Milvus is an open-source vector database that is highly reliable for similarity search. Let's look at how we can use it.
Let's start with some base definitions
Step 1: Create a collection
In Milvus, a collection is similar to a table in a traditional relational database. It's the primary organizational unit for storing data. A collection is formed of one or more (data fields) that can contain vectors.
Creating a collection is necessary because it forms the foundation for storing data in Milvus.
Step 2: create vectors and save them in the collection
Embedding vectors are essentially lists of numbers used to represent various types of data, like images, words, or even collections of text. We create embedding vectors when we want to convert our text data into a form that a machine-learning model can understand and process.
After creating these vectors, we then load them into a vector database, which is a type of database that uses vectors (instead of text) as its primary data representation format. This is useful as it allows for efficient storage and retrieval of high-dimensional vector data.
Step 2: search and query your data from your vector db
Finding and searching data in Milvus involves using the search function on our collection.
What we search for depends on the specific use case - for example, if we're working with a book collection, we might want to search for a word inside book intros that are saved in Milvus as vectors.
After finding the data, we can proceed to query the DB to find the books that contain a specific word.
Querying involves extracting specific data from our search results according to certain conditions.
Here's a flow diagram to visualize this process better:
Let's start using python
Creating a collection in Milvus
A collection is like a table in a relational database and is a basic data organization unit in Milvus. A collection contains a group of vectors and attributes that are logically grouped. Let's start by creating one:
- Imports necessary
import configparser
import time
import random
from pymilvus import connections, utility
from pymilvus import Collection, DataType, FieldSchema, CollectionSchema
- Retrieving Bert model directory and the collection name
bert_model_dir='/pyapp/models/multi-qa-MiniLM-L6-cos-v1/'
collection_name = "book"
- Establishing connection with env variables
cfp = configparser.RawConfigParser()
cfp.read('config.ini')
milvus_uri = cfp.get('milvus_env', 'uri')
token = cfp.get('milvus_env', 'token')
connections.connect("default",
uri=milvus_uri,
token=token)
- Creating if the collection exists. if yes drop it. Then recreate it defining the collection columns ( fields ).
# Check if the collection exists
check_collection = utility.has_collection(collection_name)
if check_collection:
drop_result = utility.drop_collection(collection_name)
print("Success!")
# create a collection with customized primary field: book_id_field
dim = 384
book_id_field = FieldSchema(name="book_id", dtype=DataType.INT64, is_primary=True, description="customized primary id")
word_count_field = FieldSchema(name="word_count", dtype=DataType.INT64, description="word count")
book_intro_field = FieldSchema(name="book_intro", dtype=DataType.FLOAT_VECTOR, dim=dim)
sentence_field = FieldSchema(name="sentence", dtype=DataType.VARCHAR,description="book sentence", max_length=2000,default_value="")
schema = CollectionSchema(fields=[book_id_field, word_count_field, book_intro_field, sentence_field],
auto_id=False,
description="my first collection")
collection = Collection(name=collection_name, schema=schema)
Creating Embedding Vectors with BERT
BERT, which stands for Bidirectional Encoder Representations from Transformers, is a neural network-based technique for natural language processing. As we have text data from books, let's create embeddings using BERT:
from sentence_transformers import SentenceTransformer
- BERT model instantiation
model = SentenceTransformer(bert_model_dir)
- Sentences to be converted into embeddings
sentences = [
"this is a intro about a book with name mela",
"this is a intro about a book with name arachide",
"this is a intro about a book with name banana",
"this is a intro about a book with name mango",
"this is a intro about a book with name pera"
]
- Converting the sentences into embeddings
model = SentenceTransformer(bert_model_dir)
sentence_embeddings = model.encode(sentences)
We now have our embeddings! The next step is inserting these into our Milvus collection.
Inserting Embeddings into the Collection
Let's now insert the embeddings that we created using BERT into our Milvus collection:
- Inserting the data into our collection
nb = 5
sentences = [
"this is a intro about a book with name mela",
"this is a intro about a book with name arachide",
"this is a intro about a book with name banana",
"this is a intro about a book with name mango",
"this is a intro about a book with name pera"
]
model = SentenceTransformer(bert_model_dir)
sentence_embeddings = model.encode(sentences)
book_ids = list(range(nb))
book_ids = [i for i in range(nb)]
word_counts = [random.randint(1, 100) for i in range(nb)]
book_intros = [embedding.tolist() for embedding in sentence_embeddings]
entities = [book_ids, word_counts, book_intros,sentences]
Now that all our embeddings are safely inside our Milvus collection, let's create an index for them.
- then flush the collection
collection.flush()
Indexing the Collection
A collection index enables efficient search operations. Now that we've inserted the vectors, let's index them:
- Creating index
index_params = {"index_type": "AUTOINDEX", "metric_type": "L2", "params": {}}
- Self-constructed index
collection.create_index(field_name=book_intro_field.name, index_params=index_params)
Well done, we just indexed our collection. Now, let's conduct a search within our indexed collection.
Searching the Collection
Let's try to perform a search operation on our collection:
- Searching
search_str = "pera"
nq = 1
search_params = {"metric_type": "L2", "params": {"level": 2}}
topk = 1
search_vec = [model.encode([search_str])[0].tolist()]
results = collection.search(search_vec,
anns_field=book_intro_field.name,
param=search_params,
limit=topk,
)
print(f"Result:{results}")
Great! We found the results from our search operation. Let's end this tutorial with a bang by performing a data query on our collection.
Search and then querying data
Now that we have our collection and its index, we can use the query method to search for specific vectors:
- Search for a specific word contained inside the book intros that we saved inside Milvus db
search_str = "pera"
nq = 1
search_params = {"metric_type": "L2", "params": {"level": 2}}
topk = 1
fruit_embedding = model.encode([search_str])[0].tolist() # encode the text "mela"
search_vec = [fruit_embedding] # use this embedding as the search vector
results = collection.search(search_vec,
anns_field=book_intro_field.name,
param=search_params,
limit=topk,
guarantee_timestamp=1)
- Found the list of id that contain the word, we query the DB to get back the original text
matched_book_id = 0
for hit in results[0]:
matched_book_id = hit.id
matched_sentence = id_sentence_map[matched_book_id]
print(f"Matched Sentence: {matched_sentence}")
given_id = matched_book_id
res = collection.query(
expr = f"book_id == {given_id}",
output_fields = ["book_id", "word_count","sentence"]
)
# Check the query result
for item in res:
print(f"book_id: {item['book_id']}, word_count: {item['word_count']} , sentence: {item['sentence']}")
And that's it! We've successfully gone through all the steps involved in using the Milvus vector database for a book collection, from creating a collection, creating embeddings with BERT, inserting these into our collection, creating an index, performing a search, and finally querying the data.
Let's use the GitHub project
You should have docker installed in your environment
git clone https://github.com/morandalex/vector-search-example.git
make up This will build the python dependencies. You will need some minutes to build, so be patient. Run this command only once.
make test This command will test the test.py script
You will have the following output
make test
docker compose run -w /pyapp python python test.py --remove-orphans
[+] Building 0.0s (0/0) docker:desktop-linux
[+] Building 0.0s (0/0) docker:desktop-linux
Connecting to DB: https://.......api.gcp-us-west1.zillizcloud.com
Success!
Creating ristobot collection: book
Schema: {'auto_id': False, 'description': 'my first book collection', 'fields': [{'name': 'book_id', 'description': 'customized primary id', 'type': <DataType.INT64: 5>, 'is_primary': True, 'auto_id': False}, {'name': 'word_count', 'description': 'word count', 'type': <DataType.INT64: 5>}, {'name': 'book_intro', 'description': '', 'type': <DataType.FLOAT_VECTOR: 101>, 'params': {'dim': 384}}, {'name': 'sentence', 'description': 'book sentence', 'type': <DataType.VARCHAR: 21>, 'params': {'max_length': 2000}}]}
Success!
Inserting 10 entities...
Generate sentence embeddings using the SentenceTransformer model
Prepare data
Succeed in 0.321 seconds!
Flushing...
Succeed in 3.6852 seconds!
Building AutoIndex...
Succeed in 1.6364 seconds!
Loading collection...
Succeed in 4.3029 seconds!
Searching for 'pera'...
Result:["['id: 1, distance: 0.7175177335739136, entity: {}']"]
search latency: 0.2129 seconds!
hit id: id: 1, distance: 0.7175177335739136, entity: {}
hit id: 1
hit distance: 0.7175177335739136
Matched Sentence: this is a intro about a book with name pera
book_id: 1, word_count: 8 , sentence: this is a intro about a book with name pera
The logs show that we created a collection with 4 columns with id, words_count, book_intro, sentence
id is the primary key words_count is a fake field used to test a random inserted number book_intro is the sentence transformed into a vector sentence is the sentence. for this test, we used three sentences that contain 5 different fruits: mela, pera, banana, mango, caco
When the collection of data is loaded then it is performed the indexing process.
After that, we load the collection and make a search request embedding a word.
In this case "pera" word is searched from Milvus and it gives back a vector with an id.
We use the given id to query the DB and get back the sentence as varchar.