01
Thousands of documents a month
An LLM on its own
An LLM waits to be asked. Someone pastes in every single file.
With Mango
Mango takes an API call or a watched folder and runs unattended.
The document platform
Any model can read a document. Mango turns documents into a system.
Intake, splitting, classification, tagging, extraction, review, delivery. Thousands of documents, unattended.
Running in production for finance, operations and supply-chain teams.
01The argument
Why not just use an LLM?
You've probably tried. It worked — models read documents well, and we use them too. The problem starts at document two hundred.
01 / 08
01
Thousands of documents a month
An LLM on its own
An LLM waits to be asked. Someone pastes in every single file.
With Mango
Mango takes an API call or a watched folder and runs unattended.
02
A 400-page batch, 60 documents inside
An LLM on its own
An LLM sees one wall of text, with no document boundaries in it.
With Mango
Mango splits the batch, classifies each document and routes it onward.
03
The same output shape, every time
An LLM on its own
Prompt an LLM twice and the shape can change. So can an upgrade.
With Mango
Mango extracts against a schema you define, version and control.
04
Knowing which fields to check
An LLM on its own
An LLM sounds exactly as certain when it is wrong as when it is right.
With Mango
Mango scores each field and sends only the doubtful ones to review.
05
Keeping the source document
An LLM on its own
The file leaves with the chat. Months later there is nothing to check.
With Mango
Mango keeps the original beside the data, one click from any value.
06
Proving where a number came from
An LLM on its own
An LLM hands back a value with no citation standing behind it.
With Mango
Every Mango field carries its document, page and schema version.
07
Feeding several systems at once
An LLM on its own
Each team re-uploads the same PDF and pays an LLM to read it again.
With Mango
Mango reads once and serves that one result to every system.
08
Wiring it into production
An LLM on its own
A chat window is not an integration. Output gets copied out by hand.
With Mango
Mango is a versioned REST API, with webhooks, retries and keys.
02The platform
Everything between the file arriving and your system acting on it
Takes in whatever arrives
Scanner batches, attachments, folder drops, API uploads. The original is stored unchanged and referenced from then on.
PDFMAILAPI
IN
3 input routesOriginal preserved
Splits a batch into documents
One 400-page scan is not one document. Mango finds the boundaries and gives each its own identity.
1 file · 400 pages5 documents found
Classifies and tags on arrival
Type identified and attributes tagged before anyone looks. That's the difference between a queue and a pile.
INVOICEVENDORAPPROVED
Type detected3 tags applied
Extracts against your schema
Fields, tables and line items in the same shape this month, next quarter, and after the model changes.
TOTAL$4,280
DATE08/20
PO1842
Schema v123 fields mapped
Puts a person where one is needed
Machine output is never edited. Corrections sit alongside it, so both answers stay permanent.
F0198%
F0262%
F0396%
1 field flaggedHuman routed
Hands it to your systems
One versioned REST API. Build accounts payable, expense management, or something we haven't thought of.
DATA
APIERPS3
Structured output3 destinations connected
03Confidence & review
A confidently wrong model is worse than one that says it isn't sure
A model is exactly as fluent when it's wrong as when it's right. So a person ends up checking every field of every document — the work you were removing.
Mango scores each result. Below your threshold it becomes review work, shown beside the source. Above it, nobody ever sees it.
Review threshold
Minimum confidence to auto-approve
7/10 fields auto-approved
INV-2389.pdf
below 90% → review
Invoice number
auto-approvedINV-2389
99%Invoice date
auto-approved24 May 2025
98%Supplier name
auto-approvedABC Pvt Ltd.
96%Supplier GSTIN
auto-approved27AABCU9603R1ZM
93%Place of supply
auto-approvedMaharashtra
90%Taxable value
auto-approved38,830.51
97%IGST
auto-approved6,989.49
92%Line 3 — description
needs reviewM8 hex bolt, zinc
78%Line 3 — quantity
needs review1,200
71%Handwritten note
needs reviewshort shipped 40
52%
04Validation
Checks with a definite answer
Some questions about a document have exactly one right answer, and a language model is the wrong tool for all of them.
- Line items reconciled against header totals
- Tax recomputed against taxable value
- Tax and registration identifiers validated per jurisdiction
- Duplicates matched against what you've already processed
- Currency normalised, rate and source recorded
- Required fields present, formats correct, dates coherent
AI extracts, suggests and explains. It never decides the number.
05What teams build on it
One platform. Every process that starts with a document.
Most tools are one application with extraction underneath. Mango is the other way round.
Supplier invoices in, structured data out. Reviewer attention only on what the machine flagged.
06Developers
One API instead of a document pipeline you maintain
Every team handling documents builds the same stack: intake, dedupe, splitting, classification, an extraction call, a review screen, retries, somewhere to keep the original. That's Mango.
- Versioned REST API, OpenAPI spec, generated clients
- Schema-bound output that doesn't drift with the model
- Tenant and workspace isolation on every path
curl https://api.getmangoai.com/v1/documents \
-H "Authorization: Bearer $MANGO_API_KEY" \
-F "file=@batch-2389.pdf" \
-F "schema=supplier_invoice@3"200 OK
{
"id": "doc_01HZX9Q2",
"type": "supplier_invoice",
"schema": "supplier_invoice@3",
"source": "file_01HZX9PK",
"fields": {
"invoice_number": { "value": "INV-2389", "confidence": 0.99 },
"total": { "value": 45820.00, "confidence": 0.97 },
"line_3.quantity":{ "value": 1200, "confidence": 0.71 }
},
"status": "needs_review"
}07Provenance
When someone asks where a number came from
Any value Mango produced traces back to the original file, the machine result before anyone touched it, every correction with its author, and the version your systems consumed.
Tax invoiceSupplier
ABC Pvt Ltd.
14 Residency Road, Bengaluru
GSTIN 29AAECA1234F1Z7
Invoice number
INV-2389
24 May 2026
Bill to
Prectech Solutions Pvt Ltd.
Due 07 Jun
DescriptionQtyRateAmount
Data capture service17,6007,600
Validation setup19,4209,420
Archive processing1,2002428,800
Payment details
HDFC Bank ···· 4821
IFSC HDFC0001234
Subtotal₹45,820.00
Tax₹0.00
Total₹45,820.00
Thank you for your business. Please quote the invoice number with your payment.
Processed by Mango
08Security & data control
Your documents, under your terms
Business documents carry commercial terms, bank details and pricing. Consumer AI tools process them under consumer terms, with retention you can't inspect.
- Isolation
- Tenant and workspace boundaries on every access path, verified by automated tests
- Encryption
- In transit and at rest, with time-limited signed file access
- Secrets
- Held outside application configuration, with rotation
- Retention & deletion
- Defined windows, legal hold, deletion covering derived files
- Residency
- Processing pinned to your region where policy requires it
- Provider control
- Extraction providers configurable per tenant, no-retention honoured
- Audit
- Actor, action, target, time recorded for every material action
- Access control
- Role-based permissions, scoped API keys and least-privilege access
09Customers
Teams already running on it
“The API took an afternoon to integrate. Two weeks later we retired an entire manual data-entry vendor.”
Karan Mehta
Engineering Lead, Firefly Batteries
“Handwritten delivery challans from depots were always the bottleneck. Mango reads them accurately.”
Rohit Deshmukh
VP Supply Chain, OM Infra Solutions
“We process documents in six regional languages. The confidence scoring tells us which fields need human review.”
Ananya Iyer
Operations Manager, Mahitri Technocrats
“Audit prep is no longer a scramble. Every extracted field has a trail back to the source document.”
Neha Kulkarni
Financial Controller, Citelum India
10Pricing
Priced around your volume and your documents
No public price list, because no two document operations look alike. What drives the number is how many documents, how many types, and what you connect Mango to.
- Unlimited Document Types and Schemas
- Unlimited Users and Reviewers
- Full API Access
- Original Document Retention
- Human Review
- A Named Implementation Contact
11FAQ
Questions teams ask before they trust this
Product
For one document, you should. A model on its own can't take a scanner batch and find where each document ends, run unattended, hold its output shape across model versions, tell you what it was unsure about, or keep the original. We use models too — we just don't leave them unsupervised.
No. Teams run accounts payable, expense management and payment posting on it, and several use the API to build something specific to their business. Any document type you define a schema for.
Nothing breaks. Extraction is bound to a schema you version, not to positions on a page. Where confidence drops, results route to review, and your integration keeps receiving the same shape.
Yes — this is what Mango is for. A 400-page batch is split into the documents inside it, each classified, tagged and extracted independently.
Yes, including mixed scripts in one document and phone photos taken at an angle. Handwriting is exactly where per-result confidence matters most.
A versioned REST API with a documented contract and generated clients. Webhooks and Tally/SAP posting are in progress.
Data & trust
Retention, deletion and residency are contractual, not defaults. Processing can be pinned to your region, and deletion covers derived files as well as originals.
Documents and extracted data are encrypted both in transit and at rest. Processing is isolated, file access is time-limited, and credentials are managed separately and rotated regularly.
Yes. Each extracted field keeps a reference to the source document and the region it came from, so reviewers and auditors can verify the value without hunting through the original file.
No. Anything with one correct answer — arithmetic, identifier validity, duplicate detection — is computed by deterministic rules that return the same result every time.
Tell us about your documents and we'll set up your workspace, usually within a working day. You run your own files, not a sample pack. Pricing comes after.
It surfaces as a visible, recoverable failure — never a silent drop. Nothing is lost, because the original is stored unchanged the moment it arrives.