Adds embeddings filter option to return scores in state (#12489)

CC @baskaryan @assafelovic

libs/langchain/langchain/retrievers/document_compressors/embeddings_filter.py

@@ -67,4 +67,6 @@ class EmbeddingsFilter(BaseDocumentCompressor):
                 similarity[included_idxs] > self.similarity_threshold
             )
             included_idxs = included_idxs[similar_enough]
+        for i in included_idxs:
+            stateful_documents[i].state["query_similarity_score"] = similarity[i]
         return [stateful_documents[i] for i in included_idxs]

libs/langchain/tests/integration_tests/retrievers/document_compressors/test_embeddings_filter.py

@@ -35,7 +35,9 @@ def test_embeddings_filter_with_state() -> None:
     state = {"embedded_doc": np.zeros(len(embedded_query))}
     docs = [_DocumentWithState(page_content=t, state=state) for t in texts]
     docs[-1].state = {"embedded_doc": embedded_query}
-    relevant_filter = EmbeddingsFilter(embeddings=embeddings, similarity_threshold=0.75)
+    relevant_filter = EmbeddingsFilter(
+        embeddings=embeddings, similarity_threshold=0.75, return_similarity_scores=True
+    )
     actual = relevant_filter.compress_documents(docs, query)
     assert len(actual) == 1
     assert texts[-1] == actual[0].page_content