I Evaluated EmbeddingGemma 2 Like an Architect Would. Here's the Decision Record.

Can one open model make a mixed-media archive searchable without a transcription pipeline, and under what conditions would I recommend it to an organisation? This is an architecture decision record, narrated. Not a benchmark review, not a product demo. Five options including no AI at all, seven criteria, every number tagged with its source, what held up, what didn't, four sub-decisions with the option rejected, and a conditional recommendation gated on an evaluation I have not yet run. The reference implementation is Footage Finder: a local app that indexes video, audio, photos and PDF pages with EmbeddingGemma 2 into one SQLite index and searches it by typed text, spoken question or image. Loopback only. Nothing leaves the machine after the weights download. Status, stated plainly: independent build on a personal test library of public clips. Not a client system. Not deployed. The labelled evaluation is proposed, not run. Code and the written decision record: 🔗 https://github.com/praveenvijayan/foo... 🔗 https://github.com/praveenvijayan/foo... Model: google/embeddinggemma-2 (open weights, Hugging Face, licence-gated) Chapters 0:00 The question 0:52 Baseline and business case 1:58 Options: no AI, caption chain, managed API, local open model, hybrid 3:03 The candidate: EmbeddingGemma 2 in a minute 3:56 Reference implementation: Footage Finder 5:12 Evaluation: what held, what broke 7:24 Decision records: four sub-decisions, each with the rejected option 8:44 Enterprise consequences: boundary, cost, operations, hybrid search 10:09 Recommendation: conditional pilot, and when not to use it 10:48 Close Evidence codes used on screen M measured on my MacBook · B observed in my build · P public Colab test · C model card · J my judgement · V unverified · I illustrative · N not run Numbers quoted from public tests (Flickr-1k photo retrieval, 50 sound categories) are not my data. My own accuracy numbers do not exist yet; the gate for producing them is in the video and the ADR. If you would weight a criterion differently, or make a different call, say which one in the comments. That is more useful to me than agreement.