Less jargon.
More useful meaning.
Twenty-five terms that help connect the source, the record, and the receiving system. Each definition leads to a complete field guide.
API
An application programming interface defines how one software component requests or exchanges information with another. An extraction API should make its accepted inputs, returned records, and failure states explicit.
Read the connected guide ↗Extraction
Selecting information from an input and mapping it into a defined record. It is distinct from retrieving the input and from delivering the accepted result to another system.
Read the connected guide ↗Source adapter
The part of a workflow that understands one source’s resources, access conditions, and field conventions. Keeping it separate from the canonical record makes source changes easier to contain.
Read the connected guide ↗Schema
A description of the expected structure of a record, including types and required properties. Structural validity does not establish that the fields are true or supported by the source.
Read the connected guide ↗Canonical record
The application’s agreed internal representation of an accepted item. Define format-specific exports from this record so that different delivery paths do not quietly develop incompatible meanings.
Read the connected guide ↗Provenance
Information about where a record or value came from and how it was obtained. A useful provenance design identifies the source and relevant processing context without retaining unnecessary sensitive material.
Read the connected guide ↗Evidence span
A reference to the source passage supporting an extracted value. The passage must not only exist; it must refer to the correct item and actually support the field’s interpretation.
Read the connected guide ↗Normalization
A documented transformation that gives source values a consistent representation, such as trimming irrelevant whitespace or using an agreed date format. It should not conceal unsupported assumptions.
Read the connected guide ↗Validation
A set of checks that determine whether a candidate record satisfies structural, semantic, and evidence requirements. Keep the reasons for rejection visible rather than treating every failure as an empty result.
Read the connected guide ↗Null
An explicit JSON value that can represent absence under a contract. A missing property, an empty string, zero, and null can mean different things to a receiving application.
Read the connected guide ↗CSV
A text representation for flat records. Delimiters, quoting, headers, encoding, and the consumer’s import behavior need to be agreed and tested; a simple comma split is not a general parser.
Read the connected guide ↗JSON
A structured interchange format using objects, arrays, strings, numbers, booleans, and null. Dates, identifiers, precision, and business relationships still require application-level conventions.
Read the connected guide ↗MIME
A framework used to describe message content and parts, including multipart email and attachments. Use a message parser to inspect these parts instead of treating an entire email as one undifferentiated string.
Read the connected guide ↗DOM
The document object model represents page content as a structured tree. A useful extractor identifies the correct record boundary rather than combining unrelated elements found anywhere in the document.
Read the connected guide ↗CORS
Cross-origin resource sharing is a browser access mechanism. A normal page cannot turn an opaque response into readable content simply by selecting no-cors mode.
Read the connected guide ↗Pagination
Retrieving a collection through bounded pages or continuation steps. Completion should describe whether the whole intended collection was traversed, not merely whether one request succeeded.
Read the connected guide ↗Cursor
A continuation value used to locate the next portion of a collection. Treat it according to the source interface rather than inventing assumptions about its contents or long-term validity.
Read the connected guide ↗Watermark
A recorded position used by an incremental workflow to track progress through changes. Commit it deliberately so a failed delivery does not cause unacknowledged records to be skipped.
Read the connected guide ↗Idempotency
A property of an operation that allows repetition without creating an unintended additional effect. In delivery, stable identities can help distinguish retries from new observations or updates.
Read the connected guide ↗Snapshot
A dataset intended to describe a particular state or observation window. Define whether the implementation provides a consistent point-in-time view or a rolling collection of observations.
Read the connected guide ↗OCR
Optical character recognition identifies text in visual inputs. Recognized text is an input to interpretation and validation, not proof that a business field has been understood correctly.
Read the connected guide ↗LLM
A large language model can help map varied text into selected fields. Keep the task narrow, treat source instructions as untrusted content, and validate candidate values outside the model.
Read the connected guide ↗Prompt injection
An attempt by content processed by a model to alter its intended behavior. Use layered controls, constrained permissions, and validated outputs rather than relying on prompt wording alone.
Read the connected guide ↗Least privilege
Giving a component only the access necessary for its task. A parser, a source fetcher, and a delivery worker need not all hold the same broad credentials.
Read the connected guide ↗Self-hosting
Running the selected software on infrastructure you control. It creates responsibilities for resource limits, updates, secrets, observability, retention, and recovery; packaging alone does not address them.
Read the connected guide ↗Technical definitions connect to the primary references cited in each full article. Core references include RFC 8259, MDN Fetch guidance, and JSON Schema object documentation.