Tesseract OCR vs Unsiloed: Complex PDF Parsing
Compare Tesseract OCR and Unsiloed for complex PDF extraction, with a scanned-invoice example, output structures, citations, costs, and deployment choices.


Tesseract turns page images into text. Unsiloed turns complex PDFs into structured data: typed layout segments, tables that keep their merged headers, and named fields in the shape of your JSON Schema, with scores and optional citations back to the page. If your application needs an invoice total rather than the words on an invoice, Unsiloed can extract it directly.
Tesseract is a mature open-source OCR engine, and its job ends at recognized words and their positions. This article compares the two on integration, output, layout and tables, citations, cost, and deployment, and shows both tools' output on the same scanned invoice. For a comparison with another local toolkit, see Docling vs Unsiloed.
Tesseract vs Unsiloed at a Glance
| Requirement | Tesseract | Unsiloed |
|---|---|---|
| What it is | Open-source OCR engine (libtesseract) and command-line program, Apache 2.0 license |
Managed document parsing and extraction API |
| Input | Images; PDFs are rendered to images first | PDF, images, Word, Excel, PowerPoint, and scans, per the pricing page |
| Integration | C/C++ API, CLI, and wrappers such as pytesseract and tesserocr | Asynchronous REST jobs: submit, receive a job_id, fetch results |
| Output | Plain text, hOCR, searchable PDF, TSV, ALTO, and PAGE XML | Parse: Markdown chunks and typed layout segments, with tables in Markdown and HTML. Extract: JSON shaped by your schema |
| Named fields | Built in your application code | Extract accepts a JSON Schema and returns values in that shape |
| Review signals | Per-word confidence and pixel boxes | Per-segment detection confidence; per-field grounding and extraction scores, with page and bounding-box citations when enabled |
| Languages | More than 100, via traineddata files | Multilingual OCR |
| Deployment | Wherever you install it | Hosted API, or VPC, on-premises, and air-gapped deployment on the Enterprise plan |
| Cost | Free software, plus compute and engineering time | Free Starter plan; usage-based Standard and volume-based Enterprise plans with custom pricing |

The Same Scanned Invoice in Both Tools
The test page is a one-page invoice with a two-column bill-to and ship-to block, a ship date next to the issue date, and a line-item table whose "Price (USD)" header spans the Unit and Amount columns. It was flattened to an image-only PDF at 200 DPI with a slight skew, so both tools had to read pixels rather than a text layer.
Tesseract 5.5.2 ran with the English traineddata on a page rendered by pdftoppm:
pdftoppm -r 300 -png -singlefile invoice-scanned.pdf page
tesseract page.png tesseract-psm3 -l eng --psm 3 txt tsv hocr
tesseract page.png tesseract-psm6 -l eng --psm 6 txt
The same PDF went to Unsiloed's POST /parse with default settings, and to POST /v2/extract with a seven-field JSON Schema and enable_citations=true.
Rendering the 200 DPI scan at 300 DPI resamples it; it does not add detail.

- Tesseract read every word correctly, with a mean word confidence of 95.9 across the page.
- Tesseract's text follows the page line by line. In the default page segmentation mode (3), the two address columns came out on shared lines such as "Accounts Payable Receiving Dock 3", and table rows came out as space-separated words, with "Price (USD)" on its own line above "Unit Amount". The TSV holds pixel coordinates for every word, which is the starting point for column and table logic in your own code.
- Unsiloed Parse returned the page's structure. The bill-to and ship-to blocks were separate
Textsegments with their own bounding boxes. The table was aTablesegment whose HTML usedrowspanandcolspanto keep the grouped header, with the subtotal, tax, and total rows folded in. - Unsiloed Extract returned every field correctly. All seven fields matched the printed values, the four line items summed to the printed subtotal, and every field, down to each cell of each line item, came back with a page, a bounding box, and grounding and extraction scores.
This is one synthetic page, not an accuracy benchmark. Unsiloed kept the address blocks intact, but its segment order was not strictly top-to-bottom: the bill-to block appeared before the invoice-number block. The Markdown table did not retain the merged header; that structure appeared in the HTML.
Integration: Command Line and Library vs Asynchronous API
Tesseract ships as the libtesseract engine and a tesseract command-line program. The Tesseract README documents the C and C++ API, and the AddOns page lists Python wrappers such as pytesseract and tesserocr. Tesseract reads images, so a PDF pipeline starts with a renderer such as pdftoppm, or with OCRmyPDF, which wraps Tesseract. The current release is 5.5.4, published on October 28, 2026.
Unsiloed is a set of HTTP endpoints that accept PDFs directly. The Parse API reference describes the flow: POST a file to /parse, receive a job_id, and fetch the result from GET /parse/{job_id} when the job completes. Extraction works the same way through /v2/extract and GET /extract/{job_id}, as described in the Extract API reference. Page rendering, layout detection, and model hosting all happen on the Unsiloed side.
Output: Words and Boxes vs Segments and Fields
The output format decides how much code sits between OCR and the system that consumes it.
Tesseract's TSV gives one row per recognized word, with its block, paragraph, line, and word numbers, a pixel bounding box, a confidence value, and the text. That is precise raw material for search and text capture. Turning it into a record, for example knowing that "$4,177.44" is the total due and "2026-10-16" is the ship date, is mapping code your team writes for each layout.
Unsiloed returns that structure directly, at two levels:
- Parse returns chunks, each with an
embedMarkdown string for retrieval, plus segments carrying asegment_type,markdown,html,bbox,page_number, and a 0 to 1 detectionconfidence, as described in the Parse response format. It suits RAG indexing, where reading order and table structure decide whether a chunk makes sense. - Extract takes a JSON Schema and returns values in that shape, such as an
invoice_numberstring and aline_itemsarray of objects. The guide to converting PDFs to structured JSON in Python walks through that path with validation.
Add business-rule checks on top of schema validation, such as line items summing to the subtotal.
Layout, Reading Order, and Tables
Tesseract handles layout through page segmentation modes. The command-line usage page lists mode 3, fully automatic page segmentation, as the default. Mode 6 treats the image as "a single uniform block of text", which suits a cropped paragraph or cell. For tables, the ImproveQuality page recommends custom segmentation or layout analysis, and it describes preprocessing steps, such as 300 DPI input, deskewing, and binarization, that improve recognition.
Unsiloed's Parse API runs layout detection as part of the request and returns each table as a dedicated segment in both Markdown and HTML, with merged headers preserved in the HTML. For tables that continue across pages, the merge_tables option combines table segments split by page breaks by matching headers and columns. The multi-page table guide covers repeated headers and continuation rows, and the roundup of layout-aware OCR solutions for complex documents covers the wider problem.
Source Citations and Confidence Scores
Tesseract's word confidence measures recognition: how sure the engine is about the characters in a word.
Unsiloed Extract scores each field instead. The job results reference shows each value with a score object containing a grounding_score and an extraction_score. With enable_citations=true, each grounded field also carries a citation with its page number and bounding box, so a reviewer can jump from a value straight to the region it came from. Turn citations on when you submit the job; they are off by default.
The same reference describes score bands, from "0.9-1.0: Very high confidence, extraction is very likely correct" down to "Below 0.6: Low confidence, likely needs manual verification." Use them as a starting point and set thresholds per field based on what a wrong value costs. The guide to confidence score thresholds for straight-through processing covers calibration, and the extraction audit trail guide covers what to keep for reviewers and auditors.
Benchmark Results
Unsiloed publishes its results on olmOCR-Bench, an Allen AI benchmark of 1,403 PDF pages covering tables, multi-column reading order, headers and footers, and old scans. In Unsiloed's own run with the unmodified upstream scorer, Unsiloed Parser scored 88.0, the top score among the 19 systems compared, and the harness and runner code are public. Tesseract was not included in that comparison.
To test on your own documents, build a sample that includes the layouts your pipeline sees most, such as multi-column filings, merged-cell tables, and low-resolution scans, and compare both pipelines on the same final output, one JSON object per document. Measure field exact match, table cells in the correct row and column, and missing values separately. The guide to evaluating document parsers on complex PDFs includes a worked example.
Pricing and Total Cost
Tesseract is free under the Apache 2.0 license. A production pipeline built on it includes PDF rendering, preprocessing, compute, the layout and field-mapping code, validation, queues and retries, and engine and traineddata upgrades.
As of November 2026, the Unsiloed pricing page lists three plans:
| Plan | Pricing |
|---|---|
| Starter | Free; pay as you go |
| Standard | Custom, usage-based pricing |
| Enterprise | Custom, volume-based pricing |
Standard adds production rate limits, webhook callbacks, zero data retention, and a HIPAA BAA. Enterprise adds VPC, on-premises, or air-gapped deployment, EU or US data residency, and a SOC 2 Type II report.
Compare cost per accepted document rather than cost per page. Count the engineering time to build and maintain the field layer, and the review time for documents that fail validation. Lower review volume can reduce cost per accepted document.
Deployment and Privacy
Tesseract runs wherever you install it, and your team operates the hosts, retention, access logging, and patching for the engine, traineddata, and renderer.
Unsiloed's hosted API runs the models, queues, and scaling for you, and Standard adds zero data retention. For data that must stay inside your network, the Enterprise plan offers VPC, on-premises, and air-gapped deployment of the same API. The guide to air-gapped document parsing and extraction covers what that setup involves.
When to Choose Unsiloed
Unsiloed is the stronger choice when:
- You need schema-shaped JSON rather than text.
- Your documents include multi-column pages, merged-cell or multi-page tables, or new layouts arriving often.
- Reviewers need the source page and bounding box for each extracted value.
- Documents must stay on your own infrastructure, through an Enterprise VPC, on-premises, or air-gapped deployment.
Unsiloed replaces much of the layout, table, and field-mapping code you would otherwise build around an OCR engine. You still need to validate the extracted fields. If your goal is a searchable PDF from clean printed scans, the guide to making scanned PDFs searchable covers that workflow.
To run your own documents through Unsiloed, create a free account. Check the pricing page for current plan terms.
FAQ
Does Tesseract return bounding boxes for extracted values?
Tesseract's TSV and hOCR output include a pixel bounding box and a confidence value for every recognized word. Linking a business field such as an invoice total to its page region takes your own mapping code. Unsiloed Extract returns a page number and bounding box for each grounded field when enable_citations is set to true.
Can Tesseract read PDF files directly?
No. Tesseract's input formats page states that it does not support reading PDF files. Render each page to an image first, for example with pdftoppm -r 300, or use a wrapper such as OCRmyPDF. Unsiloed accepts PDFs directly.
Which is better for scanned PDFs with tables and multi-column layouts?
Unsiloed. Its Parse API returns each layout block as its own segment and each table as HTML and Markdown with merged headers preserved, and Extract returns the values as schema-shaped JSON with citations.
Is Tesseract free to use commercially?
Tesseract is licensed under Apache 2.0, which permits commercial use. A production deployment still needs compute, preprocessing, and the engineering time to build field extraction, validation, and review around its text output.



