it's too bad DPR dataset are TREC, which mean it's still a supervised learning based :( hoping something like beam search like on google's meena (the awesome open domain chatbot)
Follow
and it uses FAISS, the infamous nmap indexer.
approximate vector search (neighbor search ones, like annoy/hnsw/milvus), especially on <10k docs is actually shitty (it'll return random shit, way out of the query context). despite what word / sentence embedding used.
