Rerank In Table
To maximize performance, VectorChord stores both the quantized and original vectors in the index. Although the original vectors can be fetched from the table, doing so is much slower than fetching them from our highly optimized data structure.
As a result, VectorChord's superior performance is partly due to its larger disk usage compared to disk-saving solutions like pgvectorscale. However, this is usually acceptable because disk space is affordable, and performance is more important.
This behavior is called rerank in index
and is the default option for VectorChord.
However, if you are short on disk space and don't care much about performance, you can enable rerank_in_table
option in CREATE INDEX
to reduce disk usage.
CREATE INDEX ON items USING vchordrq (embedding vector_l2_ops) WITH (options = $$
residual_quantization = true
rerank_in_table = true
[build.internal]
lists = []
$$);
Then, all queries of this index will retrieve the original vectors from the table, which will significantly reduce disk usage.
Performance Trade-offs
- Query performance: With
rerank_in_table
option, VectorChordvchordrq
index will have a 30-50% increase in latency and QPS loss at query - Index size: With
rerank_in_table
option, VectorChordvchordrq
index will save about 50% of disk usage
WARNING
Prefetch does not work when rerank_in_table
is enabled. Disk read operations, such as large indexes and cold starts, will be heavily affected.