Parseloom ships with built-in templates for the most common document types: standard invoices, bills of lading from major carriers, and a few others. For everything else, you build a custom template. This article walks through the process.
What a template actually defines
A Parseloom template tells the extraction engine two things: what fields to look for, and what data type each field contains. The engine's job is to figure out where on the document those fields appear and extract the value. You don't define positions or regions manually - the model learns from examples.
A minimal template definition looks like this:
template_name: customs_entry_import
fields:
- name: entry_number
type: text
- name: importer_name
type: text
- name: country_of_origin
type: text
- name: total_dutiable_value
type: currency
- name: duty_rate
type: percentage
- name: estimated_duty
type: currency
Training with examples
After you define the field schema, you upload example documents and annotate them: highlight the text region that corresponds to each field and confirm the value. You need at least 5 annotated examples for the model to start making reasonable extractions. The more layout variation in your examples, the better the model handles real-world documents from multiple sources.
The template builder in the Parseloom web interface makes annotation straightforward: draw a box around the field value in the document viewer, select the field name from the dropdown, and confirm. For each example, the model updates its understanding of where that field tends to appear relative to surrounding text and layout cues.
What makes a good set of examples
Diversity of source. If all five of your annotation examples came from the same vendor, the model will be good at that vendor's layout and uncertain about others. Try to include examples from at least 3-4 different sources in your initial training set.
Edge cases matter more than perfect examples. Include a document where the field you care about appears in an unusual position, or where the label uses slightly different text than usual. A model trained only on clean examples will fail more on real-world variation.
Don't over-annotate. Annotate the fields you actually need in your ERP import. Extra fields increase the chance of a mis-extraction pulling in an unrelated value. Keep the template focused.
Testing and iteration
After initial training, run a batch of test documents that you've manually verified. The batch test report shows field-level confidence and flags any low-confidence extractions. For your first round, expect 15-25% of fields to need adjustment. Annotate the correction examples, retrain, test again. Most templates reach a stable accuracy level after 2-3 rounds.
Custom templates are available on Professional and Scale plans. The Starter plan includes access to the 3 built-in templates.