Engineering at Nutrient

Engineering at Nutrient is where we share insights into how we’re building the @nutrientdocs SDKs, Document Web Services APIs, and AI capabilities in documents.

[Engineering at Nutrient @nutrient_eng]

Your agent doesn’t just need extracted text.

It needs to know:

Where did this value come from?
How confident are we?
What was its position on the page?

And most importantly...Can a human verify it?

Nutrient Data Extraction API returns structured document data with page context, coordinates, confidence, and source evidence, so downstream AI and human review can work from something traceable.

Try it with your own document: https://t.co/IqZr0rjJUC

[Engineering at Nutrient @nutrient_eng]

Document Engine 1.18.1 is out, with faster Instant sync writes and more flexible remote URL allowlists.

• Validate PDF/UA-1 and get a machine-readable report
• Use PowerPoint and Excel files with Office templating
• Improve print fidelity with per-request CMYK rendering options
• Get better metrics, distributed traces, and permission errors

Plus fixes across rendering, redaction, forms, memory use, and document conversion.

https://t.co/YIrpRztB1Z

[Engineering at Nutrient @nutrient_eng]

There’s no single best document parser. The right choice depends on what you’re processing and what you need to do with the result.

LlamaParse is strong with spatially dense forms and chart reconstruction. Nutrient is built for content-heavy documents where grounding, review, and deployment flexibility matter.

We compared the tradeoffs across document types, structured output, self-hosting, and pricing:
https://t.co/Wvwj9EnVry

[Engineering at Nutrient @nutrient_eng]

If your users annotate a PDF, they probably expect those annotations to follow the file.

With a custom PDF.js viewer, that isn’t automatic. This guide walks through embedding highlights, comments, and area annotations directly into the exported PDF.

https://t.co/Os8T4ZAcGt

[Engineering at Nutrient @nutrient_eng]

RAG has a dirty secret: it doesn't always work.

Traditional RAG usually follows a fixed path: retrieve relevant results, then answer from the selected context. The issue is that the information it returns is often wrong, or irrelevant.

Agentic RAG upgrades that process for harder questions. The agent can rethink the query, search again, choose different sources, and decide when it has enough information to answer.

But more autonomy cannot fix bad inputs. If document extraction flattens a table, scrambles the reading order, or skips a scanned page, that garbage will often poison the context and lead to poor outcomes.

Learn how agentic RAG works, where it improves on traditional retrieval, and why document quality still sets the ceiling:

https://t.co/fCVvudlKP9

[Engineering at Nutrient @nutrient_eng]

React Native Version 4.5.1 is available https://t.co/H1s9Wysg4O

[Engineering at Nutrient @nutrient_eng]

The addPage gotcha and every use case, read it. https://t.co/TZ3NtWjizh

[Engineering at Nutrient @nutrient_eng]

PDFDocument.load() then copyPages() then addPage() then save(). Copied pages aren't inserted until you call addPage, easy to forget, easy to debug for twenty minutes. Zero-indexed pages, works in Node and browser, no fonts or workers.

[Engineering at Nutrient @nutrient_eng]

CMap paths, TextContent filtering, the polyfill, all covered. https://t.co/qICiB8dvRa

[Engineering at Nutrient @nutrient_eng]

getDocument({data}) wants a Uint8Array, not your Buffer. No worker, no browser, but you still have to hand-resolve standardFontDataUrl and cMapUrl from the package on disk or CJK text comes back as mojibake. Full breakdown inside.

[Engineering at Nutrient @nutrient_eng]

PDF.js area annotations need pointer tracking, PageViewport coordinate conversion, canvas capture (derive DPI from canvas.width over canvas.clientWidth, not devicePixelRatio), and react-rnd for drag-resize. Five parts, all fragile at zoom/rotation.

[Engineering at Nutrient @nutrient_eng]

Full React component in the post. https://t.co/2RCaYCYlXQ

[Engineering at Nutrient @nutrient_eng]

viewer.currentScaleValue takes strings, even for numeric zoom levels — pass a number and it silently fails. Full PDFViewer navigation, zoom, and rotation walkthrough with EventBus listeners and a complete React toolbar component in the post.

[Engineering at Nutrient @nutrient_eng]

Tool definitions and adapters are in the post. https://t.co/ZAKtCtPgTa

[Engineering at Nutrient @nutrient_eng]

Document Authoring AI adds structured read/write tools, a transaction-API validation boundary, and tracked-changes review to Document Authoring SDK. Vercel AI SDK and LangChain adapters ship in the package; JSON export covers the rest.

[Engineering at Nutrient @nutrient_eng]

Full OCR vs. VLM comparison inside. https://t.co/tFtVfwQTSU

[Engineering at Nutrient @nutrient_eng]

Most VLMs return handwriting as ungrounded text — no coordinates, no confidence, non-deterministic at temperature zero. Nutrient Vision API adds OCR, ICR, and VLM-enhanced ICR with word-level bounding boxes and per-word confidence.

[Engineering at Nutrient @nutrient_eng]

Text search in a PDF is more than finding a word.

A useful search experience needs to scan every page, highlight matches, show a live result count, and let users move between results, even when those pages haven’t been displayed yet.

PDF.js can do all of this, but connecting it to a custom React search bar takes some work. Our guide walks through the complete implementation, then shows how Nutrient Web SDK handles it with a built-in search API and UI.

Read the guide: https://t.co/jLvCYvb3f0

[Engineering at Nutrient @nutrient_eng]

Architecture and code inside. https://t.co/D6cHYkq7zi

[Engineering at Nutrient @nutrient_eng]

Multimodal RAG retrieves text, tables, and images instead of flattening everything into plain text.

But your RAG pipeline can’t retrieve what it never extracted. Miss a table, scramble a page’s reading order, or skip a scan, and even the best embedding model won’t find the answer.

This guide breaks down the architecture and shows how to turn messy PDFs into structured, searchable inputs.