<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"
	xmlns:content="http://purl.org/rss/1.0/modules/content/"
	xmlns:wfw="http://wellformedweb.org/CommentAPI/"
	xmlns:dc="http://purl.org/dc/elements/1.1/"
	xmlns:atom="http://www.w3.org/2005/Atom"
	xmlns:sy="http://purl.org/rss/1.0/modules/syndication/"
	xmlns:slash="http://purl.org/rss/1.0/modules/slash/"
	>

<channel>
	<title>Nanonets &#8211; iAIFeed</title>
	<atom:link href="https://www.iaifeed.com/tag/nanonets/feed" rel="self" type="application/rss+xml" />
	<link>https://www.iaifeed.com</link>
	<description>iAIFeed</description>
	<lastBuildDate>Tue, 10 Nov 2026 12:57:04 +0000</lastBuildDate>
	<language>en-US</language>
	<sy:updatePeriod>
	hourly	</sy:updatePeriod>
	<sy:updateFrequency>
	1	</sy:updateFrequency>
	<generator>https://wordpress.org/?v=7.1.5</generator>

<image>
	<url>https://www.iaifeed.com/wp-content/uploads/2026/07/cropped-iaifeed1-80x80.png</url>
	<title>Nanonets &#8211; iAIFeed</title>
	<link>https://www.iaifeed.com</link>
	<width>32</width>
	<height>32</height>
</image> 
	<item>
		<title>How to Raise Nanonets Accuracy on Messy Vendor Invoices</title>
		<link>https://www.iaifeed.com/how-to-raise-nanonets-accuracy-on-messy-invoices</link>
					<comments>https://www.iaifeed.com/how-to-raise-nanonets-accuracy-on-messy-invoices#respond</comments>
		
		<dc:creator><![CDATA[iamltlb]]></dc:creator>
		<pubDate>Tue, 10 Nov 2026 12:57:04 +0000</pubDate>
				<category><![CDATA[Blog]]></category>
		<category><![CDATA[Nanonets]]></category>
		<guid isPermaLink="false">https://www.iaifeed.com/how-to-raise-nanonets-accuracy-on-messy-invoices</guid>

					<description><![CDATA[Practical fixes for failing invoice extraction: preprocessing, field-level prompts, confidence thresholds per field, and a retraining loop that actually converges.]]></description>
										<content:encoded><![CDATA[<p>The first month with any extraction tool looks great. Then Q2 arrives, three new vendors send scanned invoices from 2011-era software, and accuracy falls off a cliff.</p>
<p>This guide is about the recovery path in <a href="https://www.iaifeed.com/ai-tool/nanonets">Nanonets</a>: what to change when field accuracy drops, which fixes cost nothing, and which require retraining. It assumes you already have a working invoice model and a review queue.</p>
<h2 id="why-does-accuracy-drop-when-it-worked-before">Why does accuracy drop when it worked before?</h2>
<p>Rarely the model&#8217;s fault. In order of frequency: scan quality dropped, a new vendor layout appeared, a field definition shifted (someone started sending a PO number in the &#8220;reference&#8221; field), or the document class changed — a credit note being processed as an invoice.</p>
<p>Check the failing documents by vendor first. If failures cluster on two vendors, it is a layout problem. If they scatter across all vendors, it is a quality or threshold problem.</p>
<h2 id="fix-1-preprocess-the-images-before-extraction">Fix 1: Preprocess the images before extraction</h2>
<p>This is the cheapest fix and the most skipped. Photographed invoices and old scans improve markedly with three steps:</p>
<ul>
<li><strong>Deskew</strong> — rotate to straight. Many scanner APIs and most image tools expose this.</li>
<li><strong>Increase DPI</strong> — 300 DPI is the sweet spot. Below 200, small characters merge.</li>
<li><strong>Convert to grayscale and boost contrast</strong> — removes colored backgrounds that confuse text detection.</li>
</ul>
<p>A 150 DPI skewed phone photo and the same image deskewed at 300 DPI grayscale are, functionally, different documents for a model. Teams routinely get 10-15 accuracy points from preprocessing alone.</p>
<h2 id="fix-2-define-fields-at-the-level-the-model-needs">Fix 2: Define fields at the level the model needs</h2>
<p>&#8220;Total&#8221; is ambiguous. There is a subtotal, a tax total, a currency total, and sometimes a running total per page. Vague fields produce inconsistent outputs that look like model error.</p>
<p>Define each field with an example and a rule:</p>
<ul>
<li><code>total_gross</code> — the final amount payable including tax. Example: 1,847.50. Not the subtotal.</li>
<li><code>invoice_number</code> — vendor&#8217;s document ID. Excludes PO numbers and customer IDs.</li>
<li><code>due_date</code> — date payment is due, not issue date. If missing, derive from payment terms and mark as derived.</li>
</ul>
<p>Field-level clarity fixes a surprising share of &#8220;accuracy&#8221; problems because the model was answering a different question than you meant.</p>
<h2 id="fix-3-set-thresholds-per-field-not-globally">Fix 3: Set thresholds per field, not globally</h2>
<p>One confidence threshold for all fields is a blunt instrument. Some fields are cheap to get wrong and some are not.</p>
<ul>
<li><code>total_gross</code> — high threshold. A wrong total posts wrong money.</li>
<li><code>invoice_number</code> — high threshold, because of duplicate detection.</li>
<li><code>vendor_address</code> — low threshold. Nobody reconciles from it.</li>
<li><code>line_items</code> — medium, and accept partial extraction where the table is unclear.</li>
</ul>
<p>Per-field thresholds let you auto-post 70% of invoices while routing the risky ones to a human. That is the only configuration where automation and control coexist.</p>
<h2 id="fix-4-retrain-with-corrections-not-with-volume">Fix 4: Retrain with corrections, not with volume</h2>
<p>The review queue teaches the model, but only if corrections are diverse. Two hundred corrections on the same vendor layout teach it one layout.</p>
<p>A working cadence: weekly, take the corrections from the past seven days, sample 20 that cover different vendors, and confirm they are labeled correctly before adding them to training. Six weeks of that with diverse samples beats six months of bulk uploading.</p>
<h2 id="how-do-you-know-it-improved">How do you know it improved?</h2>
<p>Measure straight-through processing rate, not raw accuracy. STP is the share of documents that reach the ERP with no human touch. Raw accuracy can improve while STP stays flat if the improved fields were not the blocking ones.</p>
<p>Track two numbers monthly: STP rate and exception rate by vendor. When a vendor&#8217;s exception rate stays near zero for a month, you have a stable layout. When it jumps, a template changed and nobody told finance.</p>
<h2 id="what-still-cannot-be-fixed">What still cannot be fixed</h2>
<p>Handwriting. Cursive amounts on a scanned invoice are a coin flip, and no preprocessing resolves it. Route handwritten documents to a human queue from the start rather than trying to solve it with model tuning.</p>
<p>Multi-page invoices where line items continue across pages are the second hard case. Page-level extraction loses the continuation; whole-document extraction sometimes merges unrelated tables. Test both and pick whichever fails less on your actual vendors.</p>
<h2 id="quick-summary">Quick Summary</h2>
<ul>
<li>When accuracy drops, diagnose by vendor first — clustering failures mean layout problems, scattered failures mean quality or thresholds.</li>
<li>Deskew, upscale to 300 DPI, and grayscale with contrast boost routinely add 10-15 accuracy points.</li>
<li>Define each field with an example and an exclusion rule; ambiguous field names cause most &#8220;model errors&#8221;.</li>
<li>Set high confidence thresholds for total and invoice number, low for address, and measure straight-through processing rate rather than raw accuracy.</li>
<li>Handwritten amounts and multi-page line-item continuations should route to humans, not be tuned.</li>
</ul>
]]></content:encoded>
					
					<wfw:commentRss>https://www.iaifeed.com/how-to-raise-nanonets-accuracy-on-messy-invoices/feed</wfw:commentRss>
			<slash:comments>0</slash:comments>
		
		
			</item>
		<item>
		<title>How to Extract Invoice Data with Nanonets (No Code Setup)</title>
		<link>https://www.iaifeed.com/how-to-extract-invoice-data-with-nanonets</link>
					<comments>https://www.iaifeed.com/how-to-extract-invoice-data-with-nanonets#respond</comments>
		
		<dc:creator><![CDATA[iamltlb]]></dc:creator>
		<pubDate>Tue, 10 Nov 2026 12:56:43 +0000</pubDate>
				<category><![CDATA[Blog]]></category>
		<category><![CDATA[Nanonets]]></category>
		<guid isPermaLink="false">https://www.iaifeed.com/how-to-extract-invoice-data-with-nanonets</guid>

					<description><![CDATA[Build a working invoice extractor in Nanonets: upload samples, label three fields, set validation rules, and push clean rows to Google Sheets or your ERP.]]></description>
										<content:encoded><![CDATA[<p>Nanonets is an IDP (intelligent document processing) platform. You upload documents, label a handful of example fields, and it builds a model that pulls those fields from every future document of the same type.</p>
<p><a href="https://www.iaifeed.com/ai-tool/nanonets">Nanonets</a> is used mostly for invoices, the standard first project: high volume, predictable layout variance, and a clear payoff. Ten seconds saved per invoice across 2,000 monthly invoices is worth setting up. This guide builds one end to end.</p>
<h2 id="what-is-the-difference-between-ocr-and-nanonets">What is the difference between OCR and Nanonets?</h2>
<p>OCR gives you text. Nanonets gives you fields.</p>
<p>A plain OCR tool reads an invoice and returns every word on the page in a jumble. Nanonets returns <code>invoice_number: INV-2291</code>, <code>invoice_date: 2026-03-04</code>, <code>total: 1,847.50</code>. The difference is the model knows what an invoice number looks like and where to look for it.</p>
<p>That matters because the expensive part of AP is not reading — it is keying values into a system in the right format.</p>
<h2 id="step-1-choose-a-prebuilt-model-or-a-custom-one">Step 1: Choose a prebuilt model or a custom one</h2>
<p>Nanonets ships prebuilt models for invoices, receipts, purchase orders, and a long list of ID documents. Start with the prebuilt invoice model.</p>
<p>Only build a custom model when your invoices are genuinely unusual — vendor-specific layouts with industry codes in odd places, or a proprietary form that only looks like an invoice. Most teams save a week by starting prebuilt and editing the schema.</p>
<h2 id="step-2-upload-samples">Step 2: Upload samples</h2>
<p>Upload 15 to 30 real invoices from different vendors. Variation is the point: three invoices from the same vendor teach the model nothing about how to handle the other 40 vendors.</p>
<p>Include the ugly ones. A clean digital PDF and a photographed crumpled paper invoice produce very different confidence scores, and you want the model trained on both before you find out the hard way.</p>
<h2 id="step-3-review-and-correct-the-extracted-fields">Step 3: Review and correct the extracted fields</h2>
<p>The prebuilt model returns fields with confidence scores. Anything below your threshold is flagged for review, and you correct it in the review queue. Those corrections become training data.</p>
<p>Expect the first 50 to 100 documents to need heavy correction. Accuracy on clean printed invoices commonly lands above 95% after that, while photographed and heavily formatted invoices lag behind. The review queue is not a bug — plan for a human to touch the low-confidence ones indefinitely.</p>
<h2 id="step-4-add-validation-rules">Step 4: Add validation rules</h2>
<p>Extraction is half the job. The other half is catching garbage before it reaches your ledger.</p>
<ul>
<li><strong>Arithmetic check</strong> — line items sum to subtotal, subtotal plus tax equals total. Mismatches should block, not warn.</li>
<li><strong>Duplicate check</strong> — invoice number plus vendor plus amount. Duplicate invoices are a real AP risk and a common fraud vector.</li>
<li><strong>Required field check</strong> — missing PO number on a vendor that requires one should route to a person, not pass silently.</li>
</ul>
<p>These three rules catch more errors than any model improvement you will make this year.</p>
<h2 id="step-5-send-the-data-somewhere">Step 5: Send the data somewhere</h2>
<p>Nanonets exports to Google Sheets, an API endpoint, or through a connector to a finance system. For a first project, push to a Google Sheet and let a human review a week of rows before connecting anything to the ledger.</p>
<p>When you do connect to the ERP, send only validated rows and route exceptions to a person. The failure mode to avoid is a clean automated pipeline that quietly posts wrong amounts.</p>
<h2 id="where-nanonets-is-the-wrong-tool">Where Nanonets is the wrong tool</h2>
<p>If your documents are hand-filled checkboxes and irregular scribbles, extraction accuracy drops hard and the review workload can exceed manual entry. Scan quality is the dominant variable — a 150 DPI flatbed scan and a phone photo of a receipt at an angle are different problems.</p>
<p>Also, it is priced by document volume. At a few hundred documents a month, the platform fee may exceed the labor it saves. Do the arithmetic with your own hourly rate before committing.</p>
<h2 id="quick-summary">Quick Summary</h2>
<ul>
<li>Nanonets is a no-code document extraction platform that returns structured fields, not raw OCR text.</li>
<li>Start with the prebuilt invoice model and upload 15-30 invoices across different vendors, including photographed ones.</li>
<li>Expect heavy correction for the first 50-100 documents, then above 95% accuracy on clean printed invoices.</li>
<li>Add three validation rules — arithmetic, duplicate detection, required fields — before any ERP connection.</li>
<li>Poor-quality scans and very low document volume are the two cases where manual entry wins.</li>
</ul>
]]></content:encoded>
					
					<wfw:commentRss>https://www.iaifeed.com/how-to-extract-invoice-data-with-nanonets/feed</wfw:commentRss>
			<slash:comments>0</slash:comments>
		
		
			</item>
	</channel>
</rss>
