La recherche à l’ère de l’IA

A presentation at Devoxx France 2024 in April 2024 in Paris, France by David Pilato

Slide 1

Slide 1

DEVOXX FRANCE 2024 La recherche À l’ère de l’IA David Pilato | @dadoonet

Slide 2

Slide 2

Agenda ● ● ● ● Co m m er ci “Classic” search and its limitations ML model and usage Vector search or hybrid search in Elasticsearch OpenAI’s ChatGPT or LLMs with Elasticsearch al

Slide 3

Slide 3

Elasticsearch You Know, for Search

Slide 4

Slide 4

Slide 5

Slide 5

Slide 6

Slide 6

These are not the droids you are looking for.

Slide 7

Slide 7

GET /_analyze { “char_filter”: [ “html_strip” ], “tokenizer”: “standard”, “filter”: [ “lowercase”, “stop”, “snowball” ], “text”: “These are <em>not</em> the droids you are looking for.” }

Slide 8

Slide 8

These are <em>not</em> the droids you are looking for. { “tokens”: [{ “token”: “droid”, “start_offset”: 27, “end_offset”: 33, “type”: “<ALPHANUM>”, “position”: 4 },{ “token”: “you”, “start_offset”: 34, “end_offset”: 37, “type”: “<ALPHANUM>”, “position”: 5 }, { “token”: “look”, “start_offset”: 42, “end_offset”: 49, “type”: “<ALPHANUM>”, “position”: 7 }]}

Slide 9

Slide 9

Semantic search ≠ Literal matches

Slide 10

Slide 10

Elasticsearch You Know, for Search

Slide 11

Slide 11

Elasticsearch You Know, for Vector Search

Slide 12

Slide 12

What is a Vector ?

Slide 13

Slide 13

Example: 1-dimensional vector Character Vector [ 1 ] ] Realistic

[ Embeddings represent your data Cartoon 1

Slide 14

Slide 14

represent different data aspects Human Character Vector [ 1, 1 Realistic Cartoon ] ] Machine

[ Multiple dimensions 1, 0

Slide 15

Slide 15

is grouped together Human Character Vector [ 1.0, 1.0 Realistic Cartoon 1.0, 0.0 [ 1.0, 0.8 ] ] ]

Machine

[ Similar data

Slide 16

Slide 16

Vector search ranks objects by similarity (~relevance) to the query Human Rank Query 1 Realistic Cartoon 2 3 4 5 Machine Result

Slide 17

Slide 17

Choice of Embedding Model Start with Off-the Shelf Models Extend to Higher Relevance ●Text data: Hugging Face (like Microsoft’s E5 ●Apply hybrid scoring ) ●Images: OpenAI’s CLIP ●Bring Your Own Model: requires expertise + labeled data

Slide 18

Slide 18

Problem training vs actual use-case

Slide 19

Slide 19

Architecture of Vector Search

Slide 20

Slide 20

How do you index vectors ?

Slide 21

Slide 21

Data Ingestion and Embedding Generation POST /_doc { “_id”:”product-1234”, “product_name”:”Summer Dress”, “description”:”Our best-selling…”, “Price”: 118, “color”:”blue”, “fabric”:”cotton”, “fabric”:”cotton” } “desc_embedding”:[0.452,0.3242,…], “desc_embedding”:[0.452,0.3242,…] } “img_embedding”:[0.012,0.0,…] } Source data POST /_doc

Slide 22

Slide 22

Co m m er ci With Elastic ML al { } Source data POST /_doc “_id”:”product-1234”, “product_name”:”Summer Dress”, “description”:”Our best-selling…”, “Price”: 118, “color”:”blue”, “fabric”:”cotton”, “desc_embedding”:[0.452,0.3242,…]

Slide 23

Slide 23

Eland Imports PyTorch Models Co m m er ci al $ eland_import_hub_model —url https://cluster_URL —hubmodel-id BERT-MiniLM-L6 —tasktype text_embedding —start BERT-MiniLM-L6 Select the appropriate model Load it Manage models

Slide 24

Slide 24

Elastic’s range of supported NLP models Co m m er ci ● Fill mask model Mask some of the words in a sentence and predict words that replace masks ● Named entity recognition model NLP method that extracts information from text ● Text embedding model Represent individual words as numerical vectors in a predefined vector space ● Text classification model Assign a set of predefined categories to open-ended text ● Question answering model Model that can answer questions given some or no context ● Zero-shot text classification model Model trained on a set of labeled examples, that is able to classify previously unseen examples Full list at: ela.st/nlp-supported-models al

Slide 25

Slide 25

How do you search vectors ?

Slide 26

Slide 26

Vector Query GET product-catalog/_search { “query” : { “bool”: { “must”: [{ “knn”: { “field”: “desc_embbeding”, “num_candidates”: 50, “query_vector”: [0.123, 0.244,…] } }], “filter”: { “term”: { “department”: “women” } } } } }, “size”: 10

Slide 27

Slide 27

Vector Query Transformer model Co m m er ci al GET product-catalog/_search { “query” : { “bool”: { “must”: [{ “knn”: { “field”: “desc_embbeding”, “num_candidates”: 50, “query_vector_builder”: { “text_embedding”: { “model_text”: “summer clothes”, “model_id”: <text-embedding-model> } } } }], “filter”: { “term”: { “department”: “women” } } } }, “size”: 10 }

Slide 28

Slide 28

( Vector Search components Search Index Generate Query Mapping Embedding dense_vector Text embedding model kNN 3rd party, local, in Elasticsearch)

Slide 29

Slide 29

But how does it really work?

Slide 30

Slide 30

Similarity: cosine (cosine) Human q cos(θ) = d1 d2 Realistic θ q⃗ × d ⃗ | q⃗ | × | d |⃗ _score = 1 + cos(θ) 2

Slide 31

Slide 31

Similarity: cosine (cosine) 1+1 _score = =1 2 1+0 _score = = 0.5 2 1−1 _score = =0 2

Slide 32

Slide 32

Similarity: Dot Product (dot_product) q⃗ × d ⃗ = | q⃗ | × cos(θ) × | d |⃗ q d θ | q⃗ | × co s (θ ) 1 + dot_ product(q, d) scorefloat = 2 0.5 + dot product(q, d) _scorebyte = 32768 × dims

Slide 33

Slide 33

Similarity: Euclidean distance (l2_norm) y 2 n i (x ∑ 1 i= − y i) q l2_normq,d = y1 d x1 y2 x2 n ∑ i=1 (xi − yi) 1 _score = 1 + (l2_normq,d )2 x 2

Slide 34

Slide 34

Brute Force

Slide 35

Slide 35

Hierarchical Navigable Small Worlds (HNSW One popular approach HNSW: a layered approach that simplifies access to the nearest neighbor Tiered: from coarse to fine approximation over a few steps Balance: Bartering a little accuracy for a lot of scalability ) Speed: Excellent query latency on large scale indices

Slide 36

Slide 36

Filtering KNN Vector Similarity Automatically choose between brute force and HNSW Brute force (not real numbers; used only to demonstrate the linearity) HNSW fi Bound worst case to 2*(brute force) • Brute force scales O(n) of ltered • HNSW scales ~O(log(n)) of all docs

Slide 37

Slide 37

Elasticsearch + Lucene = fast progress ❤

Slide 38

Slide 38

Scaling Vector Search Vector search Best practices

  1. Needs lots of memory
  2. Avoid searches during indexing
  3. Indexing is slower
  4. Exclude vectors from _source
  5. Merging is slow
  6. Reduce vector dimensionality 4. Use byte rather than float
  • Continuous improvements in Lucene + Elasticsearch

Slide 39

Slide 39

Reduce Required Memory 2. Reduce of number of dimensions per vector

  1. Vector element size reduction (“quantize”)

Slide 40

Slide 40

Benchmarketing

Slide 41

Slide 41

Elasticsearch You Know, for Hybrid Search

Slide 42

Slide 42

Hybrid scoring Term-based score Linear Combination manual boosting Vector similarity score Combine

Slide 43

Slide 43

GET product-catalog/_search { “query” : { “bool” : { “must” : [{ “match”: { “description”: { “query”: “summer clothes”, “boost”: 0.9 } } },{ “knn”: { “field”: “desc_embbeding”, “query_vector”: [0.123, 0.244,…], “num_candidates”: 50, “boost”: 0.1, “filter”: { “term”: { “department”: “women” } } } }], “filter” : { “range” : { “price”: { “lte”: 30 } } } } } } summer clothes pre-filter post-filter

Slide 44

Slide 44

GET product-catalog/_search { “query” : { “bool” : { “must” : [{ “match”: { “description”: { “query”: “summer clothes”, “boost”: 0.9 } } },{ “knn”: { “field”: “image-vector”, “query_vector”: [54, 10, -2], “num_candidates”: 50, “boost”: 0.1 } },{ “knn”: { “field”: “title-vector”, “query_vector”: [1, 20, -52, 23, 10], “num_candidates”: 10, “boost”: 0.5 } }] } } }

Slide 45

Slide 45

ELSER Elastic Learned Sparse EncodER text_expansion Not BM25 or (dense) vector Sparse vector like BM25 Stored as inverted index Co m m er ci al

Slide 46

Slide 46

PUT /_inference/sparse_embedding/my_elser_model { “service”: “elser”, “service_settings”: { “num_allocations”: 1, “num_threads”: 1 }, “task_settings”: {} } PUT /_inference/text_embedding/openai_embeddings { “service”: “openai”, “service_settings”: { “api_key”: “<api_key>” }, “task_settings”: { “model”: “text-embedding-ada-002” } } PUT /_inference/text_embedding/hugging_face_embeddings { “service”: “hugging_face”, “service_settings”: { “api_key”: “<access_token>”, “url”: “<url_endpoint>” } } Co m m er ci al

Slide 47

Slide 47

POST /_inference/sparse_embedding/my_elser_model { “input”: [ “These are not the droids you are looking for.”, } ] “Obi-Wan never told you what happened to your father.” { “sparse_embedding”: [{ “lucas”: 0.50047517, “ship”: 0.29860738, “dragon”: 0.5300422, “quest”: 0.5974301, “dr”: 2.1055143, “space”: 0.49377063, “robot”: 0.40398192, … Co m m er ci al

Slide 48

Slide 48

Co m m er ci Hybrid ranking al Vector similarity score Reciprocal Rank Fusion (RRF blend multiple ranking methods Combine ) Term-based score ELSER score

Slide 49

Slide 49

GET product-catalog/_search { “sub_searches”: [ { “query”: { “match”: {…} } }, { “query”: { “text_expansion”: {…} } } ], “knn”: {…}, “rank”: { “rrf”: { “window_size”: 50, “rank_constant”: 20 } } } Co m m er ci al BM25f + ELSER + Vector Hybrid Ranking

Slide 50

Slide 50

Reciprocal Rank Fusion (RRF D set of docs R set of rankings as permutation on 1..|D| k - typically set to 60 by default Ranking Algorithm 1 r(d) k+r(d) A 1 1 B 0.7 C D Score r(d) k+r(d) 61 C 1,341 1 61 2 62 A 739 2 62 0.5 3 63 F 732 3 63 0.2 4 64 G 192 4 64 0.01

=

= +

E Doc 5 65 H 183 5 65 ) Score

Doc Ranking Algorithm 2 Doc RRF Score A 1/61 1/62 0,0325 C 1/63 1/61 0,0323 B 1/62 0,0161 F 1/63 0,0159 D 1/64 0,0156

Slide 51

Slide 51

https://djdadoo.pilato.fr/

Slide 52

Slide 52

https://github.com/dadoonet/music-search/

Slide 53

Slide 53

ChatGPT Elastic and LLM

Slide 54

Slide 54

Gen AI Search engines

Slide 55

Slide 55

LLM opportunities and limits your question one answer your question GAI / LLM : public internet data

Slide 56

Slide 56

Slide 57

Slide 57

Retrieval Augmented Generation your question the right answer your question + context window GAI / LLM public internet data your business data documents images audio

Slide 58

Slide 58

Atelier : RAG Time Discuter avec vos propres données Vendredi 13H30 - Paris 243 Sylvain Wallez Aline Paponaud Benjamin Dauvissat Vincent Bréhin

Slide 59

Slide 59

Demo Elastic + Azure OpenAI AWS Bedrock Google Vertex AI Mistral AI Cohere

Slide 60

Slide 60

Peut-être remplacée par votre propre source de données

Slide 61

Slide 61

Conclusion

Slide 62

Slide 62

( Vector Database Hybrid Search (text + vector) Vector Database + ML Semantic) Search Engine Store & Search Vector Embeddings Choice & Flexibility of embedding models Filtering & Faceting Create Vector Embeddings Aggregations Autocomplete Search Analytics Trained model outof-the-box Document-level Security Optimized for text, geo, & other data Ingest Tools (web crawler, connectors, Beats, Agent, API framework) On-prem / Cloud / Hybrid

Slide 63

Slide 63

Elasticsearch You Know, for Semantic Search

Slide 64

Slide 64

La recherche À l’ère de l’IA DEVOXX FRANCE 2024 David Pilato | @dadoonet