nx_search_bench.nx source

1// nx_search_bench.nx -- REAL head-to-head onsite-search benchmark (operator: benchmark vs a real 2// incumbent, get baseline from Google Scholar). BASELINE = Google Scholar's actual ranked results for 3// "web development" (fetched live via the browse tool) with real citation counts as the authority 4// signal. METRIC = citation-graded nDCG@10 (standard IR), discount table as DATA. Compares Scholar's 5// relevance ranking vs OUR BM25 ranker over the SAME 10 results. HONEST SCOPE: 1 query, 10 docs (a real 6// slice, not full BEIR -- W-BENCH-DATA-001); data + metric REAL, nothing fabricated. Hand-written = 7// motivates X-AUT-GATE-001 (team authors benchmarks). license_tier: ORIGINAL 8import "nx_syscalls.nx" 9import "nx_itoa_lib.nx" // shared MSB-first emitter (zero-alloc) 10const K_MAGIC_22000000: i64 = 22000000 11const K_MAGIC_1269: i64 = 1269 12const K_MAGIC_1717: i64 = 1717 13 14func sb_w(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} sys_write(1,s,n); return 0 } 15// MIGRATED to the shared emitter (debt 1785563586). The old body mmapped a scratch buffer 16// per call and never freed it. At PAGE granularity that is 4096B leaked PER CALL -- the 17// defect that took 28.5GB of a 36GB host in nx_ts_lumadiff (2MB input, ~3.66M calls). 18// nxi_* is MSB-first, allocates NOTHING, and emits identical bytes including the sign. 19func sb_n(v: i64) -> i64 { nxi_out(v); return 0 } 20 21func sb_doclen(s: *u8) -> i64 { var n: i64=0; var t: i64=0; var inw: i64=0; while s[n]!=(0 as u8){ if s[n]==(32 as u8){inw=0} else { if inw==0{t=t+1;inw=1} } n=n+1 } return t } 22 23func sb_tok_end(s: *u8, start: i64) -> i64 { var e: i64=start; var fe: i64=1; while fe==1 { if s[e]==(0 as u8){fe=0} else { if s[e]==(32 as u8){fe=0} else {e=e+1} } } return e } 24 25func sb_tf(s: *u8, term: *u8) -> i64 { 26 var tf: i64=0; var i: i64=0 27 var tl: i64=0; while term[tl]!=(0 as u8){tl=tl+1} 28 var go: i64=1 29 while go==1 { 30 while s[i]==(32 as u8) { i=i+1 } 31 if s[i]==(0 as u8) { go=0 } else { 32 let e: i64 = sb_tok_end(s, i) 33 let ln: i64 = e - i 34 var mt: i64 = 0 35 if ln == tl { mt=1; var q: i64=0; while q<tl { if s[i+q]!=term[q]{mt=0} q=q+1 } } 36 if mt==1 { tf=tf+1 } 37 i = e 38 if s[i]==(0 as u8){ go=0 } 39 } 40 } 41 return tf 42} 43 44func sb_bm25(tf: i64, doclen: i64, avgdl: i64) -> i64 { 45 if avgdl<=0 { return 0 } 46 let ratio: i64=(doclen*1000)/avgdl 47 let denom: i64=tf*1000+300+(900*ratio)/1000 48 if denom<=0 { return 0 } 49 return (tf*K_MAGIC_22000000)/denom 50} 51 52// fill score[] with BM25(web)+BM25(development) per title 53func sb_score_all(titles: *i64, n: i64, score: *i64) -> i64 { 54 var tot: i64=0; var i: i64=0 55 let dls: *i64 = sys_mmap(8*32) as *i64 56 while i<n { dls[i]=sb_doclen((titles[i]) as *u8); tot=tot+dls[i]; i=i+1 } 57 let avgdl: i64 = tot/n 58 i=0 59 while i<n { 60 let twf: i64=sb_tf((titles[i]) as *u8, "web" as *u8) 61 let tdf: i64=sb_tf((titles[i]) as *u8, "development" as *u8) 62 score[i]=sb_bm25(twf,dls[i],avgdl)+sb_bm25(tdf,dls[i],avgdl) 63 i=i+1 64 } 65 return 0 66} 67 68// selection sort: fill ord[] with title indices, score descending 69func sb_rank(score: *i64, n: i64, ord: *i64) -> i64 { 70 let used: *i64 = sys_mmap(8*32) as *i64 71 var i: i64=0; while i<n { used[i]=0; i=i+1 } 72 var r: i64=0 73 while r<n { 74 var best: i64 = 0 - 1 75 var bs: i64 = 0 - 1 76 var j: i64=0 77 while j<n { if used[j]==0 { if score[j]>bs { bs=score[j]; best=j } } j=j+1 } 78 ord[r]=best; used[best]=1; r=r+1 79 } 80 return 0 81} 82 83// DCG over a ranking order: sum gain[order[r]] * disc[r] / 1000 84func sb_dcg(gain: *i64, order: *i64, disc: *i64, n: i64) -> i64 { 85 var d: i64=0; var r: i64=0 86 while r<n { d = d + (gain[order[r]] * disc[r]) / 1000; r=r+1 } 87 return d 88} 89 90func main() -> i64 { 91 let N: i64 = 10 92 let titles: *i64 = sys_mmap(8*16) as *i64 93 titles[0]="php and mysql web development" as *u8 as i64 94 titles[1]="a systematic review of usability evaluation in web development" as *u8 as i64 95 titles[2]="evolutionary web development" as *u8 as i64 96 titles[3]="user centered web development" as *u8 as i64 97 titles[4]="flask web development" as *u8 as i64 98 titles[5]="web design a key factor for the website success" as *u8 as i64 99 titles[6]="web design" as *u8 as i64 100 titles[7]="python web development with django" as *u8 as i64 101 titles[8]="web development estimating quick to market software" as *u8 as i64 102 titles[9]="collaborative web development strategies and best practices for web teams" as *u8 as i64 103 let cites: *i64 = sys_mmap(8*16) as *i64 104 cites[0]=K_MAGIC_1269; cites[1]=130; cites[2]=161; cites[3]=191; cites[4]=K_MAGIC_1717 105 cites[5]=264; cites[6]=291; cites[7]=242; cites[8]=330; cites[9]=169 106 let disc: *i64 = sys_mmap(8*16) as *i64 107 disc[0]=1000; disc[1]=631; disc[2]=500; disc[3]=431; disc[4]=387 108 disc[5]=356; disc[6]=333; disc[7]=315; disc[8]=301; disc[9]=289 109 110 let sch: *i64 = sys_mmap(8*16) as *i64 111 var i: i64=0; while i<N { sch[i]=i; i=i+1 } 112 113 let score: *i64 = sys_mmap(8*16) as *i64 114 sb_score_all(titles, N, score) 115 let ours: *i64 = sys_mmap(8*16) as *i64 116 sb_rank(score, N, ours) 117 118 let idc: *i64 = sys_mmap(8*16) as *i64 119 i=0; while i<N { idc[i]=cites[i]; i=i+1 } 120 i=0; while i<N { var j: i64=i+1; while j<N { if idc[j]>idc[i] { let t: i64=idc[i]; idc[i]=idc[j]; idc[j]=t } j=j+1 } i=i+1 } 121 let ideal: *i64 = sys_mmap(8*16) as *i64 122 i=0; while i<N { ideal[i]=i; i=i+1 } 123 124 let dcg_sch: i64 = sb_dcg(cites, sch, disc, N) 125 let dcg_our: i64 = sb_dcg(cites, ours, disc, N) 126 let dcg_idl: i64 = sb_dcg(idc, ideal, disc, N) 127 var ndcg_sch: i64=0 128 var ndcg_our: i64=0 129 if dcg_idl>0 { ndcg_sch=(dcg_sch*1000)/dcg_idl } 130 if dcg_idl>0 { ndcg_our=(dcg_our*1000)/dcg_idl } 131 132 sb_w("SEARCH-BENCH query=web-development corpus=GoogleScholar-top10 metric=citation-graded-nDCG10\n" as *u8) 133 sb_w(" incumbent_GoogleScholar nDCG_permil=" as *u8); sb_n(ndcg_sch); sb_w("\n" as *u8) 134 sb_w(" ours_BM25_over_titles nDCG_permil=" as *u8); sb_n(ndcg_our); sb_w("\n" as *u8) 135 var verdict: *u8 = "BEHIND" as *u8 136 if ndcg_our > ndcg_sch { verdict="AHEAD" as *u8 } 137 if ndcg_our == ndcg_sch { verdict="PARITY" as *u8 } 138 sb_w(" HEAD-TO-HEAD=" as *u8); sb_w(verdict); sb_w(" (real data+metric; BM25-titles is the FLOOR, 2026 exceed=hybrid+rerank W-SRCH-EX-001)\n" as *u8) 139 sys_exit(0) 140 return 0 141}