Version video embeddings together with their temporal meaning

Store stable source identity and version, stream selection, interval endpoints with their rational timebase, and the sampling/preprocessing versions. Include encoder weights, output dimension, dtype, pooling rule, and whether normalization occurred before storage or search. A single vector can summarize many sampled frames, so its interval is not necessarily one frame timestamp. Re-embedding after a decoder or sampling change can alter the representation even if the encoder's parameter file remains unchanged.

An index generation should reference the exact embedding manifest and metric configuration it contains. Do not mix vectors from incompatible encoder versions simply because both have 512 coordinates. A migration can build a new generation, run compatibility and retrieval checks, then switch a declared query/encoder/index tuple together under the deployment's actual atomicity contract. Retain enough provenance to roll back and to trace a hit to its source interval; an index identifier alone cannot reconstruct missing preprocessing history.

FAISS's official metric documentation distinguishes squared Euclidean distance from inner product. Inner-product rankings are not automatically cosine rankings when vector magnitudes differ. To implement cosine through inner product, normalize both database and query vectors under a declared nonzero-vector policy. Reject, specially represent, or otherwise handle zero vectors rather than dividing by zero. Store the normalization convention with the index and verify it on the query path as well as the offline embedding job.

For unit vectors, squared L2 and inner product are related by a simple identity, so their rankings correspond with opposite ordering. That relationship does not apply unchanged to arbitrary unnormalized vectors. Quantization and approximate indexes introduce additional retrieval behavior that should be measured against exact search over the same eligible records. Recall at k of an approximate search answers whether exact neighbors were recovered, not whether neighbors have correct action labels or temporal localization for a downstream task.

Estimate vector payload and required provenance before choosing an index strategy. An index may store graph links, coarse centroids, compressed codes, or other overhead, so measure its actual serialized and resident sizes separately. Test that every returned interval decodes to the intended source moment under the recorded selection policy. A semantically related clip with an incorrect timestamp can fail a moment-retrieval task even when its vector score is high. Retention and deletion procedures must account for both payload and index generations.

The vector lab illustrates compatible two-dimensional metric arithmetic. Reset, set both angle=60°,magnitude=1, predict components (0.5,approximately 0.8660), then Play or Step. With unit query (1,0), the dot product is 0.5 and squared L2 is one. Reset and set both angle=0°,magnitude=1 for dot product one and squared L2 zero. It contains no video encoder, searchable collection, quantization, eligibility filtering, or nearest-neighbor index; it checks the geometry rather than retrieval quality.