Embedding Functions

Embeddings are high dimensional floating-point vector representations of your data or query. Anything can be embedded using some embedding model or function. For a given embedding function, the output will always have the same number of dimensions.

Creating an embedding function

Any function that takes as input a batch (list) of data and outputs a batch (list) of embeddings can be used by LanceDB as an embedding function. The input and output batch sizes should be the same.

HuggingFace example

One popular free option would be to use the sentence-transformers library from HuggingFace. You can install this using pip: pip install sentence-transformers.

from sentence_transformers import SentenceTransformer

name="paraphrase-albert-small-v2"
model = SentenceTransformer(name)

# used for both training and querying
def embed_func(batch):
    return [model.encode(sentence) for sentence in batch]

OpenAI example

You can also use an external API like OpenAI to generate embeddings

import openai
import os

# Configuring the environment variable OPENAI_API_KEY
if "OPENAI_API_KEY" not in os.environ:
    # OR set the key here as a variable
    openai.api_key = "sk-..."

# verify that the API key is working
assert len(openai.Model.list()["data"]) > 0

def embed_func(c):
    rs = openai.Embedding.create(input=c, engine="text-embedding-ada-002")
    return [record["embedding"] for record in rs["data"]]

Applying an embedding function

Using an embedding function, you can apply it to raw data to generate embeddings for each row.

Say if you have a pandas DataFrame with a text column that you want to be embedded, you can use the with_embeddings function to generate embeddings and add create a combined pyarrow table:

import pandas as pd
from lancedb.embeddings import with_embeddings

df = pd.DataFrame([{"text": "pepperoni"},
                   {"text": "pineapple"}])
data = with_embeddings(embed_func, df)

# The output is used to create / append to a table
# db.create_table("my_table", data=data)

If your data is in a different column, you can specify the column kwarg to with_embeddings.

By default, LanceDB calls the function with batches of 1000 rows. This can be configured using the batch_size parameter to with_embeddings.

LanceDB automatically wraps the function with retry and rate-limit logic to ensure the OpenAI API call is reliable.

Searching with an embedding function

At inference time, you also need the same embedding function to embed your query text. It's important that you use the same model / function otherwise the embedding vectors don't belong in the same latent space and your results will be nonsensical.

query = "What's the best pizza topping?"
query_vector = embed_func([query])[0]
tbl.search(query_vector).limit(10).to_df()

The above snippet returns a pandas DataFrame with the 10 closest vectors to the query.

Roadmap

In the near future, we'll be integrating the embedding functions deeper into LanceDB
. The goal is that you just have to configure the function once when you create the table, and then you'll never have to deal with embeddings / vectors after that unless you want to. We'll also integrate more popular models and APIs.

3.3 KiB Raw Blame History