refactor: improve posting cursor

[silver-ymz]

bench

machine: 8 vCPU, 32 GB, AMD Ryzen 7 7700, hyper-v Virtualization

VectorChord-bm25 (unicode tokenizer)

dataset	topk	qps	NDCG@10
trec-covid	10	175.19	0.64728
trec-covid	1000	51.87	0.64665
webis-touche2020	10	121.26	0.33956
webis-touche2020	1000	45.83	0.33956
fiqa	10	303.44	0.25465
fiqa	1000	61.45	0.25465

elastic-search

dataset	topk	qps	NDCG@10
trec-covid	10	138.73	0.68803
trec-covid	1000	48.77	0.68803
webis-touche2020	10	182.16	0.34707
webis-touche2020	1000	38.07	0.34707
fiqa	10	326.91	0.25364
fiqa	1000	60.07	0.25364

[VoVAllen]

What's the tokenizer in bench?

[silver-ymz]

unicode tokenizer

[kemingy]

comparing to #10 (comment)

VectorChord-bm25

dataset	topk	qps	NDCG@10
trec-covid	10	242.72/s	0.64665
trec-covid	1000	228.10/s	0.64665
webis-touche2020	10	210.03/s	0.33956
webis-touche2020	1000	203.15/s	0.33956
fiqa	10	614.63/s	0.25465
fiqa	1000	478.72/s	0.25465

While this pre-computes the query embedding, the following computes the query embedding on the fly:

dataset	topk	qps	NDCG@10
trec-covid	10	207.38/s	0.64665
trec-covid	1000	74.87/s	0.64665
webis-touche2020	10	280.00/s	0.33956
webis-touche2020	1000	85.90/s	0.33956
fiqa	10	381.58/s	0.25465
fiqa	1000	96.47/s	0.25465

[silver-ymz]

bench after commit 908eb62

machine: 4 vCPU, 6 GB, Apple M1, lxc Virtualization

VectorChord-bm25 (unicode tokenizer)

lucene + nltk stopwords:

dataset	topk	qps	NDCG@10
trec-covid	10	140.25	0.64832
trec-covid	1000	47.02	0.64705
webis-touche2020	10	184.07	0.31007
webis-touche2020	1000	17.09	0.30877
fiqa	10	333.16	0.23657
fiqa	1000	58.47	0.23657

lucene stopwords:

dataset	topk	qps	NDCG@10
trec-covid	10	116.52	0.60726
trec-covid	1000	34.20	0.60726
webis-touche2020	10	132.25	0.33888
webis-touche2020	1000	43.50	0.33888
fiqa	10	210.70	0.2524
fiqa	1000	53.11	0.2524

elastic-search

dataset	topk	qps	NDCG@10
trec-covid	10	85.37	0.68803
trec-covid	1000	26.75	0.68803
webis-touche2020	10	13.64	0.34707
webis-touche2020	1000	8.67	0.34707
fiqa	10	134.47	0.25364
fiqa	1000	26.52	0.25364