The easiest way to get started with LanceDB is the open source version, which is an embedded database that
runs in-process (like SQLite). Let’s get started in just a few steps!
1. Install LanceDB
Install LanceDB in your client SDK.
pip install lancedb # or uv add lancedb
2. Connect to a LanceDB database
Using LanceDB’s open source version is as simple as importing LanceDB as a library
and pointing to a local path — no servers needed!
Optional: LanceDB Cloud or Enterprise versions
If you’re looking for a fully-managed solution,
you can use LanceDB Cloud, which provides managed infrastructure,
security, and automatic backups. Simply replace the local path with a remote uri
that points to where your data is stored, and you’re ready to go.
For enormous scale and more advanced use cases beyond just search — like
feature engineering, model training and more, check out LanceDB Enterprise.
3. Create table and ingest data
LanceDB uses Lance tables under the hood. In a Lance table, each row represents a record, and each column represents a field
and/or its metadata. The simplest way to create a table and ingest data to it, is to define a list of objects. Each object
in this list contains a vector field (list of floats) and optional fields for metadata.
An example is shown below. We obtain records of characters in an adventure board game. The vector field is
a list of floats, representing an embedding of the text for a given character. For brevity, only three dimensions are
shown, but in n the real world, the embeddings would be much larger (thousands of dimensions) and be generated via an
embedding model.
The code above creates a table named adventurers and ingests four records into it, by overwriting
the table if it already exists.
The table schema is inferred from the data you provide, if not provided explicitly via the schema parameter.
If you do not want to overwrite the existing table, you can
omit the mode="overwrite" parameter — this will raise a ValueError when trying to create a table with the same name.
LanceDB requires that you provide either data or a schema (via PyArrow) during table creation.
You can learn more on how to provide schemas explicitly in the “working with tables” page.
4. Vector search
Now, let’s perform a vector similarity search. The query vector should have the same dimensionality as
your data vectors. The search returns the most similar vectors based on Euclidean distance.
Our query is a vector that represents a warrior, which isn’t in the data we ingested.
Let’s try and find the result that’s most similar to it!
The example above outputs the top 2 most similar results of the vector search query
as a Polars DataFrame, as it’s easier to interpret the results when displayed in tabular format.
| id | text | vector | _distance |
|---|
| 1 | knight | [0.9, 0.4, 0.8] | 0.02 |
| 2 | ranger | [0.8, 0.4, 0.7] | 0.02 |
knight is the most similar adventurer to the warrior from our query!
If you prefer Pandas, you can use the to_pandas() method to display the results as a Pandas DataFrame, see below.
5. Add data and run more queries
As you obtain more data, you can append it to an existing table. In the same script or a new one, you
can connect to the LanceDB database and open an existing table as follows.
Append more records to the existing table using the add() method:
To verify that our new data was added, we can run a different query that looks for adventurers similar to wizard.
| id | text | vector | _distance |
|---|
| 7 | mage | [0.6, 0.3, 0.4] | 0.02 |
| 9 | priest | [0.6, 0.2, 0.6] | 0.03 |
mage is the most magical of all our characters!
6. What’s next?
You’ve learned how to create a LanceDB database, add data to it, and run vector search to find the most
similar results to your given query embedding. In the real world, embeddings capture meaning and vector
search allows you to find the most relevant data based on semantic similarity.
As you explore LanceDB further, you can combine vector search with other techniques like filtering based
on metadata fields, full-text search, hybrid search, and more.
Note that LanceDB is much more than “just a vector database” — it’s
a multimodal lakehouse.
There’s a lot more you can do with it! Learn more about what LanceDB is
by continuing on to the next page, and have fun building with LanceDB! 
