Models

SEED calls foundation models through Amazon Bedrock. Each pipeline stage can use a different model, selected by a CLI flag, so you can trade off speed, cost, and quality per stage.

Available Models

Key	Model	Best For
gpt-oss	GPT-OSS 120B	Doc generation, augmentation (fast tool calling)
sonnet	Claude Sonnet 4.6	Critics (respects scope rules, good at evaluation)
nova2-lite	Amazon Nova 2 Lite	Data generation (fast, cheap, works with the calculator)
nemotron-super	Nemotron Super 120B	Doc generation (slow but high first-pass quality)
haiku	Claude Haiku 4.5	Fast critic (less reliable at scope enforcement than sonnet)

Per-Stage Model Selection

Each stage has its own flag and default:

Flag	Default	Stage
--data-model	nova2-lite	Data generation
--doc-model	gpt-oss	PDF generation
--critic-model	sonnet	All critics
--aug-model	gpt-oss	Augmentation decisions
--context-model	gpt-oss	Shared context resolution (packets only)

Example, overriding models per stage:

python -m seed_data \
  --schema-dir src/seed_data/schemas/fcc-invoice \
  --scenario "Office supplies" \
  --data-model nova2-lite \
  --doc-model gpt-oss \
  --critic-model sonnet

Critic Model Constraint

The critic stages evaluate rendered PDFs with a vision model, so --critic-model must support document/PDF input. Currently that means an Anthropic model: haiku, sonnet, or opus. All other model flags can use any available model.

Model-Agnostic Design

The pipeline is model-agnostic: the seven agents (scenario generator, data generator, data critic, doc generator, doc critic, augmentor, aug critic) each read a model key and construct a Bedrock-backed model, so you can mix and match. The default assignments reflect what works well for each role:

Agent	Role	Default Model
Scenario Generator	Creates diverse scenarios from brief + guidance	gpt-oss
Data Generator	Produces JSON data from schema. Has a calculator tool for math verification.	nova2-lite
Data Critic	Validates data against schema and domain rules. Has a calculator tool. Structured output.	sonnet
Doc Generator	Writes HTML/CSS and renders to PDF via WeasyPrint. Has an editor for targeted fixes.	gpt-oss
Doc Critic	Vision model evaluates PDF quality: layout, typography, truncation, math. Free-text output.	sonnet
Augmentor	Picks an augraphy augmentation config and applies document aging effects.	gpt-oss
Aug Critic	Evaluates the augmented doc for legibility and realism. Structured output.	sonnet

Cost Notes

Generation runs make repeated Bedrock calls and use critic-driven retry loops. Large batch counts or a high --max-attempts increase token spend. Keep --max-attempts and --timeout conservative and monitor Bedrock usage.