haystack-us23/index.html at master · metarank/haystack-us23 · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!doctype html>
<html lang="en">
	<head>
		<meta charset="utf-8">
		<meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no">

		<title>Learning to Hybrid Search</title>

		<link rel="stylesheet" href="dist/reset.css">
		<link rel="stylesheet" href="dist/reveal.css">
		<link rel="stylesheet" href="dist/theme/dracula.css">

		<!-- Theme used for syntax highlighted code -->
		<link rel="stylesheet" href="plugin/highlight/monokai.css">
	</head>
	<body>
		<div class="reveal">
			<div class="slides">
				<section>
					<h1>Learning to Hybrid Search:</h1>
					<h3>mixing BM25, LLMs and metadata into an ultimate ranking ensemble</h3>
					<br>
					<small>Grebennikov Roman | DeliveryHero SE | Haystack US 2023</small>
				</section>
				<section>
          <h2>This is me</h2>
          <img src="img/screaming_sun.png" height="300px">
          <ul>
              <li>Long ago: PhD in CS, quant trading, credit scoring</li>
              <li>Search &amp; personalization for ~7 years</li>
              <li><strong>Metarank</strong> maintainer</li>
              <li><strong>DeliveryHero</strong>: search & relevance</li>
          </ul>
				</section>
				<section>
					<h2>A typical "hybrid search" talk:</h2>
					<img src="img/opinion.gif">
					<ul class="fragment">
						<li>A is <strong>[maybe]</strong> better than B</li>
						<li>Large C is <strong>[a bit]</strong> smarter than small D</li>
						<li>You <strong>[probably]</strong> need E</li>
					</ul>
				</section>
				<section>
					<h2>No numbers, no comparisons</h2>
					<ul>
						<li class="fragment"><strong>Reproducability</strong>: proprietary tools and data</li>
						<li class="fragment"><strong>Legal</strong>: insider information</li>
					</ul>
				</section>
				<section>
					<h2>Many companies, same problems</h2>
					<p><img src="img/breast.png" height="400px" style="margin: 0px;"></p>
					<ul>
						<li class="fragment"><strong>Relevance</strong>: making visitor happier</li>
						<li class="fragment"><strong>CTR/Conversion</strong>: making company happier</li>
						<li class="fragment"><strong>Query understanding</strong>: language, intent, semantics</li>
					</ul>
				</section>
				<section>
					<h2>Open data + open tools</h2>
					<p>One step closer to reproducability?</p>
					<ul>
						<li class="fragment">TREC = ❤️</li>
						<li class="fragment">There <s>are</s> were no e-commerce datasets!</li>
					</ul>
				</section>
				<section>
					<h2>Amazon ESCI: open e-com dataset</h2>
					<p><img src="img/esci.png" height="350px"></p>
					<ul>
						<li>130k "hard queries" in EN/ES/JP</li>
						<li>1.7M products</li>
						<li>2.6M explicit relevance labels</li>
					</ul>
				</section>
				<section>
					<h2>130k hard queries</h2>
					<p><img src="img/queries.png" height="400px"></p>
				</section>
				<section>
					<h2>2.6M explicit ESCI labels</h2>
					<table>
						<tr>
							<td><img src="img/labels.png" height="400px"></td>
							<td style="vertical-align: middle;">
								<ul>
									<li><strong>E</strong>: exact match</li>
									<li><strong>S</strong>: substitute</li>
									<li><strong>C</strong>: complementary</li>
									<li><strong>I</strong>: irrelevant</li>
								</ul>

							</td>
						</tr>
					</table>
				</section>
				<section>
					<h2>1.7M real products</h2>
					<pre><code data-trim data-noescape data-line-numbers class="json">
  {
    "product_id": "B07PT4BXWK",
    "product_title": "Country Carrot, Artificial Vegetable Fake Food,
                      Bag of 12, 2 Sizes",
    "product_description": "12 pcs vegetables atractive Artificial product.
                      Orange color with green stem, great for decorating.
                      Large size, approximately 6-inches (without stem).",
    "product_bullet_point": "",
    "product_brand": "Mezly",
    "product_color": "Orange",
    "product_locale": "us"
  }
					</code></pre>
					<ul class="fragment">
						<li>Text focused: no categories, metadata, reviews</li>
						<li>product_id = ASIN</li>
					</ul>
				</section>
				<section>
					<img src="img/amazon.png" height="600px">
				</section>
				<section>
					<h2>ESCI-S: an extension to the ESCI</h2>
					<img src="img/escis.png" height="400px" style="margin: 0px;">
					<p>Scraped metadata for 92% of ESCI ASINs</p>
					<p><a href="https://github.com/shuttie/esci-s">https://github.com/shuttie/esci-s</a></p>
				</section>
				<section>
					<h2>Agenda</h2>
					<p>Take the ESCI and do the "hybrid search"</p>
					<ul>
						<li class="fragment"><strong>Baseline</strong>: BM25</li>
						<li class="fragment"><strong>LambdaMART</strong>: BM25 and metadata</li>
						<li class="fragment"><strong>Bi-encoders</strong>: the notorious <strong>all-MiniLM-L6-v2</strong></li>
						<li class="fragment"><strong>Fine-tuning</strong> bi-encoders</li>
						<li class="fragment"><strong>Cross-encoders</strong>: off-the-shelf and fine-tuned</li>
					</ul>
					<p class="fragment">Minimal amount of ad-hoc Python code: <strong>Metarank</strong>!</p>
				</section>
				<section>
					<p><img src="img/welcome.png" height="350px"></p>
					<p>Hybrid search = a glorified Ensebmle Learning</p>
					<ul>
						<li class="fragment">Take N weak predictors</li>
						<li class="fragment">Combine into an ensemble model</li>
						<li class="fragment">"number of stars" = weak predictor</li>
					</ul>
				</section>
				<section>
					<h2>Linear hybrid search</h2>
					<p>query-title score distribution on ESCI:</p>
					<table>
						<tr>
							<td><img src="img/bm25-dist.png"></td>
							<td><img src="img/cos-dist.png"></td>
						</tr>
					</table>
					<p class="fragment">score = w<sub>1</sub> * norm(bm25) + w<sub>2</sub> * norm(cos)</p>
				</section>
				<section>
					<h2>Linear hybrid search</h2>
					<p>Typical problems of</p>
					<small>score = w<sub>1</sub> * norm(bm25) + w<sub>2</sub> * norm(cos)</small>
					<br><br>
					<ul>
						<li class="fragment"><strong>Normalization</strong>: ad-hoc based on distribution</li>
						<li class="fragment"><strong>Correlations</strong>: weights for 7-30-90 days CTR</li>
						<li class="fragment"><strong>Linear</strong>: misses non-linear dependencies</li>
					</ul>
				</section>
				<section>
					<h2>Why Metarank?</h2>
					<p><img src="img/reranking.png" height="250px"></p>
					<ul>
						<li class="fragment">LambdaMART: NDCG, non-linear, correlations</li>
						<li class="fragment">An open-source secondary reranker</li>
						<li class="fragment"><strong>Search</strong>: focus on recall and speed</li>
						<li class="fragment"><strong>Metarank</strong>: focus on top-N precision</li>
					</ul>
				</section>
				<section>
					<h2>How it works</h2>
					<p><img src="img/deployment.png" height="250px"></p>
					<ul>
						<li class="fragment">Standalone Mode: no state, just files</li>
						<li class="fragment">Import existing dataset, build ranking features</li>
						<li class="fragment">Get NDCG over test set</li>
					</ul>
				</section>
				<section>
					<h2>Baseline: random ranking</h2>
					<p>Why do we need it?</p>
					<ul>
						<li>High rate of relevant products = high NDCG</li>
						<li>NDCG lower bound</li>
					</ul>
					<pre><code data-trim data-noescape data-line-numbers class="yaml">
features:
  - type: random
    name: random

models:
  haystack-demo:
    type: lambdamart
    backend:
      type: xgboost
    features:
      - random
					</code></pre>
				</section>
				<section>
					<video data-autoplay src="img/metarank-demo.mp4"></video>
				</section>
				<section>
					<h2>Baseline: random ranking</h2>
					<table>
						<thead>
							<td></td>
							<td>NDCG@5</td>
							<td>NDCG@10</td>
							<td>NDCG@20</td>
						</thead>
						<tr>
							<td>Random</td>
							<td>0.6588</td>
							<td>0.7220</td>
							<td>0.8234</td>
						</tr>
					</table>
				</section>
				<section>
					<h2>Baseline: BM25</h2>
					<ul>
						<li class="fragment"><strong>Multiple fields in ES</strong>: optimal boosts?</li>
						<li class="fragment"><strong>ES-LTR</strong>: hard to setup</li>
						<li class="fragment">Compute BM25 without Lucene/ES? 🤔</li>
					</ul>
				</section>
				<section>
					<h2>Down the BM25 hole</h2>
					<img src="img/bm25.png">
					<ul>
						<li>Only term frequencies are needed</li>
						<li>k<sub>1</sub>=1.2, b=0.75 as in Lucene/ES</li>
					</ul>
				</section>
				<section>
					<h2>BM25 without Lucene</h2>
					<pre><code data-trim data-noescape data-line-numbers class="yaml">
 - type: field_match
   name: query_desc_bm25
   itemField: item.title
   rankingField: ranking.query
   method:
     type: bm25
     language: english
     termFreq: termfreq.json

					</code></pre>
					<table class="fragment">
						<thead>
							<td></td>
							<td>NDCG@5</td>
							<td>NDCG@10</td>
							<td>NDCG@20</td>
						</thead>
						<tr>
							<td>Random</td>
							<td>0.6588</td>
							<td>0.7220</td>
							<td>0.8234</td>
						</tr>

						<tr>
							<td><strong>BM25(title)</strong></td>
							<td><strong>0.7555</strong></td>
							<td><strong>0.7966</strong></td>
							<td><strong>0.8711</strong></td>
						</tr>
					</table>
				</section>
				<section>
					<h2>BM25 over multiple fields</h2>
					<p>per-field score = weak predictor</p>
					<pre><code data-trim data-noescape data-line-numbers class="yaml">
features:
  - query_title_bm25
  - query_desc_bm25
  - query_bullets_bm25
					</code></pre>
					<table class="fragment">
						<thead>
							<td></td>
							<td>NDCG@5</td>
							<td>NDCG@10</td>
							<td>NDCG@20</td>
						</thead>
						<tr>
							<td>Random</td>
							<td>0.6588</td>
							<td>0.7220</td>
							<td>0.8234</td>
						</tr>
						<tr>
							<td>BM25(title)</td>
							<td>0.7555</td>
							<td>0.7966</td>
							<td>0.8711</td>
						</tr>
						<tr>
							<td><strong>BM25(title, desc, bullets)</strong></td>
							<td><strong>0.7620</strong></td>
							<td><strong>0.8023</strong></td>
							<td><strong>0.8740</strong></td>
						</tr>
					</table>
				</section>
				<section>
					<img src="img/pathetic.jpg" height="300px">
					<pre><code data-trim data-noescape data-line-numbers class="yaml">
query: lawnmower tires without rims

title: Dr.Roc Tire Spoon Lever Dirt Bike Lawn Mower Motorcycle
       Tire Changing Tools with Durable Bag 3 Tire Irons 2
       Rim Protectors 1 Valve Stems Set TR412 TR413

title: MaxAuto 13x5.00-6 Lawn Mower Tires with Rim 13x5.00-6
       Tire and Wheel 13x5-6 NHS Tire 13x5x6 Pneumatic Tire
					</code></pre>
					<p>Years of Amazon-specific SEO optimization</p>
				</section>

				<section>
					<h2>LTR: categorical features</h2>
					<p><img src="img/dino.jpg" height="200px" style="margin: 0px;"></p>
					<pre><code data-trim data-noescape data-line-numbers class="yaml">
  - type: string
    name: material
    field: item.material // color, category[1,2,3]
    scope: item
    encode: index // Supports XGBoost 1.7+/LightGBM 3+/Catboost
                  // native categorical features
    values:
    - fabric
    - glass
    - leather
    - ...
</code></pre>
				</section>
				<section>
					<h2>LTR: numeric features</h2>
					<p><img src="img/dino2.jpg" height="250px" style="margin: 0px;"></p>
					<pre><code data-trim data-noescape data-line-numbers class="yaml">
  - type: number
    name: stars
    field: item.stars
    scope: item
					</code></pre>
					<p># stars, # ratings, price, weight</p>
				</section>
				<section>
					<h2>LTR: 2017 edition</h2>
					<pre><code data-trim data-noescape data-line-numbers class="yaml">
    features: [query_*_bm25, category*,
               color,        material,
               price,        ratings,
               stars,        template,
               weight]
					</code></pre>
					<table>
						<thead>
							<td></td>
							<td>NDCG@5</td>
							<td>NDCG@10</td>
							<td>NDCG@20</td>
						</thead>
						<tr>
							<td>Random</td>
							<td>0.6588</td>
							<td>0.7220</td>
							<td>0.8234</td>
						</tr>
						<tr>
							<td>BM25(title)</td>
							<td>0.7555</td>
							<td>0.7966</td>
							<td>0.8711</td>
						</tr>
						<tr>
							<td>BM25(title, desc, bullets)</td>
							<td>0.7620</td>
							<td>0.8023</td>
							<td>0.8740</td>
						</tr>
						<tr>
							<td><strong>LMart(bm25, meta)</strong></td>
							<td><strong>0.7675</strong></td>
							<td><strong>0.8075</strong></td>
							<td><strong>0.8769</strong></td>
						</tr>
					</table>
				</section>
				<section>
					<h1>Hybrid Search</h1>
				</section>
				<section>
					<h2>Traditional hybrid search</h2>
					<p><img src="img/hybrid-search.png" height="300px" style="margin: 0px;"></p>
					<p>Crafting embeddings:</p>
					<ul>
						<li class="fragment"><strong>Commercial providers</strong>: Cohere, OpenAI</li>
						<li class="fragment"><strong>Pre-trained</strong>: sentence-transformers</li>
						<li class="fragment"><strong>DIY</strong>: fine-tuning</li>
					</ul>
				</section>
				<section>
					<h2>Siamese BERT networks</h2>
					<p><img src="img/sbert.png" height="300px"></p>
					<ul>
						<li class="fragment">Document indexing can happen <strong>offline</strong>!</li>
						<li class="fragment">Only need to embed <strong>query</strong> during search</li>
						<li class="fragment">HNSW for Approximate k-NN</li>
					</ul>
				</section>
				<section>
					<h2>Pre-trained all-MiniLM-L6-v2</h2>
					<img src="img/st.png" height="400px">
					<ul>
						<li>Small and fast for CPU inference</li>
						<li>Still precise on semantic search</li>
					</ul>
				</section>
				<section>
					<h2>all-MiniLM-L6-v2 in Python</h2>
					<pre><code data-trim data-noescape data-line-numbers class="python">
from sentence_transformers import SentenceTransformer, util
model = SentenceTransformer('all-MiniLM-L6-v2')

#Sentences are encoded by calling model.encode()
emb1 = model.encode("This is a red cat with a hat.")
emb2 = model.encode("Have you seen my red cat?")

cos_sim = util.cos_sim(emb1, emb2)
print("Cosine-Similarity:", cos_sim)
					</code></pre>
					<p class="fragment">Fine, but our search stack is in Java!</p>
				</section>
				<section>
					<h2>all-MiniLM-L6-v2 in Java/ONNX</h2>
					<img src="img/owl.jpg" height="300px" style="margin: 0px;">
					<pre><code data-trim data-noescape data-line-numbers class="scala fragment">
val text = "This is a red cat with a hat"
val vocab =  DefaultVocabulary.builder().addFromTextFile(...).build()
val bert = new BertFullTokenizer(vocab, true)
val tokensStrings = "[CLS]" :: bert.tokenize(text) :: "[SEP]"

val tokens = tokensStrings.map(t => vocab.getIndex(t)).toArray
val attmask  = Array.fill(tokens.length)(1L)
val tokenTypes = Array.fill(tokens.length)(0L)
					</code></pre>
				</section>
				<section>
					<h2>all-MiniLM-L6-v2 in Java/ONNX</h2>
					<pre><code data-trim data-noescape data-line-numbers class="scala">
val env     = OrtEnvironment.getEnvironment()
val session = env.createSession("minilm-v2.onnx", new SessionOptions())
val size = Array(1, tokens.length)

val tensor1 = OnnxTensor.createTensor(env, LongBuffer.wrap(tokens), size)
val tensor2 = OnnxTensor.createTensor(env, LongBuffer.wrap(attmask), size)
val tensor3 = OnnxTensor.createTensor(env, LongBuffer.wrap(tokenTypes), size)
val out = session.run(Map(
  "input_ids" -> tensor1,
  "token_type_ids" -> tensor3,
  "attention_mask" -> tensor2).asJava)

val preds = out.get(0).getValue.asInstanceOf[Array[Array[Array[Float]]]

					</code></pre>
					<p class="fragment">As seen in: Vespa, OpenSearch Model Serving, Metarank</p>
				</section>
				<section>
					<h2>Bi-encoders in practice</h2>
					<pre><code data-trim data-noescape data-line-numbers class="yaml">
  - type: field_match
    name: query_title_minilm6
    itemField: item.title
    rankingField: ranking.query
    method:
      type: bi-encoder
      model: metarank/all-MiniLM-L6-v2
      dim: 384
					</code></pre>
					<table class="fragment">
						<thead>
							<td></td>
							<td>NDCG@5</td>
							<td>NDCG@10</td>
							<td>NDCG@20</td>
						</thead>
						<tr>
							<td>Random</td>
							<td>0.6588</td>
							<td>0.7220</td>
							<td>0.8234</td>
						</tr>
						<tr>
							<td>BM25(title, desc, bullets)</td>
							<td>0.7620</td>
							<td>0.8023</td>
							<td>0.8740</td>
						</tr>
						<tr>
							<td>LMart(bm25, meta)</td>
							<td>0.7675</td>
							<td>0.8075</td>
							<td>0.8769</td>
						</tr>
						<tr>
							<td><strong>all-MiniLM-L6-v2</strong></td>
							<td><strong>0.7670</strong></td>
							<td><strong>0.8069</strong></td>
							<td><strong>0.8775</strong></td>
						</tr>
					</table>
				</section>
				<section>
					<h2>Bi-encoder models</h2>
					<ul>
						<li>metarank/<strong>all-MiniLM-L6-v2</strong>: 10M, 386 dims</li>
						<li>metarank/<strong>all-MiniLM-L12-v2</strong>: 15M, more layers</li>
						<li>metarank/<strong>all-mpnet-base-v2</strong>: 100M, 768 dims</li>
					</ul>
					<br><br>
					<table class="fragment">
						<thead>
							<td></td>
							<td>NDCG@5</td>
							<td>NDCG@10</td>
							<td>NDCG@20</td>
						</thead>
						<tr>
							<td>all-MiniLM-L6-v2</td>
							<td>0.7670</td>
							<td>0.8069</td>
							<td>0.8775</td>
						</tr>
						<tr>
							<td>all-MiniLM-L12-v2</td>
							<td>0.7685</td>
							<td>0.8077</td>
							<td>0.8786</td>
						</tr>
						<tr>
							<td><strong>all-mpnet-base-v2</strong></td>
							<td><strong>0.7694</strong></td>
							<td><strong>0.8089</strong></td>
							<td><strong>0.8794</strong></td>
						</tr>
					</table>
				</section>
				<section>
					<h2>Bring-your-own-model with ONNX</h2>
					<p>ONNX: Vespa, OpenSearch Model Serving, Metarank</p>
					<pre><code data-trim data-noescape class="bash">

$> pip install transformers[onnx] sentence-transformers
$> python -m transformers.onnx \
          --model=sentence-transformers/all-MiniLM-L6-v2 \
          export_dir
					</code></pre>
				</section>
				<section>
					<h2>LTR with bi-encoders</h2>
					<p>Cosine similarity = yet another ranking feature</p>
					<table class="fragment">
						<thead>
							<td></td>
							<td>NDCG@5</td>
							<td>NDCG@10</td>
							<td>NDCG@20</td>
						</thead>
						<tr>
							<td>BM25(title, desc, bullets)</td>
							<td>0.7620</td>
							<td>0.8023</td>
							<td>0.8740</td>
						</tr>
						<tr>
							<td>LM(bm25, meta)</td>
							<td>0.7675</td>
							<td>0.8075</td>
							<td>0.8769</td>
						</tr>
						<tr>
							<td>all-MiniLM-L6-v2</strong></td>
							<td>0.7670</td>
							<td>0.8069</td>
							<td>0.8775</td>
						</tr>
						<tr>
							<td><strong>LM(bm25, minilm)</strong></td>
							<td><strong>0.7816</strong></td>
							<td><strong>0.8168</strong></td>
							<td><strong>0.8840</strong></td>
						</tr>
						<tr>
							<td><strong>LM(bm25, minilm, meta)</strong></td>
							<td><strong>0.7825</strong></td>
							<td><strong>0.8186</strong></td>
							<td><strong>0.8846</strong></td>
						</tr>
					</table>
				</section>
				<section>
					<h2>Fine-tuning?</h2>
				</section>
				<section>
					<img src="img/easy.png" height="650px">
				</section>
				<section>
					<h2>What is fine-tuning?</h2>
					<img src="img/fine-tuning.png" height="350px" style="margin: 0px;">
					<ul>
						<li><strong>Existing LLM</strong> as a source</li>
						<li><strong>Custom loss</strong>: minimize cosine distance</li>
						<li><strong>Slightly</strong> update weights on a new dataset</li>
					</ul>
				</section>
				<section>
					<h2>Which LLM to pick?</h2>
					<ul>
						<li class="fragment">
							<strong>MS MiniLM, T5, GPT-J</strong>:
							<ul>
								<li>pro: any model from HuggingFace</li>
								<li>con: no clue about sentence similarity</li>
								<li>con: needs a large dataset, slow</li>
							</ul>
						</li>
						<li class="fragment"><strong>sentence-transformers</strong>:
							<ul>
								<li>pro: 1K queries is a good start</li>
								<li>con: no recent models</li>
							</ul>
						</li>
					</ul>
				</section>
				<section>
					<h2>CosineSimilarityLoss</h2>
					<p>Minimizes the distance within each sample</p>
					<p class="fragment">Needs triplets of query-document-score:</p>
					<ul class="fragment">
						<li><strong>Linear mapping</strong>: E-S-C-I to [1.0, 0.66, 0.33, 0.0]</li>
						<li><strong>Exponential mapping</strong>: E-S-C-I to [1.0, 0.1, 0.001, 0.0]</li>
					</ul>
					<p class="fragment">From docs: no need for hard negatives, slow to converge</p>
				</section>
				<section>
					<h2>MultipleNegativesRankingLoss</h2>
					<p class="fragment">Needs triplets of query-positive-negative:</p>
					<ul class="fragment">
						<li><strong>E/I</strong>: Exact = positive, Irrelevant = negative</li>
						<li><strong>E/SCI</strong>: Exact = positive, other = negative</li>
					</ul>
					<p style="margin: 0px" class="fragment"><img src="img/hn.png" height="300px" style="margin: 0px"></p>
					<p class="fragment">From docs: hard negatives needed, fast to converge</p>
				</section>
				<section>
					<h2>Quality of negatives</h2>
					<ul>
						<li>Easy: distinguish dogs from trains</li>
						<li>Hard: distinguish breeds of dogs</li>
					</ul>
					<p><img src="img/chi.jpg" height="400px"></p>
				</section>
				<section>
					<h2>Implicit feedback as labels</h2>
					<p><img src="img/implicit.png" height="550px"></p>
				</section>
				<section>
					<h2>Implicit feedback as labels</h2>
					<p><img src="img/implicit.png" height="350px" style="margin: 0px;"></p>
					<ul>
						<li><strong>Green</strong>: high per-query CTR, high confidence</li>
						<li><strong>Yellow</strong>: low confidence</li>
						<li><strong>Red</strong>: low per-query CTR, high confidence</li>
					</ul>
				</section>
				<section>
					<h2>Fine-tuning is easy</h2>
					<p>(when you have a GPU)</p>
					<pre><code data-trim data-noescape class="python">
model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')

train_dataset = ESCIDataset(input='train-small.json.gz')
eval_dataset = ESCIDataset(input='test-small.json.gz')
train_dataloader = DataLoader(train_dataset, shuffle=True)
train_loss = losses.CosineSimilarityLoss(model=model)


model.fit(train_objectives=[(train_dataloader, train_loss)],
          epochs=1,
          optimizer_params = {'lr': 1e-5},
          output_path=model_save_path)

model.save(model_save_path)
					</code></pre>
				</section>
				<section>
					<h2>Things to watch for</h2>
					<ul>
						<li><strong>Batch size</strong>: the largest value your GPU can fit</li>
						<li><strong>Epochs</strong>: 1-2, quick over-fitting</li>
						<li><strong>Cheap GPUs</strong>: Google Colab, Kaggle, vast.ai</li>
					</ul>
				</section>
				<section>
					<h2>CPU vs GPU</h2>
					<p><img src="img/not-hard.png" height="350px" style="margin: 0px;"></p>
					<p>all-MiniLM-L6-v2 over ESCI dataset:</p>
					<ul>
						<li><strong>RTX 3060</strong>: 6 minutes</li>
						<li><strong>Ryzen 7 2600</strong>: 240 minutes</li>
					</ul>
				</section>
				<section>
					<h2>Fine-tuned bi-encoders</h2>
					<table>
						<thead>
							<td></td>
							<td>NDCG@5</td>
							<td>NDCG@10</td>
							<td>NDCG@20</td>
						</thead>
						<tr>
							<td>BM25(title, desc, bullets)</td>
							<td>0.7620</td>
							<td>0.8023</td>
							<td>0.8740</td>
						</tr>
						<tr>
							<td>all-MiniLM-L6</strong></td>
							<td>0.7670</td>
							<td>0.8069</td>
							<td>0.8775</td>
						</tr>
						<tr class="fragment">
							<td>esci-MiniLM-L6, MN e/i</td>
							<td>0.7783</td>
							<td>0.8161</td>
							<td>0.8834</td>
						</tr>
						<tr class="fragment">
							<td>esci-MiniLM-L6, MN e/sci</td>
							<td>0.7838</td>
							<td>0.8212</td>
							<td>0.8864</td>
						</tr>
						<tr class="fragment">
							<td>esci-MiniLM-L6, cos+lin</td>
							<td>0.7800</td>
							<td>0.8184</td>
							<td>0.8845</td>
						</tr>
						<tr class="fragment">
							<td><strong>esci-MiniLM-L6, cos+exp</strong></td>
							<td><strong>0.7881</strong></td>
							<td><strong>0.8240</strong></td>
							<td><strong>0.8874</strong></td>
						</tr>
					</table>

				</section>
				<section>
					<h2>L12 and MPNET fine-tuning</h2>
					<img src="img/deeper.jpg" height="200px" style="margin: 0px;">
					<table>
						<thead>
							<td></td>
							<td>NDCG@5</td>
							<td>NDCG@10</td>
							<td>NDCG@20</td>
						</thead>
						<tr>
							<td>BM25(title, desc, bullets)</td>
							<td>0.7620</td>
							<td>0.8023</td>
							<td>0.8740</td>
						</tr>
						<tr>
							<td><strong>esci-MiniLM-L6, cos+exp</strong></td>
							<td><strong>0.7881</strong></td>
							<td><strong>0.8240</strong></td>
							<td><strong>0.8874</strong></td>
						</tr>
						<tr class="fragment">
							<td>esci-MiniLM-L12, MN e/sci</td>
							<td>0.7818</td>
							<td>0.8199</td>
							<td>0.8858</td>
						</tr>
						<tr class="fragment">
							<td><strong>esci-mpnet, cos+exp</strong></td>
							<td><strong>0.7990</strong></td>
							<td><strong>0.8329</strong></td>
							<td><strong>0.8926</strong></td>
						</tr>
					</table>
				</section>
				<section>
					<h2>LTR and fine-tuned bi-encoders</h2>
					<p>LTR never makes things worse</p>
					<table>
						<thead>
							<td></td>
							<td>NDCG@5</td>
							<td>NDCG@10</td>
							<td>NDCG@20</td>
						</thead>
						<tr>
							<td>BM25(title, desc, bullets)</td>
							<td>0.7620</td>
							<td>0.8023</td>
							<td>0.8740</td>
						</tr>
						<tr>
							<td>LM(bm25, minilm, meta)</td>
							<td>0.7825</td>
							<td>0.8186</td>
							<td>0.8846</td>
						</tr>
						<tr>
							<td><strong>LM(bm25, mpn-ft, meta)</strong></td>
							<td><strong>0.8053</strong></td>
							<td><strong>0.8390</strong></td>
							<td><strong>0.8959</strong></td>
						</tr>
					</table>
				</section>
				<section>
					<h2>Bi-encoder performance</h2>
					<pre><code data-trim data-noescape class="csv">
Benchmark     (batch)            (model)  Mode  Cnt   Score   Error  Units
encode              1   all-MiniLM-L6-v2  avgt   30  13.350 ± 1.813  ms/op
encode              1  all-MiniLM-L12-v2  avgt   30  25.954 ± 2.210  ms/op
encode              1  all-mpnet-base-v2  avgt   30  64.208 ± 2.932  ms/op
					</code></pre>
				</section>
				<section>
					<h2>Conclusions</h2>
					<ul>
						<li class="fragment">Larger the LLM - smaller the training set</li>
						<li class="fragment">Fine-tuning: simple way for a HUGE relevance boost</li>
						<li class="fragment">ESCI != e-commerce</li>
						<li class="fragment">Large LLMs are slow on CPU</li>
					</ul>
				</section>

				<section>
					<h1>Cross-encoders</h1>
				</section>
				<section>
					<h2>Cross-encoders</h2>
					<p><img src="img/cross.png" height="300px"></p>
					<ul>
						<li>Classifier head on top of BERT</li>
						<li>No cosine similarity: let the LLM work</li>
					</ul>
				</section>
				<section>
					<h2>Pre-trained CE</h2>
                     <img src="img/st-ce.png">
					<pre><code data-trim data-noescape class="python fragment">
from sentence_transformers import CrossEncoder

model = CrossEncoder('model_name', max_length=512)
scores = model.predict([('How many people live in Berlin?',
                         'Berlin had a population of 3,520,031 people.'),
                        ('How many people live in Berlin?',
                         'Berlin is well known for its museums.')])
                     </code></pre>
				</section>
				<section>
					<h2>CE: pros & cons</h2>
					<p><img src="img/ce-cat.jpg" height="400px"></p>
					<ul>
						<li>ChatGPT: "how relevant is text A for query B"</li>
						<li>No Approximate k-NN search, only top-N</li>
					</ul>
				</section>
				<section>
					<h2>Generic SBERT</h2>
					<pre><code data-trim data-noescape class="python">
  - type: field_match
    name: query_title_ce
    itemField: item.title
    rankingField: ranking.query
    method:
      type: cross-encoder
      model: metarank/ce-all-MiniLM-L6-v2
      dim: 384

					</code></pre>
						<table>
						<thead>
							<td></td>
							<td>NDCG@5</td>
							<td>NDCG@10</td>
							<td>NDCG@20</td>
						</thead>
						<tr>
							<td>BM25(title, desc, bullets)</td>
							<td>0.7620</td>
							<td>0.8023</td>
							<td>0.8740</td>
						</tr>
						<tr>
							<td>all-MiniLM-L6</td>
							<td>0.7670</td>
							<td>0.8069</td>
							<td>0.8775</td>
						</tr>
						<tr>
							<td>esci-mpnet</td>
							<td>0.7990</td>
							<td>0.8329</td>
							<td>0.8926</td>
						</tr>
						<tr>
							<td><strong>ce-msmarco-MiniLM-L6</strong></td>
							<td><strong>0.7767</strong></td>
							<td><strong>0.8120</strong></td>
							<td><strong>0.8815</strong></td>
						</tr>
					</table>
				</section>
				<section>
					<h2>Fine-tuning cross-encoders</h2>
					<pre><code data-trim data-noescape class="python fragment">
model = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2', num_labels=1)

train_dataset = ESCIDataset(input='train-small.json.gz')
train_dataloader = DataLoader(train_dataset, shuffle=True)

model.fit(train_dataloader=train_dataloader,
          epochs=1,
          optimizer_params = {'lr': 1e-5},
          output_path=model_save_path)

model.save(model_save_path)
					</code></pre>
				</section>
				<section>
					<h2>CE fine-tuning results</h2>
						<table>
						<thead>
							<td></td>
							<td>NDCG@5</td>
							<td>NDCG@10</td>
							<td>NDCG@20</td>
						</thead>
						<tr>
							<td>BM25(title, desc, bullets)</td>
							<td>0.7620</td>
							<td>0.8023</td>
							<td>0.8740</td>
						</tr>
						<tr>
							<td>all-MiniLM-L6</td>
							<td>0.7670</td>
							<td>0.8069</td>
							<td>0.8775</td>
						</tr>
						<tr>
							<td>esci-mpnet</td>
							<td>0.7990</td>
							<td>0.8329</td>
							<td>0.8926</td>
						</tr>
						<tr>
							<td>ce-msmarco-MiniLM-L6</td>
							<td>0.7767</td>
							<td>0.8120</td>
							<td>0.8815</td>
						</tr>
						<tr class="fragment">
							<td><strong>ce-esci-MiniLM-L6</strong></td>
							<td><strong>0.8062</strong></td>
							<td><strong>0.8364</strong></td>
							<td><strong>0.8963</strong></td>
						</tr>
						<tr class="fragment">
							<td><strong>ce-esci-MiniLM-L12</strong></td>
							<td><strong>0.8134</strong></td>
							<td><strong>0.8426</strong></td>
							<td><strong>0.9001</strong></td>
						</tr>
					</table>