[
  {"id":"memory","title":"Conversation memory","text":"During text generation, the KV cache stores attention keys and values. More context tokens or concurrent conversations require additional memory for these states."},
  {"id":"experts","title":"Conditional computation in MoE models","text":"A router selects a few experts for each token. Unselected experts still need weight storage. When experts live on different GPUs, activations must reach the devices hosting them."},
  {"id":"local","title":"Offline document retrieval","text":"An application can calculate document and query embeddings on the user’s machine. A locally stored index retrieves passages without sending their contents to a remote API."},
  {"id":"precision","title":"Numerical formats for EmbeddingGemma 2","text":"EmbeddingGemma 2 activations are incompatible with FP16. The documentation recommends BF16 on supported hardware and FP32 otherwise. Weight quantization requires a compatible recipe and runtime."},
  {"id":"video","title":"Finding a moment in a video","text":"To locate an event, an application segments the video and retains timestamps. It chooses a frame sampling rate appropriate to the events being retrieved."},
  {"id":"ranking","title":"Improving result ranking","text":"An initial vector search retrieves candidate passages. A reranker can then examine the question together with each passage to refine their ranking before display or generation."}
]