Document Converter Agent Skill

Document Converter

TypeScript facade (@blackbelt-technology/pi-dashboard-document-converter) over

the Dockerized pi-doc-engine Python engine. **Always call the facade — never

Python, docling, pandoc, or the nano-banana CLI directly.** The engine is

quarantined in Docker; the facade is the only call surface.

Internals reference: packages/document-converter/README.md (facade API),

packages/document-converter/engine/README.md (JSON docker run contract).

Do NOT read the vendored Python under engine/document_converter/.

Prerequisites

• Docker available; image built: cd packages/document-converter && npm run build:image.
• Styled diagrams (nano-banana) need GEMINI_API_KEY at run time; absent or on

failure, rendering falls back to mmdc (never hard-fails).

Two directions

Ingest — any format → Markdown for kb

import { createDocumentConverter } from "@blackbelt-technology/pi-dashboard-document-converter";
const dc = createDocumentConverter({ image: "pi-doc-engine:0.1.0", stagingDir: "/abs/kb-staging" });

await dc.convertToMarkdown("/docs/report.pdf");                       // digital PDF (auto: native-first)
await dc.convertToMarkdown("/docs/scan.pdf", {                        // scanned HU PDF
  ocr: { mode: "force", lang: ["hungarian", "english"], engine: "tesseract" },
});
await dc.convertToMarkdown("/docs/huge.pdf", { tables: "off" });      // escape hatch: skip TableFormer

Output .md lands in stagingDir with provenance frontmatter. Feed kb via its

existing filesystem source (kb is NOT modified):

kb index  --source /abs/kb-staging
kb search "<query>" --source /abs/kb-staging

OCR lang takes canonical names ("hungarian"); the facade maps to per-engine

codes. A wrong name raises OCR_LANG_UNSUPPORTED — no silent empty OCR. Default

mode: auto skips OCR on digital PDFs.

Produce — Markdown → DOCX/PDF

await dc.renderDocx("/spec.md", { output: "/spec.docx", template: "default" });
await dc.renderDocx("/spec.md", { output: "/spec.docx", nanoBanana: { enabled: true, style: "ros-3d" } });
await dc.renderPdf("/spec.docx", { output: "/spec.pdf", pageSize: "a4" });

const { meta } = await dc.extractForEdit("/in.docx", "/edit.md");     // round-trip
await dc.mergeBack("/in.docx", "/edit.md", { output: "/out.docx", meta });

Template vars come from frontmatter. Diagrams render with mmdc by default; opt in

to styled images per-doc via nanoBanana (md5-cached .mermaid-cache/<md5>.png,

deterministic, mmdc fallback).

Frontmatter tooling (the shared bus)

await dc.fillFrontmatter(["spec/**/*.md"], { language: "hu" });       // defaults + language packs + globs
await dc.profileTables(["spec/**/*.md"]);                             // inject table_profiles width ratios

All stages read/write one unified schema (./schema): template vars ·

table_profiles · nano_banana · provenance.

Errors

Every method rejects with DocConverterError (.code, .stderr, .exitCode):

UNSUPPORTED_FORMAT, OCR_LANG_UNSUPPORTED, INGEST_FAILED, PRODUCE_FAILED,

DOCKER_UNAVAILABLE, … Surface .code + .stderr; do not retry blindly.