Chunkr
Document parsing and extraction service, run from their cloud or self-hosted from the open-source release.
Jobs Extract
Operated by Lumina AI, Inc. (d/b/a Chunkr)
Opens the vendor's own site.
Jobs covered
Open a task to see exactly what the vendor documentation supports.
Extract
Browse all extract toolsextracting structured data from documents with the Chunkr Extract feature, filling a JSON Schema you define and returning the result in that schema shape with a source citation and a confidence score for each extracted value, accepting a raw document from a URL, file upload or base64, or an existing parse task
Where your files are processed
Mixed - depends on version and feature
The vendor's own documentation separates them. On the device: The AGPL release. On their servers: The managed API. A single answer for the whole product would be wrong for one of them. That is a property of the library itself: what an application built with it does with your file depends on that application, not on this.
Genuinely two products with two answers, so left null rather than guessed. The managed API is a hosted service - documents are uploaded and processed on Chunkr's infrastructure, which the security page describes as SOC 2 and HIPAA compliant with "customizable zero data retention by default" and deletion on an expiration the customer sets.
The AGPL release is deployed by the customer with Docker Compose and can be run entirely inside their own infrastructure, and on-premises deployment is additionally offered to enterprise customers. uploadEndpointHost is deliberately null: setting it would label the self-hosted path as cloud-only, which is the mistake the schema's note on that field warns about. The honest reading is that locality here is a deployment choice, not a product property.
What it does
Chunkr turns documents into data for whatever comes next: the Parse API returns HTML or Markdown with bounding boxes for every element, and the Extract API returns structured JSON against a schema you supply, with a citation back to the region each value came from and a confidence score attached. The engineering bet is on vision models rather than text extraction - reading-order detection, layout segmentation, handwriting, formulas and complex tables are all treated as vision problems - and the output is shaped for retrieval pipelines rather than for human reading.
It can be used as a managed API with published monthly pricing, or self-hosted from an AGPL-3.0 release, though the two do not run the same models, which is the distinction that matters most on this record.
Best for
Developers building document-heavy pipelines who need machine-readable output with provenance: RAG ingestion, form and invoice extraction, and any workflow where a downstream system has to trust a value and be able to point back to where it came from. The confidence scores and per-field citations are the features that make it usable in a pipeline with a human review step, and the self-hosted release is there for teams that cannot send documents out.
Worth knowing
The open-source release and the paid API are not the same product. The repository states it plainly: the AGPL version uses community and open-source models while the cloud API runs proprietary in-house models "for higher accuracy, speed, and enterprise reliability".
Anyone who evaluates the self-hosted version and buys on that basis would be buying different accuracy. The free allowance is described inconsistently - the pricing page says 200 credits, while the home page and the page's own structured data say 200 pages - and the two are only the same for PDFs, because a spreadsheet is charged by the cell rather than the page and vision-model processing costs double credits.
Pricing is credit-based and scales with volume: the cheapest published plan is $375 a month, which is a commitment rather than a trial, and on-premises deployment is an enterprise conversation rather than a download. Finally, this is extraction software: it reads documents and returns data, and it does not write, modify or redact a PDF at all.
The hosted API also has a published end date. Chunkr's documentation states that Chunkr is being acquired by Extend, that the hosted Chunkr API stays fully live through 14 November 2026 and shuts down on 15 November 2026, and that a migration guide covers moving a Parse or Extract integration to Extend.
That shutdown applies to the hosted API only: the same documentation says the open-source version stays up, remains available under its existing licence, and needs nothing done if it is run with Docker Compose, so a self-hosted deployment is not affected. Anyone pricing the monthly plans above should read the migration guide first, because an API key does not carry over to Extend and documents and task outputs are not migrated, they have to be retrieved before the shutdown date.
Details, and where each comes from
Vendor statements, our own reading of their pages, and the things nobody has established yet, each answered separately. Where we have not confirmed something it says so rather than guessing, and a row nobody has established is greyed out.
- Versions covered
- API · Self-hosted
- Website
- chunkr.ai
- Requires an account
- Not applicable
- Free tier
- Yes
- Watermark on the free tier
- No
- Maximum file size
- Not verified
- Batch processing
- Yes
- Data residency
- Not verified
- Open source
- Yes
- Self-hostable
- Yes
- Licence
-
AGPL-3.0 for the self-hosted release; the managed cloud API is proprietary.
AGPL-3.0 for the self-hosted release; the managed cloud API is proprietary. The repository states that the open-source version uses community models while the cloud runs in-house models.
- Redaction
- No redaction feature
- Programmatic access
- HTTP API · Embeddable library
- Last availability check
- Responded in 130 ms 21 October 2026
Checked automatically by our own runner, not reported by the vendor. 100% of our last 30 checks got a response.
What the redaction verdict rests on
From the vendor's own documentation, read 22 September 2026. Not tested by us.
What that does not establish
Recorded as `none` structurally rather than on absence: this is a reader. It parses documents into HTML, Markdown or JSON and extracts values with citations; nothing in the feature list, the API surface or the pricing model describes writing to a document or removing anything from one.
The confidentiality question a reader should actually ask about this tool is different and is not answered by this verdict - it is what happens to the documents that are uploaded, which is why processesLocally and the zero-retention claim are recorded in the fields above.
What the vendor says
Claims are labelled rather than repeated as fact. Where we have checked something ourselves, it says so and shows when. Checked means we read the vendor's own page and it says what we recorded - not that we tested the behaviour, unless the note says we did.
-
“The open-source AGPL version is different from our fully managed Cloud API. The open-source release uses community/open-source models, while the Cloud API runs proprietary in-house models for higher accuracy, speed, and enterprise reliability.”
Checked by us - the source says this Read 22 September 2026
Confirmed verbatim in the repository README, and corroborated by the licence file, whose header is the GNU Affero General Public License version 3, 19 November 2007 - read from the licence itself rather than inferred from GitHub's sidebar label. This is the most decision-relevant sentence the vendor publishes, because the free self-hosted path and the paid path are not the same models, so an accuracy evaluation of one does not transfer to the other.
Recorded as confirmed in the sense that the vendor says it clearly and the licence matches; whether the cloud models are in fact more accurate is a separate question nobody should infer from this.
Standing: Re-read and matched byte for byte
Check it yourself: the source, the SHA-256, and how we checked
- From
- the text a browser renders
- SHA-256
- 78c4bb673a89cc773c414098807c85d85ca51cf7b129a6ff6398c18edfc1f88c
How we checked
https://github.com/lumina-ai-inc/chunkr rendered in headless Chrome with scripts/render-page.mjs on 2026-09-22 (visible text 6279 bytes); the licence itself fetched from raw.githubusercontent.com and read as a file (34,523 bytes, SHA-256 8486a10c4393cee1c25392769ddd3b2d6c242d6ec7928e1414efff7dfb2f07ef). sourceHash is the SHA-256 of the rendered repository text.
-
“We are SOC2 and HIPAA compliant with customizable zero data retention by default. Your documents are processed securely and deleted immediately based on the expiration time you set.”
Checked by us - could not be settled Read 22 September 2026
Present verbatim in both the FAQ and the structured data the page publishes for search engines, and the security section adds SOC 2 Type II and Type I, HIPAA with a BAA available, "0 Data Retention" and 99.99% uptime, each pointing at a trust centre at trust.chunkr.ai. Recorded as unverifiable rather than confirmed because a compliance certification is precisely the sort of claim a vendor page cannot establish on its own: the trust centre is where the artifacts would be, and it was not opened.
The operational part of the claim is the useful part and it is testable by a customer - deletion on a customer-set expiration - and that is a setting to look for in the account rather than a statement to take on trust.
Standing: Re-read and matched byte for byte
Check it yourself: the source, the SHA-256, and how we checked
- From
- the text a browser renders
- SHA-256
- 99da36e526ca42ac482e4e6f25cbbb96ccd87daa4a377cc8815108c8ed77ba96
How we checked
https://www.chunkr.ai/ rendered in headless Chrome with scripts/render-page.mjs on 2026-09-22 (visible text 4249 bytes), including the FAQ answers embedded as FAQPage structured data; the security section and its links to a trust centre were read on the same page. sourceHash is the SHA-256 of that rendered text.
-
“order-1 State-of-the-art reading order detection that achieves 95% accuracy, approaching human-level performance.”
Checked by us - could not be settled Read 22 September 2026
Present verbatim, and the page is unusually good about showing its working: a bar chart puts order-1 at 95% against TBLR at 73.1%, XY-Cut at 68.4%, Column-aware at 65.9% and LayoutLMv3 at 40%, with the axis labelled "Baseline: 100% = human-tagged reading order". That last detail is the one that matters, and it is easy to skim past: the benchmark measures agreement with human-tagged reading order on the vendor's own data, not correctness on a customer's documents.
Recorded as unverifiable rather than confirmed because "state-of-the-art" is a claim about the field at a moment in time, and the comparison set is the vendor's. The chart is nevertheless more disclosure than most vendors in this index offer, and it is recorded as such.
Standing: Re-read and matched byte for byte
Check it yourself: the source, the SHA-256, and how we checked
- From
- the text a browser renders
- SHA-256
- 99da36e526ca42ac482e4e6f25cbbb96ccd87daa4a377cc8815108c8ed77ba96
How we checked
Same rendered page and hash as the claim above; the claim and the comparison chart beside it were read from the page, including the axis label.
-
“Start for free with 200 credits.”
Checked by us - the source only partly supports this Read 22 September 2026
The two pages disagree about the unit and both are recorded. The pricing page says "Start for free with 200 credits", the home page FAQ says "We provide 200 pages for free to get started", and the structured data on the home page encodes it as an offer of "200 pages free to get started".
They coincide for PDFs, where the rate card charges 1 credit per page, and diverge everywhere else: a spreadsheet is 1 credit per 5,000 cells and anything processed with a vision model costs double. So "200 pages" is the marketing phrasing and "200 credits" is the accounting, and a reader testing with spreadsheets or VLM processing will get less than the home page implies.
Marked partial for that reason rather than confirmed.
Standing: Re-read and matched byte for byte
Check it yourself: the source, the SHA-256, and how we checked
- From
- the text a browser renders
- SHA-256
- 181bd5b68bf4ff23ee8d4bb3ab42933df555c6a40551e2dd7231a00fecce8f55
How we checked
https://www.chunkr.ai/pricing rendered in headless Chrome with scripts/render-page.mjs on 2026-09-22 (visible text 874 bytes - small, because the page is mostly a table) and read together with the free-trial answer on https://www.chunkr.ai/ (4249 bytes) and the structured data both pages carry. sourceHash is the SHA-256 of the rendered pricing text.
Each claim is quoted from the vendor's own pages. This section describes what we have and have not checked - it is not a judgement about anyone.
Sponsorship is distribution, not an endorsement or a promise of search rankings, AI citations, traffic, or sales. This listing is a free entry, compiled from published sources.
Related reading
- Open source PDF tools, and what their licences require - what open source does and does not tell you.