Overview
VidSlice indexes third-party video material — photos, clips, B-roll — into a LanceDB vector store, signs every entry with Ed25519 for tamper detection, and serves results over HTTP. Built to plug into the OpenCut short-video pipeline (which is why the HTTP API matches OpenCut's material contract).
v0.5 read-compat / v0.6 write opt-in via --compat v0.5 flag. NDJSON progress events for streaming pipelines.
Pain → Solution → Value
Every team rolls its own frame extractor and embedding pipeline. Vendors cannot be trusted with raw production assets, yet indexing has to happen somewhere. VidSlice gives the outsourced team a standalone CLI: hand it materials, get back a signed standard index that OpenCut consumes directly.
What changes for the team
A wider view of the same loop. Each card zooms into the corresponding column above.
Outsourced indexing without leaking secrets
- Every team rolls its own frame extractor and embedding pipeline — formats drift across groups.
- Vendors cannot be trusted with raw production assets, yet indexing has to happen somewhere.
- A clip indexed by group A rarely loads cleanly in group B's pipeline.
One CLI, one signed LanceDB output
- Hand it raw materials — VidSlice walks the directory, extracts frames, computes embeddings.
- Writes a signed LanceDB index that OpenCut consumes directly.
- Outsourced teams never touch production credentials, repos, or downstream services.
Third parties index, secrets stay home
- Outsourced teams run VidSlice on materials they're given; signed index flows back to OpenCut.
- Index format consistent across vendors — no per-partner normalization layer.
- Production credentials, source repositories, and downstream services stay inside the perimeter.
Product architecture
Four CLI stages turn a directory of raw material into a signed vector index ready for OpenCut to ingest.
-
01
Discover
The CLI scans the materials directory, enumerating supported video and image files and recording their paths and hashes for downstream stages.
-
02
Extract
Frames are pulled from each video and structured metadata (codec, duration, dimensions, source hash) is captured alongside them.
-
03
Embed
A model computes per-frame vectors representing the searchable content; vectors and their metadata are written into the LanceDB store.
-
04
Sign
The output index is signed with Ed25519, producing the standardized artifact OpenCut consumes directly — tamper-evident, version-pinned, and ready for the HTTP API.
Runtime sequence
A single VidSlice CLI run walks through five participants.
Capabilities
By the numbers
- 408 tests
- LanceDB storage
- 3-frame extraction
- Ed25519 attestation
- HTTP + NDJSON streaming
- v0.5 / v0.6 dual-versioned