diff --git a/.github/workflows/pylint.yml b/.github/workflows/pylint.yml
index 4d1bcde..6233948 100644
--- a/.github/workflows/pylint.yml
+++ b/.github/workflows/pylint.yml
@@ -4,7 +4,7 @@ on: [push]
 
 jobs:
   build:
-    runs-on: node20
+    runs-on: ubuntu-latest
     strategy:
       matrix:
         python-version: ["3.12"]
diff --git a/src/interface.py b/src/interface.py
new file mode 100644
index 0000000..b961067
--- /dev/null
+++ b/src/interface.py
@@ -0,0 +1,294 @@
+"""
+@brief: Specify the project interface.
+@author: Hao Kang <haok@andrew.cmu.edu>
+"""
+
+from pathlib import Path
+from abc import ABC, abstractmethod
+from typing import Iterator, Literal, List, Type, Tuple
+from torch import Tensor
+
+
+# Define the type aliases.
+EmbeddingName = Literal["BgeBase", "MiniCPM"]
+DatasetName = Literal["MsMarco", "Beir"]
+PartitionName = Literal["Train", "Validate"]
+
+
+class Embedding(ABC):
+    """
+    The interface for an embedding model.
+
+    Attributes:
+        name (EmbeddingName): The name of the embedding.
+        size (int): The size of the embedding.
+    """
+
+    name: EmbeddingName
+    size: int
+
+    @abstractmethod
+    def __init__(self, devices: List[int]) -> None:
+        """
+        Initialize the embedding model.
+
+        :type devices: List[int]
+        :param devices: The devices to use for embedding.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def forward(self, passages: List[str]) -> Tensor:
+        """
+        Forward pass to embed the given passages.
+
+        :type passages: List[str]
+        :param passages: The list of passages to embed.
+        :rtype: torch.Tensor
+        :return: The computed embeddings in a tensor of shape (N, D), where N
+            is the number of passages and D is the embedding size.
+        """
+        raise NotImplementedError
+
+
+class Dataset(ABC):
+    """
+    The interface for a dataset.
+
+    Attributes:
+        name (DatasetName): The name of the dataset.
+    """
+
+    name: DatasetName
+
+    @abstractmethod
+    def didIter(self, batchSize: int) -> Iterator[List[str]]:
+        """
+        Iterate over the document IDs in batches.
+
+        :type batchSize: int
+        :param batchSize: The batch size for each iteration.
+        :rtype: Iterator[List[str]]
+        :return: An iterator over the document IDs. Each iteration yields a
+            list of document IDs.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def docIter(self, batchSize: int) -> Iterator[List[str]]:
+        """
+        Iterate over the document texts in batches.
+
+        :type batchSize: int
+        :param batchSize: The batch size for each iteration.
+        :rtype: Iterator[List[str]]
+        :return: The iterator over the document texts. Each iteration yields a
+            list of document texts.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def docEmbIter(
+        self,
+        embedding: Type[Embedding],
+        batchSize: int,
+        numWorkers: int,
+        shuffle: bool,
+    ) -> Iterator[Tensor]:
+        """
+        Iterate over the document embeddings in batches.
+
+        :type embedding: Type[Embedding]
+        :param embedding: The embedding model to use.
+        :type batchSize: int
+        :param batchSize: The batch size for each iteration.
+        :type numWorkers: int
+        :param numWorkers: The number of workers for data loading.
+        :type shuffle: bool
+        :param shuffle: Whether to shuffle the data during loading.
+        :rtype: Iterator[Tensor]
+        :return: The iterator over the document embeddings. Each tensor has
+            shape (N, D), where N is the batch size, or less for the last
+            batch, and D is the embedding size.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def getDocLen(self) -> int:
+        """
+        Get the number of documents.
+
+        :rtype: int
+        :return: The number of documents.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def qidIter(
+        self, split: PartitionName, batchSize: int
+    ) -> Iterator[List[str]]:
+        """
+        Iterate over the query IDs in batches.
+
+        :type split: PartitionName
+        :param split: Whether to use the training or validation split.
+        :type batchSize: int
+        :param batchSize: The batch size for each iteration.
+        :rtype: Iterator[List[str]]
+        :return: The iterator over the query IDs. Each iteration yields a list
+            of query IDs.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def qryIter(
+        self, split: PartitionName, batchSize: int
+    ) -> Iterator[List[str]]:
+        """
+        Iterate over the query texts in batches.
+
+        :type split: PartitionName
+        :param split: Whether to use the training or validation split.
+        :type batchSize: int
+        :param batchSize: The batch size for each iteration.
+        :rtype: Iterator[List[str]]
+        :return: The iterator over the query texts. Each iteration yields a
+            list of query texts.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def qryEmbIter(
+        self,
+        split: PartitionName,
+        embedding: Type[Embedding],
+        batchSize: int,
+        numWorkers: int,
+        shuffle: bool,
+    ) -> Iterator[Tensor]:
+        """
+        Iterate over the query embeddings in batches.
+
+        :type split: PartitionName
+        :param split: Whether to use the training or validation split.
+        :type embedding: Type[Embedding]
+        :param embedding: The embedding class to use.
+        :type batchSize: int
+        :param batchSize: The batch size for each iteration.
+        :type numWorkers: int
+        :param numWorkers: The number of workers for data loading.
+        :type shuffle: bool
+        :param shuffle: Whether to shuffle the data.
+        :rtype: Iterator[Tensor]
+        :return: The iterator over the query embeddings. Each tensor has shape
+            (N, D), where N is the batch size, or less for the last batch, and
+            D is the embedding size.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def getQryLen(self, split: PartitionName) -> int:
+        """
+        Get the number of queries.
+
+        :type split: PartitionName
+        :param split: Whether to use the training or validation split.
+        :rtype: int
+        :return: The number of queries.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def getQryRel(self, split: PartitionName) -> Path:
+        """
+        Get the path to the query relevance file.
+
+        :type split: PartitionName
+        :param split: Whether to use the training or validation split.
+        :rtype: Path
+        :return: The path to the query relevance file.
+        """
+
+    @abstractmethod
+    def mixEmbIter(
+        self,
+        split: PartitionName,
+        embedding: Type[Embedding],
+        relevant: int,
+        batchSize: int,
+        numWorkers: int,
+        shuffle: bool,
+    ) -> Iterator[Tuple[Tensor, Tensor]]:
+        """
+        Iterate over the embeddings of query and its retrieved documents in
+            batches.
+
+        :type split: PartitionName
+        :param split: Whether to use the training or validation split.
+        :type embedding: Type[Embedding]
+        :param embedding: The embedding class to use.
+        :type relevant: int
+        :param relevant: The number of documents to include for each query.
+        :type batchSize: int
+        :param batchSize: The batch size for each iteration.
+        :type numWorkers: int
+        :param numWorkers: The number of workers for data loading.
+        :type shuffle: bool
+        :param shuffle: Whether to shuffle the data.
+        :rtype: Iterator[Tuple[Tensor, Tensor]]
+        :return: The iterator over the query and document embeddings. The
+            first tensor is the query embeddings and has shape (N, D), where N
+            is the batch size, or less for the last batch, and D is the
+            embedding size. The second tensor is the document embeddings and
+            has shape (N, K, D), where K is the number of relevant documents.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def getMixLen(self, split: PartitionName) -> int:
+        """
+        Get the number of query-document pairs.
+        This function is equival to getQryLen.
+
+        :type split: PartitionName
+        :param split: Whether to use the training or validation split.
+        :rtype: int
+        :return: The number of query-document pairs.
+        """
+        raise NotImplementedError
+
+
+class SAE(ABC):
+    """
+    The interface for a sparse autoencoder.
+    """
+
+    def __init__(self, features: int, expandBy: int) -> None:
+        """
+        Initialize the sparse autoencoder.
+
+        :type features: int
+        :param features: The embedding size.
+        :type expandBy: int
+        :param expandBy: Expand factor for the dictionary.
+        """
+        raise NotImplementedError
+
+    def forward(self, x: Tensor, activate: int) -> Tuple[Tensor, Tensor]:
+        """
+        Forward pass to reconstruct the embedding.
+
+        :type x: Tensor
+        :param x: The original embedding. The tensor has shape (N, D), where N
+            is the batch size and D is the embedding size.
+        :type K: int
+        :param activate: The number of features to activate. This is the
+            sparsity constraint. Only the top-K features are activated. The
+            rest are set to zero.
+        :rtype: Tuple[Tensor, Tensor]
+        :return: The latent features and the reconstructed embedding. The
+            latent features have shape (N, D), where D is the dictionary size.
+            The reconstructed embedding has shape (N, E), where E is the
+            embedding size. N is the batch size in both cases.
+        """
+        raise NotImplementedError