oconvPdfToIr

.pdf (@awacloud/pdf's pdf.read() result) → oconv-ir/v1 — tier 1, text-first only.

Module oconvPdfToIr (oconvPdfToIr) | Source packages/front/office/oconv/src/read/pdf-to-ir.js | Deps oconvIr, oconvPdfTextExtract, oconvPdfStruct | Worker-safe yes

Pure transformation over the parsed pdf.read(bytes) structure plus on-demand content-stream decode through oconvPdfTextExtract/oconvPdfStruct. Real text is extracted through a font's ToUnicode CMap, the encoding-table + AGL hop, or (composite fonts only) a Unicode-coded predefined CMap; scanned/OCR is permanently out, and there is no layout inference beyond text-positioning heuristics.

Resolve

import { ModuleRuntime } from '@awacloud/fw/core/runtime.js';
import { fw_require, modules } from '@awacloud/oconv';

const runtime = new ModuleRuntime();
runtime.registerAll(fw_require);
runtime.registerAll(modules);
const pdf = runtime.resolve('pdf');
const { pdfToIr } = runtime.resolve('oconvPdfToIr');

API

Examples

Convert a .pdf read result to IR

const { ir, losses, coverage } = pdfToIr(pdf.read(bytes));
ir.kind;             // 'document'
coverage.tagged;     // false when the document has no /StructTreeRoot
// measured on the vendored facturx-minimum-sample.pdf fixture:
// coverage.total → {operators: 3585, decoded: 4128, undecodable: 0}

Notes

• Two paths: a tagged fast path (/StructTreeRoot present) derives structure from the logical tree — H1..H6 struct elements become IR headings, text-ish tags become paragraphs, keyed to text through marked-content ids; a Table/L container is flattened (loss struct/dropped). An untagged fall-back groups content-stream text pieces into lines and paragraphs by vertical position, inferring no heading; within a line it inserts one space where the horizontal gap between a piece's end and the next piece's start exceeds 0.15 em (never doubling an existing space, never on a line whose pieces go backwards). A piece that exactly repeats the previous piece of its line — same text, within 0.5 pt in x and in y — is dropped first. A piece that starts left of the previous piece's end (an overlap) gets one space iff the overlap exceeds the same 0.15 em — so a kerned apostrophe or a syllable split across two shows stays joined.
• Page boundaries become IR hr nodes in both paths — the frozen vocabulary has no section node.
• coverage is reported, never assumed: per-page and total {operators, decoded, undecodable} counts, honest-loss-matrix discipline (never claim zero loss from an unmeasured decode).
• Loss codes this module and its dependencies emit: text/undecodable (≥1 char code resolved to no Unicode), text/font-unresolved (a show op ran with no resolvable current font), text/width-approximated (a font's glyph widths fell back to a declared width — once per font resource per page), image/dropped (an image XObject or inline image, inside a Form XObject too — tier 1 keeps none), xobject/form-dropped (a Form XObject whose stream cannot be used — detail names the reason; a usable form is executed and its text extracted), xobject/form-cycle / xobject/form-depth / xobject/form-budget (a Form XObject not executed by a guard: drawn from inside itself, nested deeper than 12, or past a budget of formOpBudget form operators on the page, default 1,000,000), content/undecodable (a content stream failed to decode/parse — detail stream <objNum>: <cause>, the thrown error's message), struct/dropped (tagged path, a Table/L container flattened).
• opts.formOpBudget: absent (undefined) resolves to the default (1,000,000); present, it must satisfy Number.isSafeInteger(v) && v >= 1, else pdfToIr throws oconv: bad form op budget — checked at entry, before the first page is extracted, so a zero-page document still refuses a bad value. The option is forwarded to every page's oconvPdfTextExtract.extractPage call, which applies the identical rule.

See also

• docs/api/README.md — full module index + exports boundary note
• oconvIr — the pivot this reader produces
• oconvPdfTextExtract, oconvPdfStruct — the two dependencies this reader composes
• loss matrix