Skip to content
Page
Build or buy
Audience
Teams with engineers
Topic
LLM document extraction
Areas of work
6

“We’ll just build this in-house with an LLM”

Maybe you will. A prototype reads one document; production has to read every document, check the result and deliver it, every day.

01
The work
Between prototype and production

What an in-house build has to cover.

Six areas of work sit between a prototype and a service. For each one: what the work is, and what Parser already does.

  • 1.1

    Intake

    Accept files from people and from systems: an upload screen, an e-mail address, an API. Cope with the formats you actually receive, with files that will not open, and with the queue and the retries behind all of it.

    In ParserUpload, e-mail or API intake, on every plan.

  • 1.2

    Extraction

    Turn a document into the fields you want: a prompt for each document type, the model’s answer parsed into a fixed shape, tables that run across pages, files longer than the model accepts.

    In ParserYou name the fields in an extraction template. Documents are read whole, in one pass: up to 1,000 pages on Pro, no limit on Enterprise.

  • 1.3

    Validation

    Check values before they reach another system: a total against its line items, a tax ID, a date that makes sense. Every rule is code, and every rule needs tests.

    In ParserValidation rules check every field against what you expect, on every plan.

  • 1.4

    Review

    Decide which values a person has to look at, show them, take the correction and put it where the data goes. That is a queue, a screen, permissions, and a way to feed corrections back.

    In ParserA result can be updated through the API, for example with corrected data.

  • 1.5

    Delivery

    Send results where they are used: webhooks with retries and logs, an API to fetch them, exports to JSON, Excel or CSV.

    In ParserWebhook or REST API, and JSON, Excel and CSV output, on every plan. Webhook logs are available through the API.

  • 1.6

    Operations

    Keep it working: monitoring by document type, alerts, a re-test against your whole document set at every model upgrade, rate limits, fallbacks between providers, and the model bill.

    In ParserParser runs the service. Usage is in the API, and results include a model_version field. Enterprise can bring its own LLM key.

When to build, when to buy.

02
Decision
Build or buy
Conditions under which building document extraction in-house or buying it makes sense
Look atBuild in-house whenBuy when
What extraction is to you
It is your product: you sell it.
It supports your business; it is not what you sell.
Your documents
One stable layout that you control.
Many layouts, or parties whose layouts you do not control.
Your engineers
An ML team with room on its roadmap to own this.
Their time is committed to your own product.

Cost depends on your documents, your team and your volume; the demo is where to test it.

03
Questions
What teams ask before deciding

Questions from teams weighing a build.

  • A credit is the unit Parser bills in. One credit is about what a typical one-page invoice costs, and every extraction costs at least 1. Beyond that, an extraction is charged on the work the model actually did for it: a heavier document uses more credits and a light one uses the minimum. A 50-page text-heavy contract can cost about as little as an invoice, and a dense table costs more. Building a template with AI also uses credits (minimum 1). Free includes 10 credits a month, Starter 125 and Pro 1,500. The console shows your balance and the cost of each document.

  • New documents wait until you top up or the month renews. Plan credits renew every month and do not roll over, and top-up credits are used after the plan credits. A top-up is a one-off pack of $10, $25, $50 or $100; Starter adds a +25% bonus and Pro +50%. Nothing else is billed: the only charges are your subscription and the top-ups you choose to buy. Enterprise runs on your own model key, so it has no credits to run out.

  • Parser reads the whole document at once instead of cutting it into small pieces. Fields that depend on distant pages, like a total on page 900 that must match a line item on page 12, are extracted and validated together. The limit is pages per document: 10 on Free, 100 on Starter, 1,000 on Pro and no limit on Enterprise.

  • Yes, on Enterprise: it runs on your own LLM key (BYOK), so there are no credits to buy, and it is priced by contract. Free, Starter and Pro do not include it.

  • Cloudflare provides hosting, storage and security infrastructure, and Google (Gemini) provides the AI models used for document extraction; the privacy policy lists both.

    Documents are deleted after 7 days. Extracted data is kept until you delete it or close your account. Your documents and extracted data are never used to train AI models. A DPA is available on request, and Enterprise includes one. The Security page lists every control and where it is documented.

  • No. Free includes 10 credits a month: sign up, send documents, and upgrade or top up only if you outgrow them.

Bring your case to the demo.

In the demo we process your documents live — no slides, no fluff.

30 minutes · no commitment

  • Self-serve

    Free, Starter ($9.90/month) and Pro ($100/month) start when you sign up. No card required for Free.

  • Enterprise

    By contract · your own model key · an engineer for your team · setup done for you if you want it · POC on your documents.