Add NOOA-backed extract.dimensions wrangle

[ebhills]

Description

Add extract.dimensions, an AI-powered WranglesPY wrangle for extracting source-grounded dimensional measurements from messy product descriptions and records.

The initial scope covers:

• length
• width
• height
• diameter, including OD, ID, DIA, and Ø notation
• depth
• explicitly stated volume
• miscellaneous dimensional measurements such as thickness, radius, bore, area, clearance, and gauge

Each result retains the numeric value or range, normalized unit, optional qualifier/label, and the shortest exact source text supporting the extraction.

How NOOA works here

This wrangle is also a focused evaluation of NVIDIA NOOA as a lightweight agent harness inside WranglesPY. The prototype uses the NOOA 0.0.10 release, which is now available on PyPI.

NOOA lets an agent operation be expressed as a typed Python method:

• the method docstring supplies the extraction instructions;
• annotated method arguments provide the input data;
• the Pydantic return type becomes the structured-output contract;
• PredictStrategy performs a focused structured prediction without the CodeAct generated-code/tool loop;
• NOOA decodes and validates the response against the return type and can ask the model to correct an invalid response;
• its LLM client uses LiteLLM model/provider names, so the wrangle is not tied to one provider.

Each product row is one Predict invocation, normally one provider request. Output-validation retries can add requests. Concurrent rows use separate agent instances because a NOOA agent serializes its own generation calls.

The default in-memory event store is sufficient for this use case. We do not need NOOA long-term memory, SQLite persistence, MCP, CLI, benchmarks, sandboxes, tracing/viewer, or CodeAct.

Desired behavior

Python:

results = wrangles.extract.dimensions(
    [
        "SS sink bowl 18 x 14 x 8 in deep; drain opening DIA 3.5 in",
        "Bottle, 750 mL capacity, 9.7 in H x 3.2 in OD",
        "Shelf 12-14 in wide, 3/4 in thick",
    ],
    threads=3,
)

Recipe:

wrangles:
  - extract.dimensions:
      input:
        - Description
        - Size
        - Packaging
      output: Dimensions
      model: gpt-5.4-mini
      api_key: ${OPENAI_API_KEY}
      threads: 4

Example result:

{
  "measurements": [
    {
      "kind": "diameter",
      "label": "outside diameter",
      "value": 3.2,
      "minimum": null,
      "maximum": null,
      "unit": "in",
      "qualifier": "outside",
      "source": "3.2 in OD"
    }
  ]
}

For a range, value is null and both minimum and maximum are populated. A miscellaneous measurement requires a descriptive label. An input with no supported dimensional fact returns an empty measurements list.

Prototype design

• Add NOOA as the exact optional extra nooa==0.0.10; the base WranglesPY install remains free of this experimental dependency.
• Add the same exact pin to requirements-full.txt for Python 3.12 and 3.13. Its environment marker leaves WranglesPY's Python 3.11 base path unaffected.
• Import NOOA lazily so a normal import wrangles does not load NOOA or LiteLLM.
• Use strict Pydantic models and reject uncontracted output fields.
• Accept scalar text/records and lists in Python; accept one or multiple input columns in recipes.
• Preserve input row order while allowing bounded row-level concurrency.
• Pass provider credentials directly to the NOOA client; never add credentials to the prompt or output.
• Treat product content as untrusted data and expose no tools to the Predict agent.

Intentional boundaries

• Do not convert between units.
• Do not calculate volume from other dimensions; return volume only when explicitly stated.
• Do not interpret counts, model numbers, electrical ratings, weights, or ordinary pack quantities as dimensions.
• Do not introduce saved models, caching, long-term memory, or additional infrastructure in this prototype.
• A provider exception for any row currently fails the complete batch; the prototype does not yet add a Wrangles-level deadline or per-row error value.
• Live quality, latency, cost, and failure behavior still need evaluation on representative sanitized product data before production use.

Dependency and deployment boundary

WranglesPY's requirements-full.txt will install nooa==0.0.10 on its Python 3.12 and 3.13 validation paths. Python 3.11 skips it through the version marker.

The deployed execute-recipe-dev image is owned by Lambda-Recipes, so merging WranglesPY alone will not make NOOA available in DEV. A separate Lambda-Recipes dependency branch will add the exact nooa==0.0.10 pin. DEV testing should begin only after that companion change is deployed with the intended WranglesPY release candidate.

Current caveats

• NOOA 0.0.10 supports Python 3.12 and 3.13, not WranglesPY's Python 3.11 floor; calling this optional wrangle on 3.11 produces a clear installation/runtime error.
• NOOA 0.0.10 has native-Windows import assumptions around fcntl and SIGUSR2. The prototype includes narrow guards only for Predict with in-memory storage. SQLite persistence and Unix signal debugging remain unsupported on native Windows; Linux/DEV behavior must be verified separately.
• Every non-empty row produces a model request and validation retries can produce additional requests, so DEV testing must examine realistic concurrency, rate limits, latency, and cost.

Acceptance criteria

• The Python API returns the documented scalar/list structured contract.
• The recipe API supports one or multiple input columns and writes one structured object per row.
• Tests cover scalar values, ranges, compact L×W×H groups, written fractions, shared trailing units, OD/ID/DIA, and labeled miscellaneous dimensions.
• Representative evaluation cases exclude counts, identifiers, weights, electrical values, and packaging noise.
• Invalid output shapes fail validation or are corrected through NOOA's validation retry.
• Empty input, row order, bounded concurrency, and missing optional dependency behavior are tested.
• Base WranglesPY import and Python 3.11 installation remain unaffected.
• WranglesPY CI installs the exact pin through requirements-full.txt on Python 3.12/3.13 paths.
• The companion Lambda-Recipes branch installs nooa==0.0.10 in the DEV runtime.
• A Linux DEV smoke test confirms dependency availability, existing credential routing, one real structured extraction, and a small multi-row run.
• A sanitized golden corpus is reviewed before the wrangle is considered production-ready.

Offline validation completed

• Focused core/schema/recipe tests: 20 passed; the 2 optional NOOA runtime tests skip cleanly when NOOA is absent.
• Actual NOOA 0.0.10 contract tests using its offline fake LLM client: 2 passed, including strict structured validation and validation retry.
• The package builds successfully with the exact wrangles[nooa] pin.
• Ordinary import wrangles does not import NOOA.
• No test makes a live provider request.

Remaining DEV verification

After the WranglesPY and Lambda-Recipes dependency changes are available together in DEV:

1. Confirm nooa==0.0.10 is installed in the deployed Linux runtime.
2. Run one real recipe extraction through the existing model/key-routing path.
3. Run a small multi-row sample to observe output quality, ordering, retry behavior, latency, rate limits, and cost.
4. Record the deployed WranglesPY version/commit and Lambda image revision so the smoke evidence is reproducible.