# B-Roll Archives 0.5.0 — build 10

- Optional **Enhanced Search** uses Google EmbeddingGemma 2 locally to retrieve
  visually relevant clips and moments. It is experimental and off by default.
  Standard CLIP and metadata search remain available, including when Enhanced
  Search cannot run or fails. Results appear in separate groups.
- Download the approximately 865 MB model once, run a bounded local comparison,
  and review hardware warnings before enabling indexing. Downloading or testing
  does not start full-library indexing. Enhanced Search requires macOS 13.3 or
  later on Apple Silicon; Standard Search retains its macOS 13 deployment target.
- Enhanced indexing reports coverage, measured throughput, estimated active-work
  time remaining and separate process memory readings. Pause/resume is persistent;
  indexing adapts to memory, thermal and power conditions. Disabling retains its
  index; removing enhanced data preserves Standard search and source footage.
- Video metadata shows saved shot-list entry usage and the number of distinct
  saved lists. Limit prior usage when selecting candidates; a limit of zero
  selects unused videos. Counts reflect currently saved entries, including
  repeats, and decrease when entries are removed. MCP exposes the same controls.

Existing footage, catalogues, saved work and licence state are retained through
normal installation. Enhanced model files are downloaded separately; inference
runs locally. This release does not add audio, PDF, code or sequence embeddings.

Release evidence covers an M4 Mac with 24 GB RAM on macOS 26.3, automated
regressions, native model parity and the repository sample-library. In the
bounded cached-frame comparison, Enhanced encoding averaged about 202 ms per
frame versus 9 ms for CLIP. Those figures exclude full-library source decoding
and database writes; actual indexing depends on footage and computer resources.

Whole-app resource/UI coverage, 8 GB M1 and older macOS, battery/physical Low
Power Mode, real NAS, full cold long-video and additional codec coverage remain
unverified. Large archives can take substantial time to index, and very large
catalogues may have slow retrieval. An 8 TB archive is not certified for
interactive performance. Hardware warnings and fallback do not guarantee the
same encoding experience on an untested computer.
