Skip to content

The agent engineering platform.

fix(qdrant): allow `as_retriever` to work without embeddings in SPARSE mode (#32757)

Zhou Jing10dfeea110e3
diff --git a/libs/partners/qdrant/langchain_qdrant/qdrant.py b/libs/partners/qdrant/langchain_qdrant/qdrant.py
index 38f8e76060..d6be80dcc5 100644
--- a/libs/partners/qdrant/langchain_qdrant/qdrant.py
+++ b/libs/partners/qdrant/langchain_qdrant/qdrant.py
@@ -237,20 +237,52 @@ class QdrantVectorStore(VectorStore):
return self._client
@property
- def embeddings(self) -> Embeddings:
+ def embeddings(self) -> Optional[Embeddings]:
"""Get the dense embeddings instance that is being used.
- Raises:
- ValueError: If embeddings are ``None``.
+ Returns:
+ Embeddings: An instance of ``Embeddings``, or None for SPARSE mode.
+
+ """
+ return self._embeddings
+
+ def _get_retriever_tags(self) -> list[str]:
+ """Get tags for retriever.
+
+ Override the base class method to handle SPARSE mode where embeddings can be
+ None. In SPARSE mode, embeddings is None, so we don't include embeddings class
+ name in tags. In DENSE/HYBRID modes, embeddings is not None, so we include
+ embeddings class name.
+ """
+ tags = [self.__class__.__name__]
+
+ # Handle different retrieval modes
+ if self.retrieval_mode == RetrievalMode.SPARSE:
+ # SPARSE mode: no dense embeddings, so no embeddings class name in tags
+ pass
+ else:
+ # DENSE/HYBRID modes: include embeddings class name if available
+ if self.embeddings is not None:
+ tags.append(self.embeddings.__class__.__name__)
+
+ return tags
+
+ def _require_embeddings(self, operation: str) -> Embeddings:
+ """Require embeddings for operations that need them.
+
+ Args:
+ operation: Description of the operation requiring embeddings.
Returns:
- Embeddings: An instance of ``Embeddings``.
+ The embeddings instance.
+ Raises:
+ ValueError: If embeddings are None and required for the operation.
"""
- if self._embeddings is None:
- msg = "Embeddings are `None`. Please set using the `embedding` parameter."
+ if self.embeddings is None:
+ msg = f"Embeddings are required for {operation}"
raise ValueError(msg)
- return self._embeddings
+ return self.embeddings
@property
def sparse_embeddings(self) -> SparseEmbeddings:
@@ -517,7 +549,8 @@ class QdrantVectorStore(VectorStore):
**kwargs,
}
if self.retrieval_mode == RetrievalMode.DENSE:
- query_dense_embedding = self.embeddings.embed_query(query)
+ embeddings = self._require_embeddings("DENSE mode")
+ query_dense_embedding = embeddings.embed_query(query)
results = self.client.query_points(
query=query_dense_embedding,
using=self.vector_name,
@@ -536,7 +569,8 @@ class QdrantVectorStore(VectorStore):
).points
elif self.retrieval_mode == RetrievalMode.HYBRID:
- query_dense_embedding = self.embeddings.embed_query(query)
+ embeddings = self._require_embeddings("HYBRID mode")
+ query_dense_embedding = embeddings.embed_query(query)
query_sparse_embedding = self.sparse_embeddings.embed_query(query)
results = self.client.query_points(
prefetch=[
@@ -690,7 +724,8 @@ class QdrantVectorStore(VectorStore):
self.embeddings,
)
- query_embedding = self.embeddings.embed_query(query)
+ embeddings = self._require_embeddings("max_marginal_relevance_search")
+ query_embedding = embeddings.embed_query(query)
return self.max_marginal_relevance_search_by_vector(
query_embedding,
k=k,
@@ -1041,7 +1076,8 @@ class QdrantVectorStore(VectorStore):
texts: Iterable[str],
) -> list[models.VectorStruct]:
if self.retrieval_mode == RetrievalMode.DENSE:
- batch_embeddings = self.embeddings.embed_documents(list(texts))
+ embeddings = self._require_embeddings("DENSE mode")
+ batch_embeddings = embeddings.embed_documents(list(texts))
return [
{
self.vector_name: vector,
@@ -1063,7 +1099,8 @@ class QdrantVectorStore(VectorStore):
]
if self.retrieval_mode == RetrievalMode.HYBRID:
- dense_embeddings = self.embeddings.embed_documents(list(texts))
+ embeddings = self._require_embeddings("HYBRID mode")
+ dense_embeddings = embeddings.embed_documents(list(texts))
sparse_embeddings = self.sparse_embeddings.embed_documents(list(texts))
if len(dense_embeddings) != len(sparse_embeddings):
diff --git a/libs/partners/qdrant/tests/integration_tests/qdrant_vector_store/test_search.py b/libs/partners/qdrant/tests/integration_tests/qdrant_vector_store/test_search.py
index d95a923c72..d612c48dc9 100644
--- a/libs/partners/qdrant/tests/integration_tests/qdrant_vector_store/test_search.py
+++ b/libs/partners/qdrant/tests/integration_tests/qdrant_vector_store/test_search.py
@@ -309,3 +309,105 @@ def test_similarity_search_filters_with_qdrant_filters(
)
],
)
+
+
[email protected]("location", qdrant_locations())
+def test_embeddings_property_sparse_mode(location: str) -> None:
+ """Test that embeddings property returns None in SPARSE mode."""
+ # Use from_texts to create the vectorstore, which handles collection creation
+ texts = ["test document"]
+ vectorstore = QdrantVectorStore.from_texts(
+ texts,
+ embedding=None, # No dense embedding for SPARSE mode
+ location=location,
+ retrieval_mode=RetrievalMode.SPARSE,
+ sparse_embedding=ConsistentFakeSparseEmbeddings(),
+ sparse_vector_name="sparse",
+ )
+
+ # In SPARSE mode, embeddings should return None
+ assert vectorstore.embeddings is None
+
+
[email protected]("location", qdrant_locations())
+def test_embeddings_property_dense_mode(location: str) -> None:
+ """Test that embeddings property returns embedding object in DENSE mode."""
+ # Use from_texts to create the vectorstore, which handles collection creation
+ texts = ["test document"]
+ embedding = ConsistentFakeEmbeddings()
+ vectorstore = QdrantVectorStore.from_texts(
+ texts,
+ embedding=embedding,
+ location=location,
+ retrieval_mode=RetrievalMode.DENSE,
+ )
+
+ # In DENSE mode, embeddings should return the embedding object
+ assert vectorstore.embeddings is embedding
+
+
[email protected]("location", qdrant_locations())
+def test_as_retriever_sparse_mode(location: str) -> None:
+ """Test that as_retriever() works in SPARSE mode."""
+ # Use from_texts to create the vectorstore, which handles collection creation
+ texts = ["test document"]
+ vectorstore = QdrantVectorStore.from_texts(
+ texts,
+ embedding=None, # No dense embedding for SPARSE mode
+ location=location,
+ retrieval_mode=RetrievalMode.SPARSE,
+ sparse_embedding=ConsistentFakeSparseEmbeddings(),
+ sparse_vector_name="sparse",
+ )
+
+ # Add test documents
+ docs = [
+ Document(page_content="Python programming", metadata={"topic": "programming"}),
+ Document(page_content="Machine learning", metadata={"topic": "AI"}),
+ Document(page_content="Data analysis", metadata={"topic": "data"}),
+ ]
+ vectorstore.add_documents(docs)
+
+ # Test basic as_retriever() functionality
+ retriever = vectorstore.as_retriever()
+ results = retriever.invoke("programming")
+
+ # Should return documents
+ assert len(results) > 0
+ assert all(isinstance(doc, Document) for doc in results)
+
+ # Test that retriever has tags
+ assert hasattr(retriever, "tags")
+ assert isinstance(retriever.tags, list)
+ assert "QdrantVectorStore" in retriever.tags
+
+
[email protected]("location", qdrant_locations())
+def test_as_retriever_sparse_mode_with_search_kwargs(location: str) -> None:
+ """Test as_retriever() with custom search_kwargs in SPARSE mode."""
+ # Use from_texts to create the vectorstore, which handles collection creation
+ texts = ["test document"]
+ vectorstore = QdrantVectorStore.from_texts(
+ texts,
+ embedding=None, # No dense embedding for SPARSE mode
+ location=location,
+ retrieval_mode=RetrievalMode.SPARSE,
+ sparse_embedding=ConsistentFakeSparseEmbeddings(),
+ sparse_vector_name="sparse",
+ )
+
+ # Add test documents
+ docs = [
+ Document(page_content="Python programming", metadata={"topic": "programming"}),
+ Document(page_content="Machine learning", metadata={"topic": "AI"}),
+ Document(page_content="Data analysis", metadata={"topic": "data"}),
+ ]
+ vectorstore.add_documents(docs)
+
+ # Test with custom search_kwargs
+ retriever = vectorstore.as_retriever(search_kwargs={"k": 1})
+ results = retriever.invoke("programming")
+
+ # Should return exactly 1 document
+ assert len(results) == 1
+ assert isinstance(results[0], Document)