Bagatur/filter metadata (#9015)

Co-authored-by: Matt Robinson <mrobinson@unstructuredai.io>

libs/langchain/langchain/vectorstores/chroma.py

@@ -205,12 +205,22 @@ class Chroma(VectorStore):
                     [embeddings[idx] for idx in non_empty_ids] if embeddings else None
                 )
                 ids_with_metadata = [ids[idx] for idx in non_empty_ids]
-                self._collection.upsert(
-                    metadatas=metadatas,
-                    embeddings=embeddings_with_metadatas,
-                    documents=texts_with_metadatas,
-                    ids=ids_with_metadata,
-                )
+                try:
+                    self._collection.upsert(
+                        metadatas=metadatas,
+                        embeddings=embeddings_with_metadatas,
+                        documents=texts_with_metadatas,
+                        ids=ids_with_metadata,
+                    )
+                except ValueError as e:
+                    if "Expected metadata value to be" in str(e):
+                        msg = (
+                            "Try filtering complex metadata from the document using "
+                            "langchain.vectorstore.utils.filter_complex_metadata."
+                        )
+                        raise ValueError(e.args[0] + "\n\n" + msg)
+                    else:
+                        raise e
             if empty_ids:
                 texts_without_metadatas = [texts[j] for j in empty_ids]
                 embeddings_without_metadatas = (

libs/langchain/langchain/vectorstores/utils.py

@@ -1,10 +1,11 @@
 """Utility functions for working with vectors and vectorstores."""
 
 from enum import Enum
-from typing import List
+from typing import List, Tuple, Type
 
 import numpy as np
 
+from langchain.docstore.document import Document
 from langchain.utils.math import cosine_similarity
 
 
@@ -51,3 +52,23 @@ def maximal_marginal_relevance(
         idxs.append(idx_to_add)
         selected = np.append(selected, [embedding_list[idx_to_add]], axis=0)
     return idxs
+
+
+def filter_complex_metadata(
+    documents: List[Document],
+    *,
+    allowed_types: Tuple[Type, ...] = (str, bool, int, float)
+) -> List[Document]:
+    """Filter out metadata types that are not supported for a vector store."""
+    updated_documents = []
+    for document in documents:
+        filtered_metadata = {}
+        for key, value in document.metadata.items():
+            if not isinstance(value, allowed_types):
+                continue
+            filtered_metadata[key] = value
+
+        document.metadata = filtered_metadata
+        updated_documents.append(document)
+
+    return updated_documents

libs/langchain/tests/unit_tests/vectorstores/test_utils.py

@@ -1,7 +1,11 @@
 """Test vector store utility functions."""
 import numpy as np
 
-from langchain.vectorstores.utils import maximal_marginal_relevance
+from langchain.docstore.document import Document
+from langchain.vectorstores.utils import (
+    filter_complex_metadata,
+    maximal_marginal_relevance,
+)
 
 
 def test_maximal_marginal_relevance_lambda_zero() -> None:
@@ -52,3 +56,70 @@ def test_maximal_marginal_relevance_query_dim() -> None:
     first = maximal_marginal_relevance(query_embedding, embedding_list)
     second = maximal_marginal_relevance(query_embedding_2d, embedding_list)
     assert first == second
+
+
+def test_filter_list_metadata() -> None:
+    documents = [
+        Document(
+            page_content="",
+            metadata={
+                "key1": "this is a string!",
+                "key2": ["a", "list", "of", "strings"],
+            },
+        ),
+        Document(
+            page_content="",
+            metadata={
+                "key1": "this is another string!",
+                "key2": {"foo"},
+            },
+        ),
+        Document(
+            page_content="",
+            metadata={
+                "key1": "this is another string!",
+                "key2": {"foo": "bar"},
+            },
+        ),
+        Document(
+            page_content="",
+            metadata={
+                "key1": "this is another string!",
+                "key2": True,
+            },
+        ),
+        Document(
+            page_content="",
+            metadata={
+                "key1": "this is another string!",
+                "key2": 1,
+            },
+        ),
+        Document(
+            page_content="",
+            metadata={
+                "key1": "this is another string!",
+                "key2": 1.0,
+            },
+        ),
+        Document(
+            page_content="",
+            metadata={
+                "key1": "this is another string!",
+                "key2": "foo",
+            },
+        ),
+    ]
+
+    updated_documents = filter_complex_metadata(documents)
+    filtered_metadata = [doc.metadata for doc in updated_documents]
+
+    assert filtered_metadata == [
+        {"key1": "this is a string!"},
+        {"key1": "this is another string!"},
+        {"key1": "this is another string!"},
+        {"key1": "this is another string!", "key2": True},
+        {"key1": "this is another string!", "key2": 1},
+        {"key1": "this is another string!", "key2": 1.0},
+        {"key1": "this is another string!", "key2": "foo"},
+    ]