Elasticsearch

Elasticsearch is a distributed, RESTful search and analytics engine. It provides a distributed, multitenant-capable full-text search engine with an HTTP web interface and schema-free JSON documents. It supports keyword search, vector search, hybrid search and complex filtering.

The ElasticsearchRetriever is a generic wrapper to enable flexible access to all Elasticsearch features through the Query DSL. For most use cases the other classes (ElasticsearchStore, ElasticsearchEmbeddings, etc.) should suffice, but if they don’t you can use ElasticsearchRetriever.

%pip install --upgrade --quiet elasticsearch langchain-elasticsearch

from typing import Any, Dict, Iterable

from elasticsearch import Elasticsearch
from elasticsearch.helpers import bulk
from langchain.embeddings import DeterministicFakeEmbedding
from langchain_core.documents import Document
from langchain_core.embeddings import Embeddings
from langchain_elasticsearch import ElasticsearchRetriever

API Reference:

• DeterministicFakeEmbedding
• Document
• Embeddings
• ElasticsearchRetriever

Configure

Here we define the conncection to Elasticsearch. In this example we use a locally running instance. Alternatively, you can make an account in Elastic Cloud and start a free trial.

es_url = "http://localhost:9200"
es_client = Elasticsearch(hosts=[es_url])
es_client.info()

For vector search, we are going to use random embeddings just for illustration. For real use cases, pick one of the available LangChain Embeddings classes.

embeddings = DeterministicFakeEmbedding(size=3)

Define example data

index_name = "test-langchain-retriever"
text_field = "text"
dense_vector_field = "fake_embedding"
num_characters_field = "num_characters"
texts = [
    "foo",
    "bar",
    "world",
    "hello world",
    "hello",
    "foo bar",
    "bla bla foo",
]

Index data

Typically, users make use of ElasticsearchRetriever when they already have data in an Elasticsearch index. Here we index some example text documents. If you created an index for example using ElasticsearchStore.from_documents that’s also fine.

def create_index(
    es_client: Elasticsearch,
    index_name: str,
    text_field: str,
    dense_vector_field: str,
    num_characters_field: str,
):
    es_client.indices.create(
        index=index_name,
        mappings={
            "properties": {
                text_field: {"type": "text"},
                dense_vector_field: {"type": "dense_vector"},
                num_characters_field: {"type": "integer"},
            }
        },
    )


def index_data(
    es_client: Elasticsearch,
    index_name: str,
    text_field: str,
    dense_vector_field: str,
    embeddings: Embeddings,
    texts: Iterable[str],
    refresh: bool = True,
) -> None:
    create_index(
        es_client, index_name, text_field, dense_vector_field, num_characters_field
    )

    vectors = embeddings.embed_documents(list(texts))
    requests = [
        {
            "_op_type": "index",
            "_index": index_name,
            "_id": i,
            text_field: text,
            dense_vector_field: vector,
            num_characters_field: len(text),
        }
        for i, (text, vector) in enumerate(zip(texts, vectors))
    ]

    bulk(es_client, requests)

    if refresh:
        es_client.indices.refresh(index=index_name)

    return len(requests)

index_data(es_client, index_name, text_field, dense_vector_field, embeddings, texts)

7

Usage examples

Vector search

Dense vector retrival using fake embeddings in this example.

def vector_query(search_query: str) -> Dict:
    vector = embeddings.embed_query(search_query)  # same embeddings as for indexing
    return {
        "knn": {
            "field": dense_vector_field,
            "query_vector": vector,
            "k": 5,
            "num_candidates": 10,
        }
    }


vector_retriever = ElasticsearchRetriever.from_es_params(
    index_name=index_name,
    body_func=vector_query,
    content_field=text_field,
    url=es_url,
)

vector_retriever.invoke("foo")

[Document(page_content='foo', metadata={'_index': 'test-langchain-index', '_id': '0', '_score': 1.0, '_source': {'fake_embedding': [-2.336764233933763, 0.27510289545940503, -0.7957597268194339], 'num_characters': 3}}),
 Document(page_content='world', metadata={'_index': 'test-langchain-index', '_id': '2', '_score': 0.6770179, '_source': {'fake_embedding': [-0.7041151202179595, -1.4652961969276497, -0.25786766898672847], 'num_characters': 5}}),
 Document(page_content='hello world', metadata={'_index': 'test-langchain-index', '_id': '3', '_score': 0.4816144, '_source': {'fake_embedding': [0.42728413221815387, -1.1889908285425348, -1.445433230084671], 'num_characters': 11}}),
 Document(page_content='hello', metadata={'_index': 'test-langchain-index', '_id': '4', '_score': 0.46853775, '_source': {'fake_embedding': [-0.28560441330564046, 0.9958894823084921, 1.5489829880195058], 'num_characters': 5}}),
 Document(page_content='foo bar', metadata={'_index': 'test-langchain-index', '_id': '5', '_score': 0.2086992, '_source': {'fake_embedding': [0.2533670476638539, 0.08100381646160418, 0.7763644080870179], 'num_characters': 7}})]

BM25

Traditional keyword matching.

def bm25_query(search_query: str) -> Dict:
    return {
        "query": {
            "match": {
                text_field: search_query,
            },
        },
    }


bm25_retriever = ElasticsearchRetriever.from_es_params(
    index_name=index_name,
    body_func=bm25_query,
    content_field=text_field,
    url=es_url,
)

bm25_retriever.invoke("foo")

[Document(page_content='foo', metadata={'_index': 'test-langchain-index', '_id': '0', '_score': 0.9711467, '_source': {'fake_embedding': [-2.336764233933763, 0.27510289545940503, -0.7957597268194339], 'num_characters': 3}}),
 Document(page_content='foo bar', metadata={'_index': 'test-langchain-index', '_id': '5', '_score': 0.7437035, '_source': {'fake_embedding': [0.2533670476638539, 0.08100381646160418, 0.7763644080870179], 'num_characters': 7}}),
 Document(page_content='bla bla foo', metadata={'_index': 'test-langchain-index', '_id': '6', '_score': 0.6025789, '_source': {'fake_embedding': [1.7365927060137358, -0.5230400847844948, 0.7978339724186192], 'num_characters': 11}})]

Hybrid search

The combination of vector search and BM25 search using Reciprocal Rank Fusion (RRF) to combine the result sets.

def hybrid_query(search_query: str) -> Dict:
    vector = embeddings.embed_query(search_query)  # same embeddings as for indexing
    return {
        "query": {
            "match": {
                text_field: search_query,
            },
        },
        "knn": {
            "field": dense_vector_field,
            "query_vector": vector,
            "k": 5,
            "num_candidates": 10,
        },
        "rank": {"rrf": {}},
    }


hybrid_retriever = ElasticsearchRetriever.from_es_params(
    index_name=index_name,
    body_func=hybrid_query,
    content_field=text_field,
    url=es_url,
)

hybrid_retriever.invoke("foo")

[Document(page_content='foo', metadata={'_index': 'test-langchain-index', '_id': '0', '_score': 0.9711467, '_source': {'fake_embedding': [-2.336764233933763, 0.27510289545940503, -0.7957597268194339], 'num_characters': 3}}),
 Document(page_content='foo bar', metadata={'_index': 'test-langchain-index', '_id': '5', '_score': 0.7437035, '_source': {'fake_embedding': [0.2533670476638539, 0.08100381646160418, 0.7763644080870179], 'num_characters': 7}}),
 Document(page_content='bla bla foo', metadata={'_index': 'test-langchain-index', '_id': '6', '_score': 0.6025789, '_source': {'fake_embedding': [1.7365927060137358, -0.5230400847844948, 0.7978339724186192], 'num_characters': 11}})]

Fuzzy matching

Keyword matching with typo tolerance.

def fuzzy_query(search_query: str) -> Dict:
    return {
        "query": {
            "match": {
                text_field: {
                    "query": search_query,
                    "fuzziness": "AUTO",
                }
            },
        },
    }


fuzzy_retriever = ElasticsearchRetriever.from_es_params(
    index_name=index_name,
    body_func=fuzzy_query,
    content_field=text_field,
    url=es_url,
)

fuzzy_retriever.invoke("fox")  # note the character tolernace

[Document(page_content='foo', metadata={'_index': 'test-langchain-index', '_id': '0', '_score': 0.6474311, '_source': {'fake_embedding': [-2.336764233933763, 0.27510289545940503, -0.7957597268194339], 'num_characters': 3}}),
 Document(page_content='foo bar', metadata={'_index': 'test-langchain-index', '_id': '5', '_score': 0.49580228, '_source': {'fake_embedding': [0.2533670476638539, 0.08100381646160418, 0.7763644080870179], 'num_characters': 7}}),
 Document(page_content='bla bla foo', metadata={'_index': 'test-langchain-index', '_id': '6', '_score': 0.40171927, '_source': {'fake_embedding': [1.7365927060137358, -0.5230400847844948, 0.7978339724186192], 'num_characters': 11}})]

Complex filtering

Combination of filters on different fields.

def filter_query_func(search_query: str) -> Dict:
    return {
        "query": {
            "bool": {
                "must": [
                    {"range": {num_characters_field: {"gte": 5}}},
                ],
                "must_not": [
                    {"prefix": {text_field: "bla"}},
                ],
                "should": [
                    {"match": {text_field: search_query}},
                ],
            }
        }
    }


filtering_retriever = ElasticsearchRetriever.from_es_params(
    index_name=index_name,
    body_func=filter_query_func,
    content_field=text_field,
    url=es_url,
)

filtering_retriever.invoke("foo")

[Document(page_content='foo bar', metadata={'_index': 'test-langchain-index', '_id': '5', '_score': 1.7437035, '_source': {'fake_embedding': [0.2533670476638539, 0.08100381646160418, 0.7763644080870179], 'num_characters': 7}}),
 Document(page_content='world', metadata={'_index': 'test-langchain-index', '_id': '2', '_score': 1.0, '_source': {'fake_embedding': [-0.7041151202179595, -1.4652961969276497, -0.25786766898672847], 'num_characters': 5}}),
 Document(page_content='hello world', metadata={'_index': 'test-langchain-index', '_id': '3', '_score': 1.0, '_source': {'fake_embedding': [0.42728413221815387, -1.1889908285425348, -1.445433230084671], 'num_characters': 11}}),
 Document(page_content='hello', metadata={'_index': 'test-langchain-index', '_id': '4', '_score': 1.0, '_source': {'fake_embedding': [-0.28560441330564046, 0.9958894823084921, 1.5489829880195058], 'num_characters': 5}})]

Note that the query match is on top. The other documents that got passed the filter are also in the result set, but they all have the same score.

Custom document mapper

It is possible to cusomize the function tha maps an Elasticsearch result (hit) to a LangChain document.

def num_characters_mapper(hit: Dict[str, Any]) -> Document:
    num_chars = hit["_source"][num_characters_field]
    content = hit["_source"][text_field]
    return Document(
        page_content=f"This document has {num_chars} characters",
        metadata={"text_content": content},
    )


custom_mapped_retriever = ElasticsearchRetriever.from_es_params(
    index_name=index_name,
    body_func=filter_query_func,
    document_mapper=num_characters_mapper,
    url=es_url,
)

custom_mapped_retriever.invoke("foo")

[Document(page_content='This document has 7 characters', metadata={'text_content': 'foo bar'}),
 Document(page_content='This document has 5 characters', metadata={'text_content': 'world'}),
 Document(page_content='This document has 11 characters', metadata={'text_content': 'hello world'}),
 Document(page_content='This document has 5 characters', metadata={'text_content': 'hello'})]