Choosing an extraction approach

As you scale up and encounter document complexity, you can optimize extraction performance by choosing between LLMs or layout-based extraction methods.

See the following diagram for an overview of when to use Sensible Instruct (LLMs) or SenseML (layout-based) for extractions:

Click to enlarge

Sensible recommends using large language model (LLM) prompts for free-form, highly variable documents, and layout-based, or "rule-based" queries for structured, less-variable documents. You can combine both strategies since they're fully compatible with each other.

See the following table to learn more about extraction strategies:

extraction method	notes	get started
LLMs	You describe the document data that you want to extract, using Sensible's visual authoring tool, Sensible Instruct. Sensible uses GPT-4 and other large language models (LLMs) to extract data from your documents.	Getting started
layout-based	To extract from complex document layouts, use SenseML, a superset of Sensible Instruct. SenseML is a JSON-formatted query language that combines layout-based queries with LLM-based queries. When either strategy can work, Sensible recommends layout-based queries for the sake of fast performance and deterministic output.	Getting started with layout-based extractions
out-of-the-box	Sensible provides out-of-the-box extraction configurations for common business and tax forms. Use Sensible's pre-built, open-source configuration library to extract key information from tax forms such as 1099s, major carrier insurance declaration pages, and other documents. Then tweak the pre-built configurations for your custom data needs.	Getting started with out-of-the-box extractions

See the following table for an overview of the pros and cons of LLMs versus layout-based extraction:

	LLM (Sensible Instruct)	layout-based (SenseML)
Technical expertise required	For nontechnical users. Describe what you want to extract in a prompt to an LLM. For example, "the policy period" or "total amount invoiced".	Offers highly configurable JSON-based extraction configuration for technical users. For example, write instructions in JSON to grab the second cell in a column headed by "premium."
Workflow automation	Suited to workflows that include human review or that are fault-tolerant.	Suited to automated workflows that require predictable results and validation.
Document variability	Suited to documents that are unstructured or that have a large number of layout variations or revisions.	Suited to structured documents with a finite number of variations, where you know the layout of the document in advance.
Deterministic	No	Yes. Find the information in the document using anchoring text and layout data.
Handles repeating layouts	Use List method.	Use sections for highly complex repeating substructures, for example, loss runs.
Handles non-text images (photos, illustrations, charts, etc)	To extract data about images (`"is the building in this picture multistory?"`, use Query Group method with the Multimodal Engine parameter configured	No
Performance	Data extraction takes a few seconds for each Instruct method.	Offers faster performance in general. For more information, see Optimizing extraction performance.