Skip to main content

ExtractedDocument

@forge/monorepo

@forge/monorepo / backend/src / ExtractedDocument

Type Alias: ExtractedDocument

ExtractedDocument = object

Defined in: backend/src/documents/index.ts:73

What a parser produces when it succeeds.

Properties​

blocks​

readonly blocks: readonly DocumentBlock[]

Defined in: backend/src/documents/index.ts:74

confidence?​

readonly optional confidence?: number

Defined in: backend/src/documents/index.ts:107

How confident the extraction is, 0–1 (#132).

Present only for extraction that has a confidence — OCR and vision. A PDF's text layer is not recognised, it is read, so a confidence there would be a number with nothing behind it. Absent therefore means "not a probabilistic extraction", which is a different fact from "confidence unknown".

frontMatter?​

readonly optional frontMatter?: Readonly<Record<string, string>>

Defined in: backend/src/documents/index.ts:99

YAML front matter, when the source had a leading --- block — REQ-050 (#209), task #220.

Scalar keys only, and deliberately not a YAML parser. A nested mapping, a list of mappings or a flow collection is reported in warnings and otherwise dropped, because a half-correct parse of provenance metadata is worse than none: generated: { by: x, at: y } silently read as the string { by: x, at: y } would be recorded as if it had been understood.

It exists because the alternative was worse. Front matter used to reach the block stream as content: a document's sidebar_position and type became a paragraph, got chunked, embedded, and could be returned as a retrieval hit and cited. #220 found it while reading the Open Knowledge Format — whose concept files carry far more metadata than ours — and it was already happening to this repository's own documentation site.

pageCount?​

readonly optional pageCount?: number

Defined in: backend/src/documents/index.ts:76

Pages seen. Absent for formats with no pagination — a CSV does not have pages.

truncated​

readonly truncated: boolean

Defined in: backend/src/documents/index.ts:78

True when a limit stopped extraction early. The content is usable; it is just not all of it.

usage?​

readonly optional usage?: readonly object[]

Defined in: backend/src/documents/index.ts:117

Priced operations this extraction performed (#132, AC-4).

Transport only. The pipeline reports each entry to onPricedOperation and then strips the field before storing the blob, so a stored document never carries billing data. It lives on the result rather than in a callback because only the pipeline knows which file and tenant a parse belonged to — a callback handed to the parser at construction is shared across every file it ever parses, and correlating one would mean trusting that parses never interleave.

warnings​

readonly warnings: readonly string[]

Defined in: backend/src/documents/index.ts:85

Things the caller should know that are not failures.

Surfaced rather than logged: "three pages had no text layer" changes how much a user should trust an answer, and a log line does not reach them.