Self-hosted

Invoice capture and extraction

Invoices arrive by email or watched folder, Docling converts them (OCR included), a local vision-capable model fills a JSON schema under constrained decoding, code checks the arithmetic and the supplier against your master data, and n8n posts the clean ones while routing the rest to a person.

Save to your workspace

Source: Editorial recipe — no step evidence has been fetched yet
Verified: Evidence not verified
Confidence: Low

01Objective

Turn incoming supplier invoices — PDF, scan or email attachment — into validated structured records in the accounting system, with a human queue for anything uncertain.

Reference deployments using this recipe

Invoice capture and coding for an SME finance teamA four-person finance function re-typing about 700 supplier invoices a month from PDFs and photographs. A self-hosted extraction pipeline that proposes the fields and the coding, and a person who approves every one of them before anything is paid.Reference deployment · R-04

Field notes about this recipe

Observe the workflow before choosing the technology

02Recommended stack

6 components

Role	Component	Notes
Inference	Mistral-Small-3.2-24B-Instruct or Gemma-3-12B-IT (vision-capable)	A model that can look at the page helps on stamped, rotated and photographed invoices where text extraction alone loses the layout.
Inference	vLLM with structured outputs	Constrained decoding so every extraction is the same shape and can be validated.
Ingestion	Docling	Conversion and OCR, with table structure preserved — line items live in the tables.
Ingestionoptional	Paperless-ngx	Optional front door: mail fetching, a consume folder, tagging and an archive of the originals. Useful when the invoices also need to be findable later.
Orchestration	n8n	The workflow: trigger, convert, extract, validate, branch to post or to the exception queue, retry, alert. Visible to finance, which matters when they own the exceptions.
Storage	PostgreSQL	Extractions, validation results, the audit trail and the exception queue state.

Architecture and data flow

External data transfer · SOME

confidential content leaves your premises for your own cloud tenancy ("Accounting system"). You keep control of the account; the provider is a processor, so a DPA and a documented region apply.
Invoices are read and extracted entirely on your own machines; the only outbound flow is the posting into your own accounting system.
If the accounting system is itself a SaaS product, that link inherits its terms — worth stating explicitly when the design is reviewed.

03Suitable for

Organisation size: 10–1000 employees
Data classes: confidential, personal
Constraints: enough invoice volume for automation to pay back — typically 200+ per month; an accounting system with an API or a supported import format; someone in finance who owns the exception queue; a 24 GB GPU; a CPU-only pilot is possible at low volume
Industries: Accounting, Professional services, Manufacturing, Logistics, Retail
Jurisdictions: any

04Hardware

On-premise single 24 GB GPU server
GPU
NVIDIA RTX 4090 24 GB (or NVIDIA L4 24 GB for a rack-mounted, 72 W alternative)
VRAM
24 GB
System RAM
64 GB
Storage
2000 GB
CPU
16-core x86 server CPU (AMD EPYC 7003/9004 or Intel Xeon Scalable)
Form factor
Tower server
Indicative costUS$4,000 – US$9,000
indicative build cost for the complete machine, USD, Aug 2026 — verify with a local supplier
The default box for a 20–60 person firm. Fits a 14B model at 4-bit with roughly 8 GB of KV cache left for concurrent chat, or an 8B model at fp16. NVMe storage sized for the model cache plus a document corpus and its embeddings. Add a UPS and an offsite backup target — this machine holds the whole knowledge base.
CPU-only server (small models and embeddings)
GPU
unknown
VRAM
unknown
System RAM
64 GB
Storage
1000 GB
CPU
16–32 core x86 server CPU with AVX-512
Form factor
Tower server
Indicative costUS$1,500 – US$4,000
indicative build cost for the complete machine, USD, Aug 2026 — verify with a local supplier
No GPU. Embedding models and document parsing run acceptably here, which is enough for a search-only pilot or a nightly batch pipeline. Chat generation with a 7–8B model at 4-bit works but reads at a few tokens per second — usable for one person testing, not for a team. The honest use of this profile is to prove the retrieval quality before buying a GPU.

Indicative costUSD · one-off plus monthly

Software licences: All components are open source and self-installed. Check each licence for use limits.; US$0
Model usage: Open weights served locally: no per-token charge. The GPU cost sits under infrastructure.; US$0
GPU server (24 GB class), purchased: Hardware profile onprem-small-24gb — indicative build cost, Aug 2026, verify locally.; US$4,000 – US$9,000
Implementation (8–20 FDE-days): 8–20 FDE-days at US$760–1940 per day, converted from the HK$6,000–15,000 band at the HKMA Linked Exchange Rate band of HK$7.75–7.85 to one US dollar. One-off; excludes internal staff time.; US$6,080 – US$38,800

200–5,000 invoices per month, one accounting system, one exception queue.
The implementation band widens sharply if the accounting system has no usable API.
Excludes the finance time spent on exceptions, which dominates the running cost until the straight-through rate is high.
Every figure is an assessment from the inputs listed, not a quotation.

05Difficulty

4 / 5

Multiple weeks and a team that has shipped infrastructure

06Skills

API integrationapi-integration: development
Data engineeringdata-engineering: data
Dockerdocker: infrastructure
Linuxlinux: operations
LLM evaluationllm-evaluation: ml
LLM inferencellm-inference: ml
OCR and document parsingocr: ml
PostgreSQLpostgresql: data
Pythonpython: development
Workflow automationworkflow-automation: operations

07Deployment steps

7 steps

Commands are copied from each project’s own documentation, and the page they came from is linked under the step. 0 of 7 steps currently open an evidence record. The rest are linked to their source; §10 says which of those documents were fetched and which were fetched without their anchor being found — two different states, named differently there.

01
Collect 200 real invoices before writing anything
Assessment
Take a genuine month, including the ugly ones: photographed, multi-page, foreign currency, credit notes, handwritten annotations. Bucket them by supplier. Ten suppliers usually make up most of the volume — automating those ten well beats automating everything badly.
02
Convert and OCR with Doclingversion-sensitive
Assessment
Install Docling and convert the sample. Check line-item tables specifically: if the table structure survives conversion, extraction is straightforward; if it does not, no amount of prompting will recover it.
```
pip install docling
```
```
docling https://arxiv.org/pdf/2206.01062
```
from the Docling README
Source documentation
03
Write the invoice schema and extract against itversion-sensitive
Assessment
Fields: supplier name and tax id, invoice number, dates, currency, line items, net, tax, gross, purchase-order reference. vLLM "supports the generation of structured outputs using xgrammar or guidance as backends" — use the schema so every result parses. Require a confidence and a source quote per field.
Source documentation
04
Validate in code, never in the model
Assessment
Line items must sum to net; net plus tax must equal gross; the tax rate must be one your jurisdiction allows; the supplier must match your master data; the invoice number must not already exist. Arithmetic belongs in code — a language model is the wrong tool for a sum, and a failed check is the cheapest possible error detector.
05
Wire the workflow in n8nversion-sensitive
Assessment
Trigger on the mailbox or consume folder, call conversion and extraction, run validation, branch: clean invoices post to the accounting system, anything that fails a check or falls below the confidence threshold goes to the exception queue with the reason attached. Add retries and an alert when the queue stops draining.
```
docker volume create n8n_data

docker run -it --rm \
 --name n8n \
 -p 5678:5678 \
 -e GENERIC_TIMEZONE="<YOUR_TIMEZONE>" \
 -e TZ="<YOUR_TIMEZONE>" \
 -e N8N_ENFORCE_SETTINGS_FILE_PERMISSIONS=true \
 -e N8N_RUNNERS_ENABLED=true \
 -v n8n_data:/home/node/.n8n \
 docker.n8n.io/n8nio/n8n
```
from the n8n docs — set your timezone
Source documentation
06
Keep the original next to the extraction
Assessment
Store the source file and the extraction together, with the model and prompt version. Paperless-ngx does this well if you also need the invoices searchable; a folder plus a database row is enough if you do not. Either way, an extraction you cannot trace back to an image is not auditable.
Source documentation
07
Run in shadow mode and measure the exception rate
Assessment
For the first month, extract everything but post nothing: compare against what finance keyed manually. Report straight-through rate and the reasons for exceptions by supplier. Go live per supplier as each one clears the bar, not all at once.

08Compliance considerations

Structured issue-spotting to support your own review — not legal advice. Verify against the cited primary sources and your counsel.

Applies everywhere

Personal data · Retentionmedium
Invoices carry personal data (contacts, bank details, sometimes employee names) and are usually subject to a statutory accounting retention period. That period governs the archive; it does not automatically justify keeping every intermediate extraction.
Sector rules · Auditabilityhigh
Tax and audit rules in most jurisdictions require the original document to remain available and unaltered. Keep the source file, not just the extracted values, and record who approved each posting.
Human oversight · Automated decision-makinghigh
Automatic posting is an automated decision with a financial effect. Set a value threshold above which a person always approves, and make the exception queue someone's named job.
Security · Confidentialityhigh
Supplier bank details are a fraud target. The mailbox, the pipeline and the accounting credentials all need to be treated as payment infrastructure, including change control on bank-detail updates.

09Alternatives

A specialist capture vendor
Purpose-built invoice capture arrives pre-trained on millions of invoices and ships the exception UI. Usually more accurate on day one than anything you build in a fortnight.
- — Accurate immediately, no model to run
- — Invoices go to a processor: DPA, residency and retention questions
- — Per-document pricing that scales with volume
rossum nanonets
Your accounting system's own capture feature
Most accounting platforms now include capture. If it covers your top ten suppliers, the integration work disappears entirely.
- — Nothing to integrate
- — Little control over accuracy or exception handling
- — Data goes wherever the accounting vendor processes it

No alternative recipe is published yet.

10Evidence

0 of 6 fetched

Not yet in the ledger

Docling — README (quickstart and conversion)
Readme
not yet fetched
vLLM — Structured Outputs
Docs
not yet fetched
vLLM — Using Docker
Docs
not yet fetched
n8n — Install with Docker
Docs
not yet fetched
Paperless-ngx — Setup (installation)
Docs
not yet fetched
pgvector — README (installation and getting started)
Readme
not yet fetched

11Community

Deployed this stack, or hit something this page does not cover? Corrections, sources and implementation reports are what keep a recipe worth reading.

Improve this page

From the field

0 deployments · 0 questions

Nobody has reported deploying this here yet, and no question has been opened against this page. Both appear once a reviewer accepts them.

12Hire an FDE

If you would rather not build it, we can introduce a forward-deployed engineer who has deployed this stack before. The enquiry form starts from this recipe.

Hire an FDE for this stack

01Objective

Reference deployments using this recipe

Field notes about this recipe

02Recommended stack

Architecture and data flow

03Suitable for

04Hardware

On-premise single 24 GB GPU server

CPU-only server (small models and embeddings)

Indicative cost·USD · one-off plus monthly

05Difficulty

06Skills

07Deployment steps

Collect 200 real invoices before writing anything

Convert and OCR with Doclingversion-sensitive

Write the invoice schema and extract against itversion-sensitive

Validate in code, never in the model

Wire the workflow in n8nversion-sensitive

Keep the original next to the extraction

Run in shadow mode and measure the exception rate