the email 0.99, so min_confidence: 0.7 would have let the name through. The default 0.0 keeps everything.
Closing the gap in JSON
In JSON, short values without context ({"name": "Ravi"}) are hard for any NER model. Field-name hints and your own field_map tell Obscate what a field holds, so it doesn’t have to guess.