Add Marengo embedding predictor for zero-shot moment retrieval

[mohit-twelvelabs]

Hi! I'm Mohit, I work at TwelveLabs (@mohit-twelvelabs).

This PR adds an optional, zero-shot moment retrieval predictor backed by TwelveLabs Marengo embeddings.

What it adds

• lighthouse/marengo_predictor.py — a MarengoPredictor that segments and embeds a video server-side, embeds the text query into the same 512-dim Marengo space, and ranks clips by cosine similarity. Its output matches the existing predictors: {"pred_relevant_windows": [[start, end, score], ...]}.
• Re-exported from lighthouse.models alongside the existing predictors.
• api_example/marengo_demo.py mirroring the existing demo.py / amr_demo.py.
• A README section, a marengo optional extra in setup.py, a mypy.ini entry, and tests.

Why it helps Lighthouse

• A training-free baseline that needs no local checkpoint, feature files, or GPU.
• Not bound by the 150s benchmark limit, since embedding happens server-side — handy for longer videos.

Opt-in and non-breaking

• The twelvelabs SDK is an optional extra (pip install 'lighthouse[marengo]') and is only imported when MarengoPredictor is instantiated, so existing users pay no import or dependency cost. No defaults or existing predictors change.

How it was tested

• No-network unit tests in tests/test_marengo.py cover the cosine-similarity ranking and the empty-input path.
• A live test (skipped unless TWELVELABS_API_KEY is set) asserts a Marengo text embedding is 512-dim — verified passing locally against the API.
• ruff and mypy pass on the new files.

You can grab a free API key at https://twelvelabs.io — there's a generous free tier.

Per the contributing note, happy to open a tracking issue first if you'd prefer to discuss before reviewing — opening the PR so the proposed change is concrete.

[awkrail]

Thank you for your contribution. We will look into the code today.

[awkrail]

Does this work well? As shown in README, lighthouse is installed from github repo directly. So, correctly, pip install 'lighthouse[marengo] @ git+https://github.com/line/lighthouse.git' works?

[awkrail]

In addition, we would like to think about the naming. Now, Twelvelabs have Marengo as a video embedding model, but in the future, Twelvelabs may release other video embedding models, right? If so, I would like to use lighthouse[twelvelabs] rather than Marengo itself.

[awkrail]

@mohit-twelvelabs I added some comments to your code. I will check the logit part tomorrow by running the code...

[mohit-twelvelabs]

Thanks for the careful review, @awkrail! Both great points — addressed in f3ee766:

1. Install command — you're right, since lighthouse installs from GitHub (not PyPI), pip install 'lighthouse[marengo]' wouldn't resolve. Fixed the README to the git form you suggested:

   pip install 'lighthouse[twelvelabs] @ git+https://github.com/line/lighthouse.git'
   

2. Naming — fully agree on future-proofing. I renamed the optional extra from [marengo] to [twelvelabs] (it installs the twelvelabs SDK, which will cover future embedding models too), so the install is now lighthouse[twelvelabs].

I left the predictor class as MarengoPredictor since it's specific to the Marengo embedding model, but happy to rename it (e.g. TwelveLabsPredictor with a model_name arg) if you'd prefer the vendor-scoped naming there as well — just say the word.

No rush on the logit part — thanks for offering to run it, and let me know if anything comes up when you do.