Qdrant is not a Graph Database, but graphs are still the backbone of vector search technology since HNSW is a graph-based ANN algorithm. I stumbled upon an interesting white paper that evaluates twelve state-of-the-art graph-based methods on large datasets, up to 1B vectors.⤵️

For 𝐡𝐨𝐫𝐢𝐳𝐨𝐧𝐭𝐚𝐥 𝐬𝐜𝐚𝐥𝐚𝐛𝐢𝐥𝐢𝐭𝐲, sharding and consensus logic need to be built around search. qdrant.to/d9d For flexible 𝐦𝐞𝐦𝐨𝐫𝐲 𝐫𝐞𝐬𝐨𝐮𝐫𝐜𝐞 𝐨𝐩𝐭𝐢𝐦𝐢𝐳𝐚𝐭𝐢𝐨𝐧, quantization techniques need to be applied. qdrant.to/q10n For a more 𝐩𝐫𝐞𝐜𝐢𝐬𝐞 𝐫𝐞𝐭𝐫𝐢𝐞𝐯𝐚𝐥, Hybrid Search with sparse vectors can be used. qdrant.to/hybrid For 𝐦𝐮𝐥𝐭𝐢𝐭𝐞𝐧𝐚𝐧𝐜𝐲 𝐬𝐮𝐩𝐩𝐨𝐫𝐭, payload-based partitioning is required. qdrant.to/m10y For even more 𝐞𝐟𝐟𝐢𝐜𝐢𝐞𝐧𝐭 𝐢𝐧𝐝𝐞𝐱𝐢𝐧𝐠, GPUs can be used. qdrant.to/gpu For combining similarity search with domain-specific logic, a score boosting reranker can be applied. qdrant.to/boosting. And many, many more production-critical aspects.