Deep Lake Vector Store API

Running Vector Search in the Deep Lake Vector Store module.

Search Options for Deep Lake Vector Stores in the Deep Lake API

This tutorial requires installation of:

!pip3 install "deeplake[enterprise]" langchain openai tiktoken

Vector Search Using Python Logic (Client-Side)

Let's load the same vector store used in the Quickstart and run embeddings search based on a user prompt using the Deep Lake Vector Store module.

from deeplake.core.vectorstore.deeplake_vectorstore import DeepLakeVectorStore
import openai
import os

os.environ['OPENAI_API_KEY'] = <OPENAI_API_KEY>

vector_store_path = 'hub://activeloop/paul_graham_essay'

vector_store = DeepLakeVectorStore(
    path = vector_store_path,
    read_only = True
)

Next, let's define an embedding function using OpenAI. It must work for a single string and a list of strings, so that it can both be used to embed a prompt and a batch of texts.

def embedding_function(texts, model="text-embedding-ada-002"):
   
   if isinstance(texts, str):
       texts = [texts]

   texts = [t.replace("\n", " ") for t in texts]
   return [data['embedding']for data in openai.Embedding.create(input = texts, model=model)['data']]

Simple Vector Search

Lets run a simple vector search using default options, which performs simple cosine similarity search in Python on the client.

prompt = "What are the first programs he tried writing?"

search_results = vector_store.search(embedding_data=prompt, embedding_function=embedding_function)

The search_results is a dictionary with keys for the text, score, id, and metadata, with data ordered by score. By default, it returns 4 samples ordered by similarity score, and if we examine the first returned text, it appears to contain the text about trust and safety models that is relevant to the prompt.

search_results['text'][0]

Returns:

What I Worked On

February 2021

Before college the two main things I worked on, outside of school, were writing and programming. I didn't write essays. I wrote what beginning writers were supposed to write then, and probably still are: short stories. My stories were awful. They had hardly any plot, just characters with strong feelings, which I imagined made them deep.

The first programs I tried writing were on the IBM 1401 that our school district used for what was then called "data processing." This was in 9th grade, so I was 13 or 14. The school district's 1401 happened to be in the basement of our junior high school, and my friend Rich Draves and I got permission to use it. It was like a mini Bond villain's lair down there, with all these alien-looking machines — CPU, disk drives, printer, card reader — sitting up on a raised floor under bright fluorescent lights.

Advanced Vector Search

Vector search can be combined with other search logic for performing more advanced queries. Let's define a function compatible with deeplake.filter for filtering data prior to the vector search. The function below will filter samples that contain the word "program" in the text tensor.

def filter_fn(x):
    # x is a single row in Deep Lake, 'text' is the tensor name, .data()['value'] is the method for fetching the data
    return "program" in x['text'].data()['value'].lower()

Let's run the vector search with the filter above, return more samples (k = 10), and perform similarity search using L2 metric (distance_metric = "l2"):

prompt = "What are the first programs he tried writing?"

search_results_filter = vector_store.search(embedding_data=prompt, 
                                            embedding_function=embedding_function,
                                            filter=filter_fn,
                                            k=10,
                                            distance_metric='l2')

We can verity that the word "program" is present in all of the results:

all(["program" in result for result in search_results_filter["text"]])

# Returns True

Vector Search Using Compute Engine (Client-Side)

Deep Lake offers advanced search features using Compute Engine, which executes queries with higher performance in C++, and offers querying using Deep Lake's Tensor Query Language (TQL).

In order to use Compute Engine, Deep Lake data must be stored in Deep Lake Storage, or in the user's cloud while being connected to Deep Lake using Managed Credentials.

Let's load a larger Vector Store for running more interesting queries:

vector_store_path = "hub://activeloop/twitter-algorithm"

vector_store = DeepLakeVectorStore(
    path = vector_store_path,
    read_only = True
)

NOTE: this Vector Store is stored in us-east, and query performance may vary significantly depending on your location. In real-world use-cases, users would store their vector stores in regions optimized for their use case.

Simple Vector Search

Lets run a simple vector search and specify exec_option = "compute_engine", which will performs cosine similarity search using Compute Engine on the client.

prompt = "What do the trust and safety models do?"

search_results = vector_store.search(embedding_data = prompt, 
                                     embedding_function = embedding_function,
                                     exec_option = "compute_engine")

If we examine the first returned text, it appears to contain the text about trust and safety models that is relevant to the prompt.

search_results['text'][0]

Returns:

Trust and Safety Models
=======================

We decided to open source the training code of the following models:
- pNSFWMedia: Model to detect tweets with NSFW images. This includes adult and porn content.
- pNSFWText: Model to detect tweets with NSFW text, adult/sexual topics.
- pToxicity: Model to detect toxic tweets. Toxicity includes marginal content like insults and certain types of harassment. Toxic content does not violate Twitter's terms of service.
- pAbuse: Model to detect abusive content. This includes violations of Twitter's terms of service, including hate speech, targeted harassment and abusive behavior.

We have several more models and rules that we are not going to open source at this time because of the adversarial nature of this area. The team is considering open sourcing more models going forward and will keep the community posted accordingly.

Advanced Vector Search

Now let's run a more advanced search that includes filtering of text, metadata, and embedding tensors. We do this using TQL by combining embedding search syntax (cosine_similarity(embedding, ...)) and filtering syntax (where ....).

We are interested in answering a prompt based on the question:

prompt = "What does the python code do?"

Therefore, we apply a filter to only search for text that contains the word "python" and metadata where the source key contains ".py".

embedding = embedding_function(prompt)[0]

# Format the embedding array or list as a string, so it can be passed in the REST API request.
embedding_string = ",".join([str(item) for item in embedding])

tql_query = f"select * from (select text, metadata, cosine_similarity(embedding, ARRAY[{embedding_string}]) as score where contains(text, 'python') or contains(metadata['source'], '.py')) order by score desc limit 5"

search_results = vector_store.search(query = tql_query,
                                     exec_option = "compute_engine")

Vector Search Using the Managed Tensor Database (Server-Side)

Any of the queries above can be executed on the Managed Tensor Database (instead of the client) by specifying exec_option = "tensor_db". Note that these queries are only available for dataset stored in the Managed Tensor Database, which is done by specifying vector_store_path = hub://org_id/dataset_name and runtime = {"tensor_db": True} during Vector Store creation.

search_results = vector_store.search(query = tql_query,
                                     exec_option = "compute_engine")

If Vector Stores are not in the Managed Tensor Database, they can be migrated using these steps: