{"samples":[{"categories":["factual"],"id":"storage-durability","metadata":{"source":"database-guide"},"tags":["storage","durability"],"text":"PF Vector persists vectors using transactional storage and write-ahead logging so committed records survive a restart."},{"categories":["advice","factual"],"id":"embedding-compatibility","metadata":{"source":"database-guide"},"tags":["embeddings","configuration"],"text":"Use the same embedding model and vector dimensions for stored records and search queries. Re-embed data into a new collection when switching models."},{"categories":["factual"],"id":"partition-isolation","metadata":{"source":"database-guide"},"tags":["isolation","retrieval"],"text":"A collection can contain multiple partitions. Similarity search is restricted to the selected database, collection and partition before vectors are scored."},{"categories":["advice"],"id":"search-benchmarking","metadata":{"source":"engineering-notes"},"tags":["performance","benchmarking"],"text":"Measure retrieval latency with representative vector counts, dimensions and concurrent clients before choosing an approximate nearest-neighbor index."}]}