Embedded name and address detection

EmbeddedCnnDetector carries the statements-v1 CNN inside the extension. It detects recipient names (PERSON) and full mailing-address blocks (ADDRESS), including United States and United Kingdom layouts. Weights are compiled into the artifact: no model download, no files, no ONNX runtime, no external interpreter. The browser binding owns the same detector in a dedicated Worker. Python and Node.js expose it directly and expect the caller to keep it off the main thread for sustained throughput. Rust callers use datafog_core::cnn::CnnDetector behind the embedded-cnn feature.

from datafog_core import EmbeddedCnnDetector

detector = EmbeddedCnnDetector()  # authenticates the embedded weights once
result = detector.scan(text, {"entities": ["PERSON", "ADDRESS", "EMAIL"]})
for finding in result.findings:
    print(finding.entity_type, finding.matched_text)

import { EmbeddedCnnDetector } from "@datafog/node";

const detector = new EmbeddedCnnDetector();
const result = detector.scan(text, { entities: ["PERSON", "ADDRESS", "EMAIL"] });

ADDRESS is the full block and is distinct from STREET_ADDRESS. PERSON from this detector is prose coverage; the rule engine’s PERSON remains structured-field-only.

Methods

Member	Behaviour
`EmbeddedCnnDetector()`	Authenticates embedded weights and builds the contract
`describe()`	Model id, digest, labels, adapter version, score scale, thresholds
`scan(text, config=None)`	Rule detectors plus requested model labels. Result carries `contract_version`, `model`, `findings` and `evidence`
`scan_model(text)`	Model-only findings, for parity and evaluation
`EmbeddedCnnDetector.scan_disabled(text, config=None)`	Validates the same selection with no weight authentication or inference. Model-only selections return no findings

Omitting config equals {}: every enabled rule detector plus the model’s remaining labels. Selection is validated even while disabled, so an unknown label fails in both states.

Limits and behaviour

65,536 input bytes and 8,192 tokens per request; larger input fails rather than truncating.
At most 1,024 findings.
Inputs over 256 tokens use 256-token windows with 64-token overlap.
Inference is synchronous and in-process. ModelCancellation does not interrupt it, unlike ModelSession.scan.
Errors never return partial detections. There is no silent fallback to rule findings when model loading or inference fails.
Raw tag evidence is not calibrated confidence; the CNN publishes no calibrated score.

Ranges follow the normal finding contract: UTF-8 byte and Unicode code-point ranges everywhere, plus UTF-16 ranges in JavaScript. ADDRESS and PERSON findings deduplicate against rule findings on the same span.

Get started

Core concepts

Guides

Reference

Project

Embedded name and address detection

Methods

Limits and behaviour