infer

Classify text into one of Finetype's 251 semantic types. Use infer to explore individual values, process files of inputs, or stream column batches via JSONL.

infer runs the shipped model, which predicts 244 of the taxonomy's 251 types directly. The other 7 are taxonomy-live but not model-predicted — infer will not return them; only finetype profile's column-level guards recover them. See the type registry for the full taxonomy and how the two counts relate.

Usage

finetype infer [OPTIONS]

Options

Examples

These examples use values from contacts.csv, the quick start's sample file.

Classify a single value

The label prints first, with the confidence and sample count beneath it:

finetype infer -i "alice@example.com" --confidence

identity.person.email
  confidence: 0.9900 (1 samples)
  disambiguation: deterministic_fast_path

deterministic_fast_path means the value's structure settled the type without the model, and the confidence is then a fixed 0.99. It applies to a narrow set of types, among them email addresses, IPv4 addresses and delimited ISO 8601 timestamps, and only when no --header is given. Other values go to the model:

finetype infer -i "not-an-email" --confidence

representation.text.entity_name
  confidence: 0.4900 (1 samples)

A single value carries less signal than a full column. For anything beyond a quick check, reach for profile.

Classify a column from a file

A file of values, one per line, is treated as a single column and resolved to one type. This writes the email column of contacts.csv to a file and classifies it with its column name as the header hint:

tail -n +2 contacts.csv | cut -d, -f3 > emails.txt
finetype infer -f emails.txt --header "email" --confidence

identity.person.email
  confidence: 0.6383 (12 samples)

With --header the model classifies the column, as profile does, and reads it at the same confidence. Without --header, 11 of the 12 values pass the email check, which is enough for the structural path:

finetype infer -f emails.txt --confidence

identity.person.email
  confidence: 0.9900 (12 samples)
  disambiguation: deterministic_fast_path

Classify a stream of columns

--batch reads one column per JSONL line from stdin and writes one JSON line per column. Progress lines go to stderr:

printf '%s\n' \
  '{"header":"email","values":["alice@example.com","bob@company.org","not-an-email"]}' \
  '{"header":"created_at","values":["2024-01-15T09:30:00Z","2024-02-20T14:15:00Z"]}' \
  | finetype infer --batch

{"confidence":0.7841140627861023,"disambiguation_rule":"header_hint_cross_domain:email","label":"identity.person.email","samples_used":3}
{"confidence":0.9971720576286316,"label":"datetime.timestamp.iso_8601","samples_used":2}

Row mode

The shipped model is column-level, so --mode row stops with an error:

finetype infer -i "alice@example.com" --mode row

Error: Row mode is unsupported: the shipped model is column-level. Use --mode column (the default) or `finetype profile`.

To profile a whole CSV in one step, every column at once, use profile.

Output formats

-o json emits the full classification, including the vote distribution and which disambiguation rule fired. The vote distribution is empty when the structural path settled the type:

finetype infer -i "alice@example.com" -o json --confidence

{
  "confidence": 0.9900000095367432,
  "disambiguation_applied": true,
  "disambiguation_rule": "deterministic_fast_path",
  "label": "identity.person.email",
  "samples_used": 1,
  "vote_distribution": []
}

-o csv gives a compact label,confidence,samples row, handy for piping:

finetype infer -i "alice@example.com" -o csv --confidence

identity.person.email,0.9900,1

See also

• profile — profile an entire CSV file using column-mode inference
• taxonomy — browse the semantic type taxonomy
• Type Registry — browse all 251 semantic types