Add rerank document compressor

libs/aws/langchain_aws/__init__.py

@@ -14,6 +14,7 @@
     InMemorySemanticCache,
     InMemoryVectorStore,
 )
+from langchain_aws.document_compressors.rerank import BedrockRerank
 
 
 def setup_logging():
@@ -48,4 +49,5 @@ def setup_logging():
     "NeptuneGraph",
     "InMemoryVectorStore",
     "InMemorySemanticCache",
+    "BedrockRerank"
 ]

libs/aws/langchain_aws/document_compressors/rerank.py

@@ -0,0 +1,136 @@
+from copy import deepcopy
+from typing import Any, Dict, List, Optional, Sequence, Union
+
+import boto3
+from langchain_core.callbacks.manager import Callbacks
+from langchain_core.documents import BaseDocumentCompressor, Document
+from langchain_core.utils import from_env
+from pydantic import ConfigDict, Field, model_validator
+from typing_extensions import Self
+
+
+class BedrockRerank(BaseDocumentCompressor):
+    """Document compressor that uses AWS Bedrock Rerank API."""
+
+    model_arn: str
+    """The ARN of the reranker model."""
+    client: Any = None
+    """Bedrock client to use for compressing documents."""
+    top_n: Optional[int] = 3
+    """Number of documents to return."""
+    region_name: str = Field(
+        default_factory=from_env("AWS_DEFAULT_REGION", default=None)
+    )
+    """AWS region to initialize the Bedrock client."""
+    credentials_profile_name: Optional[str] = Field(
+        default_factory=from_env("AWS_PROFILE", default=None)
+    )
+    """AWS profile for authentication, optional."""
+
+    model_config = ConfigDict(
+        extra="forbid",
+        arbitrary_types_allowed=True,
+    )
+
+    @model_validator(mode="after")
+    def initialize_client(self) -> Self:
+        """Initialize the AWS Bedrock client."""
+        if not self.client:
+            session = self._get_session()
+            self.client = session.client(
+                "bedrock-agent-runtime",
+                region_name=self.region_name
+                )
+        return self
+
+    def _get_session(self):
+        return (
+            boto3.Session(profile_name=self.credentials_profile_name)
+            if self.credentials_profile_name
+            else boto3.Session()
+        )
+
+    def rerank(
+        self,
+        documents: Sequence[Union[str, Document, dict]],
+        query: str,
+        top_n: Optional[int] = None,
+        extra_model_fields: Optional[Dict[str, Any]] = None,
+    ) -> List[Dict[str, Any]]:
+        """Returns an ordered list of documents based on their relevance to the query.
+
+        Args:
+            query: The query to use for reranking.
+            documents: A sequence of documents to rerank.
+            top_n: The number of top-ranked results to return. Defaults to self.top_n.
+            extra_model_fields: A dictionary of additional fields to pass to the model.
+
+        Returns:
+            List[Dict[str, Any]]: A list of ranked documents with relevance scores.
+        """
+        if len(documents) == 0:
+            return []
+
+        # Serialize documents for the Bedrock API
+        serialized_documents = [
+            {"textDocument": {"text": doc.page_content}, "type": "TEXT"}
+            if isinstance(doc, Document)
+            else {"textDocument": {"text": doc}, "type": "TEXT"}
+            if isinstance(doc, str)
+            else {"jsonDocument": doc, "type": "JSON"}
+            for doc in documents
+        ]
+
+        request_body = {
+            "queries": [{"textQuery": {"text": query}, "type": "TEXT"}],
+            "rerankingConfiguration": {
+                "bedrockRerankingConfiguration": {
+                    "modelConfiguration": {
+                        "modelArn": self.model_arn,
+                        "additionalModelRequestFields": extra_model_fields
+                        or {},
+                    },
+                    "numberOfResults": top_n or self.top_n,
+                },
+                "type": "BEDROCK_RERANKING_MODEL",
+            },
+            "sources": [
+                {"inlineDocumentSource": doc, "type": "INLINE"}
+                for doc in serialized_documents
+            ],
+        }
+
+        response = self.client.rerank(**request_body)
+        response_body = response.get("results", [])
+
+        results = [
+            {"index": result["index"], "relevance_score": result["relevanceScore"]}
+            for result in response_body
+        ]
+
+        return results
+
+    def compress_documents(
+        self,
+        documents: Sequence[Document],
+        query: str,
+        callbacks: Optional[Callbacks] = None,
+    ) -> Sequence[Document]:
+        """
+        Compress documents using Bedrock's rerank API.
+
+        Args:
+            documents: A sequence of documents to compress.
+            query: The query to use for compressing the documents.
+            callbacks: Callbacks to run during the compression process.
+
+        Returns:
+            A sequence of compressed documents.
+        """
+        compressed = []
+        for res in self.rerank(documents, query):
+            doc = documents[res["index"]]
+            doc_copy = Document(doc.page_content, metadata=deepcopy(doc.metadata))
+            doc_copy.metadata["relevance_score"] = res["relevance_score"]
+            compressed.append(doc_copy)
+        return compressed

libs/aws/tests/unit_tests/document_compressors/test_rerank.py

@@ -0,0 +1,55 @@
+from unittest.mock import MagicMock, patch
+
+import pytest
+from langchain_core.documents import Document
+
+from langchain_aws.document_compressors.rerank import BedrockRerank
+
+
+@pytest.fixture
+def reranker():
+    reranker = BedrockRerank(
+        model_arn="arn:aws:bedrock:us-west-2::foundation-model/amazon.rerank-v1:0"
+        )
+    reranker.client = MagicMock()
+    return reranker
+
+@patch("boto3.Session")
+def test_initialize_client(mock_boto_session, reranker):
+    session_instance = MagicMock()
+    mock_boto_session.return_value = session_instance
+    session_instance.client.return_value = MagicMock()
+    reranker.initialize_client()
+    assert reranker.client is not None
+
+@patch("langchain_aws.document_compressors.rerank.BedrockRerank.rerank")
+def test_rerank(mock_rerank, reranker):
+    mock_rerank.return_value = [
+        {"index": 0, "relevance_score": 0.9},
+        {"index": 1, "relevance_score": 0.8},
+    ]
+
+    documents = [Document(page_content="Doc 1"), Document(page_content="Doc 2")]
+    query = "Example Query"
+    results = reranker.rerank(documents, query)
+
+    assert len(results) == 2
+    assert results[0]["index"] == 0
+    assert results[0]["relevance_score"] == 0.9
+    assert results[1]["index"] == 1
+    assert results[1]["relevance_score"] == 0.8
+
+@patch("langchain_aws.document_compressors.rerank.BedrockRerank.rerank")
+def test_compress_documents(mock_rerank, reranker):
+    mock_rerank.return_value = [
+        {"index": 0, "relevance_score": 0.95},
+        {"index": 1, "relevance_score": 0.85},
+    ]
+
+    documents = [Document(page_content="Content 1"), Document(page_content="Content 2")]
+    query = "Relevant query"
+    compressed_docs = reranker.compress_documents(documents, query)
+
+    assert len(compressed_docs) == 2
+    assert compressed_docs[0].metadata["relevance_score"] == 0.95
+    assert compressed_docs[1].metadata["relevance_score"] == 0.85