Worked example: legal as document-vision composition

This is an illustrative example with placeholder candidates and scores, not a captured Hub search or runtime result. Actual output depends on current metadata.

This example shows the domain-agnostic engine applied to a legal use case by composing the document-vision foundation (OCR / layout). No legal-specific code exists in the plugin; only the interview answers differ.

Interview transcript

What are you building? Describe the domain in a sentence.
> On-device legal contract review with scanned PDFs

# document-vision is detected from "scanned PDFs" / OCR language — no modality ask

Which model roles do you need? (choose one or more) [Vision / OCR, Embeddings, Coding, Reasoning, General chat, Tool use / function calling]
> Vision / OCR, General chat

Any domain keywords to prioritize? (comma-separated, optional)
> ocr, contracts, redaction

Restrict to specific licenses? (comma-separated, optional)
> apache-2.0, mit

Memory budget in GB? (optional, e.g. 32)
> 32

Any other constraints or notes? (optional)
> must run fully offline

Resulting intent

{
  "domain": "On-device legal contract review with scanned PDFs",
  "roles": ["vision", "general"],
  "keywords": ["ocr", "contracts", "redaction"],
  "license_allow": ["apache-2.0", "mit"],
  "memory_gb": 32.0,
  "notes": "must run fully offline",
  "modalities": ["document-vision"],
  "facets": ["ocr"]
}

Resulting pack (excerpt)

# MLX Research Pack: On-device legal contract review with scanned PDFs

- Roles: vision, general
- Keywords: ocr, contracts, redaction
- License filter: apache-2.0, mit
- Memory budget (GB): 32.0
- Modalities: document-vision
- Facets: ocr

## Modality foundations

- `document-vision` — Document extraction / vision
  - Facets: `ocr` (OCR / text extraction)

## Runtime preference

- Preferred: `mlx-vlm`
- Alternates: `lmstudio`
- Rationale: The selected vision workflow uses mlx-vlm; review runtime support for the chosen model.

## Candidates

### 1. `mlx-community/<an-ocr-vision-model>` — score 82.0/100

- Role: vision
- Suggested wiring: mlx-vlm server -> ...
- Why it ranked:
  - role_match: matched roles: vision
  - keyword_match: matched keywords: ocr
  - license_ok: license: apache-2.0
  - memory_fit: est_ram_gb ... vs budget 32.0

## Adapters / LoRAs

### 1. `org/<legal-ocr-lora>` — score 71.0/100

- Kind: adapter
- Why it ranked:
  - keyword_match: matched keywords: ocr, contracts

## Datasets

No datasets matched this intent on the Hub. See the dataset blueprint below.

## Dataset blueprint

No Hub datasets ranked for this intent. Use this guidance-only blueprint
to create a local dataset — nothing is downloaded or trained by research.

### Goal

Create a small, domain-specific dataset for **On-device legal contract review...**

Legal is a composition of the document-vision foundation plus domain keywords — not a separate engine. The same foundations power audio and video packs (see sibling examples). When Hub datasets are empty, the pack still includes an explicit dataset blueprint so another agent can draft labeled data locally without any download or training from research.