community[patch]: LanceDB integration improvements/fixes (#16173)

Hi, I'm from the LanceDB team.

Improves LanceDB integration by making it easier to use - now you aren't
required to create tables manually and pass them in the constructor,
although that is still backward compatible.

Bug fix - pandas was being used even though it's not a dependency for
LanceDB or langchain

PS - this issue was raised a few months ago but lost traction. It is a
feature improvement for our users kindly review this , Thanks !

docs/docs/modules/data_connection/vectorstores/index.mdx

@@ -131,7 +131,7 @@ table = db.create_table(
 raw_documents = TextLoader('../../../state_of_the_union.txt').load()
 text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
 documents = text_splitter.split_documents(raw_documents)
-db = LanceDB.from_documents(documents, OpenAIEmbeddings(), connection=table)
+db = LanceDB.from_documents(documents, OpenAIEmbeddings())
 ```
 
   </TabItem>

libs/community/langchain_community/vectorstores/lancedb.py

@@ -12,6 +12,18 @@ class LanceDB(VectorStore):
     """`LanceDB` vector store.
 
     To use, you should have ``lancedb`` python package installed.
+    You can install it with ``pip install lancedb``.
+
+    Args:
+        connection: LanceDB connection to use. If not provided, a new connection
+                    will be created.
+        embedding: Embedding to use for the vectorstore.
+        vector_key: Key to use for the vector in the database. Defaults to ``vector``.
+        id_key: Key to use for the id in the database. Defaults to ``id``.
+        text_key: Key to use for the text in the database. Defaults to ``text``.
+        table_name: Name of the table to use. Defaults to ``vectorstore``.
+
+
 
     Example:
         .. code-block:: python
@@ -25,33 +37,43 @@ class LanceDB(VectorStore):
 
     def __init__(
         self,
-        connection: Any,
-        embedding: Embeddings,
+        connection: Optional[Any] = None,
+        embedding: Optional[Embeddings] = None,
         vector_key: Optional[str] = "vector",
         id_key: Optional[str] = "id",
         text_key: Optional[str] = "text",
+        table_name: Optional[str] = "vectorstore",
     ):
-        """Initialize with Lance DB connection"""
+        """Initialize with Lance DB vectorstore"""
         try:
             import lancedb
         except ImportError:
             raise ImportError(
                 "Could not import lancedb python package. "
                 "Please install it with `pip install lancedb`."
             )
-        if not isinstance(connection, lancedb.db.LanceTable):
-            raise ValueError(
-                "connection should be an instance of lancedb.db.LanceTable, ",
-                f"got {type(connection)}",
-            )
-        self._connection = connection
+        self.lancedb = lancedb
         self._embedding = embedding
         self._vector_key = vector_key
         self._id_key = id_key
         self._text_key = text_key
+        self._table_name = table_name
+
+        if self._embedding is None:
+            raise ValueError("embedding should be provided")
+
+        if connection is not None:
+            if not isinstance(connection, lancedb.db.LanceTable):
+                raise ValueError(
+                    "connection should be an instance of lancedb.db.LanceTable, ",
+                    f"got {type(connection)}",
+                )
+            self._connection = connection
+        else:
+            self._connection = self._init_table()
 
     @property
-    def embeddings(self) -> Embeddings:
+    def embeddings(self) -> Optional[Embeddings]:
         return self._embedding
 
     def add_texts(
@@ -74,7 +96,7 @@ def add_texts(
         # Embed texts and create documents
         docs = []
         ids = ids or [str(uuid.uuid4()) for _ in texts]
-        embeddings = self._embedding.embed_documents(list(texts))
+        embeddings = self._embedding.embed_documents(list(texts))  # type: ignore
         for idx, text in enumerate(texts):
             embedding = embeddings[idx]
             metadata = metadatas[idx] if metadatas else {}
@@ -86,7 +108,6 @@ def add_texts(
                     **metadata,
                 }
             )
-
         self._connection.add(docs)
         return ids
 
@@ -102,14 +123,23 @@ def similarity_search(
         Returns:
             List of documents most similar to the query.
         """
-        embedding = self._embedding.embed_query(query)
-        docs = self._connection.search(embedding).limit(k).to_df()
+        embedding = self._embedding.embed_query(query)  # type: ignore
+        docs = (
+            self._connection.search(embedding, vector_column_name=self._vector_key)
+            .limit(k)
+            .to_arrow()
+        )
+        columns = docs.schema.names
         return [
             Document(
-                page_content=row[self._text_key],
-                metadata=row[docs.columns != self._text_key],
+                page_content=docs[self._text_key][idx].as_py(),
+                metadata={
+                    col: docs[col][idx].as_py()
+                    for col in columns
+                    if col != self._text_key
+                },
             )
-            for _, row in docs.iterrows()
+            for idx in range(len(docs))
         ]
 
     @classmethod
@@ -134,3 +164,23 @@ def from_texts(
         instance.add_texts(texts, metadatas=metadatas, **kwargs)
 
         return instance
+
+    def _init_table(self) -> Any:
+        import pyarrow as pa
+
+        schema = pa.schema(
+            [
+                pa.field(
+                    self._vector_key,
+                    pa.list_(
+                        pa.float32(),
+                        len(self.embeddings.embed_query("test")),  # type: ignore
+                    ),
+                ),
+                pa.field(self._id_key, pa.string()),
+                pa.field(self._text_key, pa.string()),
+            ]
+        )
+        db = self.lancedb.connect("/tmp/lancedb")
+        tbl = db.create_table(self._table_name, schema=schema, mode="overwrite")
+        return tbl

libs/community/tests/integration_tests/vectorstores/test_lancedb.py

@@ -1,8 +1,11 @@
+import pytest
+
 from langchain_community.vectorstores import LanceDB
 from tests.integration_tests.vectorstores.fake_embeddings import FakeEmbeddings
 
 
-def test_lancedb() -> None:
+@pytest.mark.requires("lancedb")
+def test_lancedb_with_connection() -> None:
     import lancedb
 
     embeddings = FakeEmbeddings()
@@ -23,22 +26,23 @@ def test_lancedb() -> None:
     assert "text 1" in result_texts
 
 
-def test_lancedb_add_texts() -> None:
-    import lancedb
+@pytest.mark.requires("lancedb")
+def test_lancedb_without_connection() -> None:
+    embeddings = FakeEmbeddings()
+    texts = ["text 1", "text 2", "item 3"]
+
+    store = LanceDB(embedding=embeddings)
+    store.add_texts(texts)
+    result = store.similarity_search("text 1")
+    result_texts = [doc.page_content for doc in result]
+    assert "text 1" in result_texts
 
+
+@pytest.mark.requires("lancedb")
+def test_lancedb_add_texts() -> None:
     embeddings = FakeEmbeddings()
-    db = lancedb.connect("/tmp/lancedb")
-    texts = ["text 1"]
-    vectors = embeddings.embed_documents(texts)
-    table = db.create_table(
-        "my_table",
-        data=[
-            {"vector": vectors[idx], "id": text, "text": text}
-            for idx, text in enumerate(texts)
-        ],
-        mode="overwrite",
-    )
-    store = LanceDB(table, embeddings)
+
+    store = LanceDB(embedding=embeddings)
     store.add_texts(["text 2"])
     result = store.similarity_search("text 2")
     result_texts = [doc.page_content for doc in result]