You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
=== Tutorial: semantic search with a deployed model
3
+
4
+
++++
5
+
<titleabbrev>Semantic search with deployed model</titleabbrev>
6
+
++++
7
+
8
+
[IMPORTANT]
9
+
====
10
+
* For the easiest way to perform semantic search in the {stack}, refer to the <<semantic-search-semantic-text, `semantic_text`>> end-to-end tutorial.
11
+
* This tutorial was written before the <<inference-apis,{infer} endpoint>> and <<semantic-text,`semantic_text` field type>> was introduced.
12
+
Today we have simpler options for performing semantic search.
13
+
====
14
+
15
+
This guide shows you how to implement semantic search with models deployed in {es}: from selecting an NLP model, to writing queries.
16
+
17
+
18
+
[discrete]
19
+
[[deployed-select-nlp-model]]
20
+
==== Select an NLP model
21
+
22
+
{es} offers the usage of a {ml-docs}/ml-nlp-model-ref.html#ml-nlp-model-ref-text-embedding[wide range of NLP models], including both dense and sparse vector models.
23
+
Your choice of the language model is critical for implementing semantic search successfully.
24
+
25
+
While it is possible to bring your own text embedding model, achieving good search results through model tuning is challenging.
26
+
Selecting an appropriate model from our third-party model list is the first step.
27
+
Training the model on your own data is essential to ensure better search results than using only BM25.
28
+
However, the model training process requires a team of data scientists and ML experts, making it expensive and time-consuming.
29
+
30
+
To address this issue, Elastic provides a pre-trained representational model called {ml-docs}/ml-nlp-elser.html[Elastic Learned Sparse EncodeR (ELSER)].
31
+
ELSER, currently available only for English, is an out-of-domain sparse vector model that does not require fine-tuning.
32
+
This adaptability makes it suitable for various NLP use cases out of the box.
33
+
Unless you have a team of ML specialists, it is highly recommended to use the ELSER model.
34
+
35
+
In the case of sparse vector representation, the vectors mostly consist of zero values, with only a small subset containing non-zero values.
36
+
This representation is commonly used for textual data.
37
+
In the case of ELSER, each document in an index and the query text itself are represented by high-dimensional sparse vectors.
38
+
Each non-zero element of the vector corresponds to a term in the model vocabulary.
39
+
The ELSER vocabulary contains around 30000 terms, so the sparse vectors created by ELSER contain about 30000 values, the majority of which are zero.
40
+
Effectively the ELSER model is replacing the terms in the original query with other terms that have been learnt to exist in the documents that best match the original search terms in a training dataset, and weights to control how important each is.
41
+
42
+
43
+
[discrete]
44
+
[[deployed-deploy-nlp-model]]
45
+
==== Deploy the model
46
+
47
+
After you decide which model you want to use for implementing semantic search, you need to deploy the model in {es}.
Depending on the type of model you have deployed, you can query rank features with a <<query-dsl-sparse-vector-query, sparse vector>> query, or dense vectors with a kNN search.
In some situations, lexical search may perform better than semantic search.
92
+
For example, when searching for single words or IDs, like product numbers.
93
+
94
+
Combining semantic and lexical search into one hybrid search request using <<rrf,reciprocal rank fusion>> provides the best of both worlds.
95
+
Not only that, but hybrid search using reciprocal rank fusion {blog-ref}improving-information-retrieval-elastic-stack-hybrid[has been shown to perform better in general].
Copy file name to clipboardExpand all lines: docs/reference/search/search-your-data/semantic-search.asciidoc
+54-70Lines changed: 54 additions & 70 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -7,120 +7,104 @@ Semantic search is a search method that helps you find data based on the intent
7
7
Using an NLP model enables you to extract text embeddings out of text.
8
8
Embeddings are vectors that provide a numeric representation of a text.
9
9
Pieces of content with similar meaning have similar representations.
10
-
NLP models can be used in the {stack} various ways, you can:
11
10
12
-
* deploy models in {es}
13
-
* use the <<semantic-search-semantic-text, `semantic_text` workflow>> (recommended)
14
-
* use the <<semantic-search-inference, {infer} API workflow>>
11
+
You have several options for using NLP models in the {stack}:
15
12
13
+
* use the `semantic_text` workflow (recommended)
14
+
* use the {infer} API workflow
15
+
* deploy models directly in {es}
16
16
17
-
[[semantic-search-diagram]]
18
-
.A simplified representation of encoding textual concepts as vectors
19
-
image::images/search/vector-search-oversimplification.png[A simplified representation of encoding textual concepts as vectors,align="center"]
17
+
Refer to <<using-nlp-models,this section>> to choose your workflow.
20
18
21
-
At query time, {es} can use the same NLP model to convert a query into embeddings, enabling you to find documents with similar text embeddings.
19
+
You can also store your own embeddings in {es} as vectors.
20
+
Refer to <<using-query,this section>> for guidance on which query type to use for semantic search.
22
21
23
-
This guide shows you how to implement semantic search with {es}: From selecting an NLP model, to writing queries.
22
+
At query time, {es} can use the same NLP model to convert a query into embeddings, enabling you to find documents with similar text embeddings.
24
23
25
-
IMPORTANT: For the easiest way to perform semantic search in the {stack}, refer to the <<semantic-search-semantic-text, `semantic_text`>> end-to-end tutorial.
26
24
27
25
[discrete]
28
-
[[semantic-search-select-nlp-model]]
29
-
=== Select an NLP model
30
-
31
-
{es} offers the usage of a
32
-
{ml-docs}/ml-nlp-model-ref.html#ml-nlp-model-ref-text-embedding[wide range of NLP models], including both dense and sparse vector models.
33
-
Your choice of the language model is critical for implementing semantic search successfully.
34
-
35
-
While it is possible to bring your own text embedding model, achieving good search results through model tuning is challenging.
36
-
Selecting an appropriate model from our third-party model list is the first step.
37
-
Training the model on your own data is essential to ensure better search results than using only BM25.
38
-
However, the model training process requires a team of data scientists and ML experts, making it expensive and time-consuming.
39
-
40
-
To address this issue, Elastic provides a pre-trained representational model called {ml-docs}/ml-nlp-elser.html[Elastic Learned Sparse EncodeR (ELSER)].
41
-
ELSER, currently available only for English, is an out-of-domain sparse vector model that does not require fine-tuning.
42
-
This adaptability makes it suitable for various NLP use cases out of the box.
43
-
Unless you have a team of ML specialists, it is highly recommended to use the ELSER model.
44
-
45
-
In the case of sparse vector representation, the vectors mostly consist of zero values, with only a small subset containing non-zero values.
46
-
This representation is commonly used for textual data.
47
-
In the case of ELSER, each document in an index and the query text itself are represented by high-dimensional sparse vectors.
48
-
Each non-zero element of the vector corresponds to a term in the model vocabulary.
49
-
The ELSER vocabulary contains around 30000 terms, so the sparse vectors created by ELSER contain about 30000 values, the majority of which are zero.
50
-
Effectively the ELSER model is replacing the terms in the original query with other terms that have been learnt to exist in the documents that best match the original search terms in a training dataset, and weights to control how important each is.
26
+
[[using-nlp-models]]
27
+
=== Choose a semantic search workflow
51
28
52
29
[discrete]
53
-
[[semantic-search-deploy-nlp-model]]
54
-
=== Deploy the model
30
+
==== `semantic_text` workflow
55
31
56
-
After you decide which model you want to use for implementing semantic search, you need to deploy the model in {es}.
32
+
The simplest way to use NLP models in the {stack} is through the <<semantic-search-semantic-text, `semantic_text` workflow>>.
33
+
We recommend using this approach because it abstracts away a lot of manual work.
34
+
All you need to do is create an {infer} endpoint and an index mapping to start ingesting, embedding, and querying data.
35
+
There is no need to define model-related settings and parameters, or to create {infer} ingest pipelines.
36
+
Refer to the <<put-inference-api, Create an {infer} endpoint API>> documentation for a list of supported services.
The <<semantic-search-semantic-text, Semantic search with `semantic_text`>> tutorial shows you the process end-to-end.
59
39
60
40
[discrete]
61
-
[[semantic-search-field-mappings]]
62
-
=== Map a field for the text embeddings
41
+
==== {infer} API workflow
63
42
64
-
Before you start using the deployed model to generate embeddings based on your input text, you need to prepare your index mapping first.
65
-
The mapping of the index depends on the type of model.
43
+
The <<semantic-search-inference, {infer} API workflow>> is more complex but offers greater control over the {infer} endpoint configuration.
44
+
You need to create an {infer} endpoint, provide various model-related settings and parameters, define an index mapping, and set up an {infer} ingest pipeline with the appropriate settings.
The <<semantic-search-inference, Semantic search with the {infer} API>> tutorial shows you the process end-to-end.
68
47
69
48
[discrete]
70
-
[[semantic-search-generate-embeddings]]
71
-
=== Generate text embeddings
49
+
==== Model deployment workflow
72
50
73
-
Once you have created the mappings for the index, you can generate text embeddings from your input text.
74
-
This can be done by using an
75
-
<<ingest,ingest pipeline>> with an <<inference-processor,inference processor>>.
76
-
The ingest pipeline processes the input data and indexes it into the destination index.
77
-
At index time, the inference ingest processor uses the trained model to infer against the data ingested through the pipeline.
78
-
After you created the ingest pipeline with the inference processor, you can ingest your data through it to generate the model output.
51
+
You can also deploy NLP in {es} manually, without using an {infer} endpoint.
52
+
This is the most complex and labor intensive workflow for performing semantic search in the {stack}.
53
+
You need to select an NLP model from the {ml-docs}/ml-nlp-model-ref.html#ml-nlp-model-ref-text-embedding[list of supported dense and sparse vector models], deploy it using the Eland client, create an index mapping, and set up a suitable ingest pipeline to start ingesting and querying data.
The <<semantic-search-deployed-nlp-model, Semantic search with a model deployed in {es}>> tutorial shows you the process end-to-end.
81
56
82
-
Now it is time to perform semantic search!
83
57
84
58
[discrete]
85
-
[[semantic-search-search]]
86
-
=== Search the data
59
+
[[using-query]]
60
+
=== Using the right query
87
61
88
-
Depending on the type of model you have deployed, you can query rank features with a <<query-dsl-sparse-vector-query, sparse vector>> query, or dense vectors with a kNN search.
62
+
Crafting the right query is crucial for semantic search.
63
+
Which query you use and which field you target in your queries depends on your chosen workflow.
64
+
If you're using the `semantic_text` workflow it's quite simple.
65
+
If not, it depends on which type of embeddings you're working with.
| <<semantic-text,`semantic_text`>> | <<query-dsl-semantic-query,`semantic`>> | The `semantic_text` field handles generating embeddings for you at index time and query time.
71
+
| <<sparse-vector,`sparse_vector`>> | <<query-dsl-sparse-vector-query,`sparse_vector`>> | The `sparse_vector` query can generate query embeddings for you, but you can also provide your own. You must provide embeddings at index time.
72
+
| <<dense-vector,`dense_vector`>> | <<query-dsl-knn-query,`knn`>> | The `knn` query can generate query embeddings for you, but you can also provide your own. You must provide embeddings at index time.
If you want {es} to generate embeddings at both index and query time, use the `semantic_text` field and the `semantic` query.
76
+
If you want to bring your own embeddings, use the `sparse_vector` or `dense_vector` field type and the associated query depending on the NLP model you used to generate the embeddings.
95
77
96
-
In some situations, lexical search may perform better than semantic search.
97
-
For example, when searching for single words or IDs, like product numbers.
98
-
99
-
Combining semantic and lexical search into one hybrid search request using
100
-
<<rrf,reciprocal rank fusion>> provides the best of both worlds.
101
-
Not only that, but hybrid search using reciprocal rank fusion {blog-ref}improving-information-retrieval-elastic-stack-hybrid[has been shown to perform better in general].
78
+
IMPORTANT: For the easiest way to perform semantic search in the {stack}, refer to the <<semantic-search-semantic-text, `semantic_text`>> end-to-end tutorial.
** <<semantic-search-elser,Semantic search with ELSER>>
86
+
** <<semantic-search-semantic-text, Semantic search with `semantic_text`>>
87
+
** <<semantic-search-inference, Semantic search with the {infer} API>>
88
+
** <<semantic-search-elser,Semantic search with ELSER>> using the model deployment workflow
89
+
** <<semantic-search-deployed-nlp-model, Semantic search with a model deployed in {es}>>
111
90
** {ml-docs}/ml-nlp-text-emb-vector-search-example.html[Semantic search with the msmarco-MiniLM-L-12-v3 sentence-transformer model]
91
+
* Interactive examples:
92
+
** The https://github.com/elastic/elasticsearch-labs[`elasticsearch-labs`] repo contains a number of interactive semantic search examples in the form of executable Python notebooks, using the {es} Python client
93
+
** https://github.com/elastic/elasticsearch-labs/blob/main/notebooks/search/03-ELSER.ipynb[Semantic search with ELSER using the model deployment workflow]
94
+
** https://github.com/elastic/elasticsearch-labs/blob/main/notebooks/search/09-semantic-text.ipynb[Semantic search with `semantic_text`]
112
95
* Blogs:
96
+
** https://www.elastic.co/search-labs/blog/semantic-search-simplified-semantic-text[{es} new semantic_text mapping: Simplifying semantic search]
113
97
** {blog-ref}may-2023-launch-sparse-encoder-ai-model[Introducing Elastic Learned Sparse Encoder: Elastic's AI model for semantic search]
114
98
** {blog-ref}lexical-ai-powered-search-elastic-vector-database[How to get the best of lexical and AI-powered search with Elastic's vector database]
115
99
** Information retrieval blog series:
116
100
*** {blog-ref}improving-information-retrieval-elastic-stack-search-relevance[Part 1: Steps to improve search relevance]
** The https://github.com/elastic/elasticsearch-labs[`elasticsearch-labs`] repo contains a number of interactive semantic search examples in the form of executable Python notebooks, using the {es} Python client
0 commit comments