feat: add opt-in TwelveLabs video modality (Pegasus + Marengo)

[mohit-twelvelabs]

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

Description

This PR adds an opt-in video modality to RAG-Anything, backed by the TwelveLabs platform. It lets RAG-Anything ingest and retrieve video the same way it already handles images, tables and equations.

• Pegasus (analyze) generates a detailed transcript/description of a video. That text flows through the existing knowledge-graph + chunk pipeline via BaseModalProcessor._create_entity_and_chunk — no new retrieval path, no special-casing downstream.
• Marengo (embed) produces a 512-dim multimodal embedding, returned on the item's entity_info (tl_video_embedding) for semantic video retrieval. A companion embed_text() lets you score a text query in the same 512-dim space.

Why it helps this project

RAG-Anything is explicitly multimodal, but video isn't yet a first-class modality. Pegasus + Marengo give a turnkey way to bring video understanding and video↔text semantic search into the existing graph/RAG pipeline, with no extra infra on the user's side.

Changes Made

• raganything/twelvelabs.py: new TwelveLabsModalProcessor (mirrors GenericModalProcessor). Accepts video_url, video_path (uploaded as a TwelveLabs asset), or video_id.
• config.py: enable_video_processing flag (env ENABLE_VIDEO_PROCESSING), default False.
• raganything.py: registers the video processor only when enabled and twelvelabs is installed and a key is present; otherwise it logs a warning and skips — initialization never breaks.
• utils.py: dispatch routes "video" to the processor (falls back to generic when unregistered, never KeyError).
• __init__.py: optional TwelveLabsModalProcessor export (same try/except pattern as other optional features).
• pyproject.toml / setup.py: new [video] extra (twelvelabs>=1.2.8), also folded into [all].
• env.example, requirements.txt, README.md: docs + config entries.

Opt-in / non-breaking

Disabled by default. With ENABLE_VIDEO_PROCESSING unset, there are zero behavioural changes and no new hard dependency (twelvelabs only installs via the [video]/[all] extra).

How it was tested

tests/test_twelvelabs_integration.py:

• No-network unit tests: config default, dispatch routing/fallback, and Pegasus/Marengo request wiring against a mocked TwelveLabs client.
• A live smoke test (gated on TWELVELABS_API_KEY, skipped in CI without it) asserting a real Marengo text embedding is 512-dim — verified locally (9 passed).

Ran the repo's ruff format --check and ruff check --ignore=E402 on all changed files — clean.

Note: full Pegasus/Marengo video runs are server-side and can be slow (Marengo video embedding is an async, polled task). The video paths are wiring-verified against the SDK contract; the synchronous Marengo text-embedding path is verified end-to-end against the live API.

Checklist

• Changes tested locally
• Documentation updated
• Unit tests added

Additional Notes

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