Utility_Bills

Utility Bill Extraction — LandingAI ADE

This use case extracts structured fields from utility bills (electric / gas) using LandingAI's Agentic Document Extraction (ADE). Utility bills are a common proof-of-address document in KYC and onboarding workflows. Fields include provider info, account details, billing summary, and electric/gas charges.

It ships two independent implementations of the same task, driven by the same shared schema (utility_bill.json), so you can compare approaches side by side:

	v1/	v2/
Model family	DPT-2	DPT-3
Interface	landingai-ade Python SDK	v2 REST APIs (direct calls)
APIs used	Parse + Extract (SDK)	Parse Jobs + Extract Jobs (async)
Form factor	Jupyter notebook	Standalone Python scripts
Service tier	n/a	standard

Both read the same inputs and the same utility_bill.json schema, so the results are directly comparable.

Folder layout

Utility_Bills/
├── input_folder/                 # 9 utility bills (6 PDF + 3 JPG), shared by both versions
├── utility_bill.json             # shared extraction schema (used by v1 AND v2)
├── images/                       # supporting images
├── README.md                     # this file
├── v1/                           # DPT-2, via the Python SDK
│   ├── parse_extract_utility_bills.ipynb   # walkthrough notebook
│   └── results_folder/           # generated outputs
└── v2/                           # DPT-3, via the v2 REST APIs
    ├── README.md                 # detailed v2 walkthrough
    ├── process_utility_bills.py  # end-to-end pipeline
    ├── ade_v2_client.py          # thin REST client for the v2 Jobs APIs
    └── results_folder/           # generated outputs (parse / extract / csv_summaries)

The schema is intentionally not duplicated — both versions load the single utility_bill.json at the folder root.

The two versions

v1/ — DPT-2 via the Python SDK

The original sample. It uses the landingai-ade SDK to parse each bill and extract fields against the JSON schema, driven from a Jupyter notebook. Open v1/parse_extract_utility_bills.ipynb to run it.

v2/ — DPT-3 via the v2 REST APIs

A newer sample that calls the v2 REST endpoints directly — submitting each bill to Parse Jobs, polling for results, then running Extract Jobs against the same schema — using the DPT-3 model family. It runs as a plain Python script. See v2/README.md for the full walkthrough.

Cost comparison (high level)

Both versions were run over the same 9 utility bills (17 pages total) with the same schema. Costs below are credits as reported by each API.

Stage	v1 (DPT-2)	v2 (DPT-3)	Difference
Parse	51.0	24.2	−53%
Extract	23.5	20.3	−14%
Total	74.5	44.5	−40%
Credits per page	4.4	2.6	−40%

For this set of bills, v2 (DPT-3) costs about 40% fewer credits than v1 (DPT-2). Most of the savings come from the parse step: DPT-2 charged a flat 3.0 credits per page, while DPT-3 Parse Jobs (standard tier) is complexity-aware — and because this corpus includes multi-page bills (up to 4 pages), the per-page parse savings add up.

Notes. This is a credits-to-credits comparison on one small sample, not a dollar quote. v2 runs on the standard service tier (the lowest-cost tier). The two versions use different model families and pricing, so treat the numbers as directional guidance for this dataset rather than a universal benchmark.

Getting started

Each version authenticates with a VISION_AGENT_API_KEY (environment variable or a local .env file). Pick a folder and follow its instructions:

• v1: open v1/parse_extract_utility_bills.ipynb.
• v2: see v2/README.md, then run python v2/process_utility_bills.py.