The API server’s late‑interaction scoring for ColBERT models (e.g., answerdotai/answerai-colbert-small-v1 )…
The API server’s late‑interaction scoring for ColBERT models (e.g., answerdotai/answerai-colbert-small-v1 )…: a task in LegoFlow-SWE (Harbor dataset). When a scoring request arrives, the server must cache the query embeddings, associating them with a key, and track how many documents will be…
The task
When a scoring request arrives, the server must cache the query embeddings, associating them with a key, and track how many documents will be scored against that query. Each document’s embedding is then scored against the cached query embeddings using the MaxSim operation: for each query token, compute the maximum…
Part of Lego-X/LegoFlow-SWE.