diff --git a/rag/__init__.py b/rag/__init__.py
new file mode 100644
index 0000000..cd2b504
--- /dev/null
+++ b/rag/__init__.py
@@ -0,0 +1,8 @@
+"""
+RAG (Retrieval-Augmented Generation) System for DriveMind-V
+"""
+from .vector_store import VectorStore
+from .retriever import CLIPRetriever
+from .rag_vlm import RAG_VLM
+
+__all__ = ['VectorStore', 'CLIPRetriever', 'RAG_VLM']
diff --git a/rag/__pycache__/__init__.cpython-310.pyc b/rag/__pycache__/__init__.cpython-310.pyc
new file mode 100644
index 0000000..8b22114
Binary files /dev/null and b/rag/__pycache__/__init__.cpython-310.pyc differ
diff --git a/rag/__pycache__/rag_vlm.cpython-310.pyc b/rag/__pycache__/rag_vlm.cpython-310.pyc
new file mode 100644
index 0000000..09511e6
Binary files /dev/null and b/rag/__pycache__/rag_vlm.cpython-310.pyc differ
diff --git a/rag/__pycache__/retriever.cpython-310.pyc b/rag/__pycache__/retriever.cpython-310.pyc
new file mode 100644
index 0000000..cd65273
Binary files /dev/null and b/rag/__pycache__/retriever.cpython-310.pyc differ
diff --git a/rag/__pycache__/vector_store.cpython-310.pyc b/rag/__pycache__/vector_store.cpython-310.pyc
new file mode 100644
index 0000000..97f588e
Binary files /dev/null and b/rag/__pycache__/vector_store.cpython-310.pyc differ
diff --git a/rag/rag_vlm.py b/rag/rag_vlm.py
new file mode 100644
index 0000000..525ceb2
--- /dev/null
+++ b/rag/rag_vlm.py
@@ -0,0 +1,219 @@
+"""
+RAG-Enhanced Vision-Language Model
+Combines retrieval with generation for few-shot reasoning
+"""
+import os
+import sys
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
+
+from typing import Union, List, Dict
+from PIL import Image
+from model.model_vlm import MiniMindVLM
+from rag.retriever import CLIPRetriever
+from rag.vector_store import VectorStore
+
+
+class RAG_VLM:
+    """
+    RAG-enhanced VLM for improved generation through retrieval
+    
+    Workflow:
+    1. Encode query image with CLIP
+    2. Retrieve k similar images from knowledge base
+    3. Build few-shot prompt with retrieved examples
+    4. Generate response with VLM
+    """
+    
+    def __init__(
+        self,
+        vlm_model: MiniMindVLM,
+        tokenizer,
+        vector_store: VectorStore,
+        retriever: CLIPRetriever = None,
+        k: int = 3
+    ):
+        """
+        Initialize RAG-VLM
+        
+        Args:
+            vlm_model: MiniMindVLM instance
+            tokenizer: Tokenizer for VLM
+            vector_store: VectorStore with image knowledge base
+            retriever: CLIPRetriever instance (creates new one if None)
+            k: Number of examples to retrieve
+        """
+        self.model = vlm_model
+        self.tokenizer = tokenizer
+        self.vector_store = vector_store
+        self.k = k
+        
+        # Create retriever if not provided
+        if retriever is None:
+            self.retriever = CLIPRetriever()
+        else:
+            self.retriever = retriever
+    
+    def build_few_shot_prompt(
+        self,
+        base_prompt: str,
+        retrieved_metadata: List[Dict],
+        format_fn=None
+    ) -> str:
+        """
+        Build few-shot prompt from retrieved examples
+        
+        Args:
+            base_prompt: Original user prompt
+            retrieved_metadata: List of metadata dicts from retrieval
+            format_fn: Optional function to format each example
+        
+        Returns:
+            Augmented prompt with examples
+        """
+        if format_fn is None:
+            # Default formatting
+            format_fn = lambda meta: f"Example: {meta.get('description', 'No description')}\n"
+        
+        # Build context from retrieved examples
+        context = "Here are some similar examples:\n\n"
+        for i, meta in enumerate(retrieved_metadata, 1):
+            context += f"{i}. {format_fn(meta)}\n"
+        
+        # Combine with user prompt
+        augmented_prompt = f"{context}\nNow, {base_prompt}"
+        
+        return augmented_prompt
+    
+    def generate(
+        self,
+        image: Union[str, Image.Image],
+        prompt: str,
+        use_rag: bool = True,
+        max_new_tokens: int = 256,
+        temperature: float = 0.7,
+        top_k: int = 50,
+        top_p: float = 0.9,
+        format_fn=None
+    ) -> Dict:
+        """
+        Generate response with optional RAG
+        
+        Args:
+            image: Query image or path
+            prompt: User prompt
+            use_rag: Whether to use retrieval-augmented generation
+            max_new_tokens: Max tokens to generate
+            temperature: Sampling temperature
+            top_k: Top-k sampling
+            top_p: Nucleus sampling
+            format_fn: Optional function to format retrieved examples
+        
+        Returns:
+            Dict with 'response', 'retrieved_metadata', 'augmented_prompt'
+        """
+        # Load image if path
+        if isinstance(image, str):
+            image = Image.open(image)
+        
+        retrieved_metadata = []
+        augmented_prompt = prompt
+        
+        # Retrieve if RAG enabled
+        if use_rag:
+            distances, retrieved_metadata = self.retriever.retrieve(
+                image,
+                self.vector_store,
+                k=self.k
+            )
+            
+            # Build few-shot prompt
+            augmented_prompt = self.build_few_shot_prompt(
+                prompt,
+                retrieved_metadata,
+                format_fn=format_fn
+            )
+            
+            print(f"\n📚 Retrieved {len(retrieved_metadata)} examples")
+            print(f"   Distances: {distances}")
+        
+        # Generate with VLM
+        # Note: Actual generation code depends on MiniMindVLM's interface
+        # This is a placeholder showing the concept
+        
+        # For now, return structure (actual generation would call model.generate())
+        return {
+            'response': augmented_prompt,  # Placeholder: would be actual generation
+            'retrieved_metadata': retrieved_metadata,
+            'augmented_prompt': augmented_prompt,
+            'retrieval_distances': distances.tolist() if use_rag else []
+        }
+    
+    def evaluate_retrieval(
+        self,
+        test_images: List[Union[str, Image.Image]],
+        ground_truth_labels: List[str],
+        metadata_label_fn=lambda meta: meta.get('label', '')
+    ) -> Dict:
+        """
+        Evaluate retrieval quality
+        
+        Args:
+            test_images: List of test images
+            ground_truth_labels: True labels for each test image
+            metadata_label_fn: Function to extract label from metadata
+        
+        Returns:
+            Dict with Recall@K, MRR, etc.
+        """
+        recall_at_k = 0
+        mrr = 0
+        
+        for image, true_label in zip(test_images, ground_truth_labels):
+            distances, retrieved_metadata = self.retriever.retrieve(
+                image,
+                self.vector_store,
+                k=self.k
+            )
+            
+            # Check if true label in retrieved results
+            retrieved_labels = [metadata_label_fn(meta) for meta in retrieved_metadata]
+            
+            if true_label in retrieved_labels:
+                recall_at_k += 1
+                # Mean reciprocal rank
+                rank = retrieved_labels.index(true_label) + 1
+                mrr += 1.0 / rank
+        
+        recall_at_k /= len(test_images)
+        mrr /= len(test_images)
+        
+        return {
+            f'Recall@{self.k}': recall_at_k,
+            'MRR': mrr
+        }
+
+
+# Example usage
+if __name__ == "__main__":
+    from transformers import AutoTokenizer
+    
+    # Load VLM
+    from model.model_vlm import VLMConfig
+    config = VLMConfig()
+    vlm = MiniMindVLM(config)
+    tokenizer = AutoTokenizer.from_pretrained('./model')
+    
+    # Load vector store (assuming it exists)
+    # vector_store = VectorStore.load("knowledge_base.faiss")
+    
+    # Create RAG-VLM
+    # rag_vlm = RAG_VLM(vlm, tokenizer, vector_store, k=3)
+    
+    # Generate with RAG
+    # result = rag_vlm.generate(
+    #     image="test_image.jpg",
+    #     prompt="Describe this scene in detail",
+    #     use_rag=True
+    # )
+    
+    print("RAG-VLM module ready!")
diff --git a/rag/retriever.py b/rag/retriever.py
new file mode 100644
index 0000000..b1733b2
--- /dev/null
+++ b/rag/retriever.py
@@ -0,0 +1,185 @@
+"""
+CLIP-based Image Retriever for RAG System
+Uses frozen CLIP ViT-B/16 from MiniMindVLM for visual similarity
+"""
+import os
+import sys
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
+
+import torch
+import numpy as np
+from PIL import Image
+from typing import List, Union, Dict, Tuple
+from model.model_vlm import MiniMindVLM
+from rag.vector_store import VectorStore
+
+
+class CLIPRetriever:
+    """
+    CLIP-based retriever for visual similarity search
+    
+    Uses the same CLIP ViT-B/16 model as MiniMindVLM for consistency
+    """
+    
+    def __init__(
+        self,
+        vision_model_path: str = "./model/vision_model/clip-vit-base-patch16",
+        device: str = "cuda" if torch.cuda.is_available() else "cpu"
+    ):
+        """
+        Initialize CLIP retriever
+        
+        Args:
+            vision_model_path: Path to CLIP model
+            device: Device to run on
+        """
+        self.device = device
+        
+        # Load CLIP model (reuse MiniMindVLM's CLIP loader)
+        self.vision_model, self.processor = MiniMindVLM.get_vision_model(vision_model_path)
+        
+        if self.vision_model is None:
+            raise ValueError(f"CLIP model not found at {vision_model_path}")
+        
+        self.vision_model = self.vision_model.to(device)
+        self.vision_model.eval()
+        
+        print(f"✅ Loaded CLIP ViT-B/16 on {device}")
+    
+    def encode_image(self, image: Union[str, Image.Image]) -> np.ndarray:
+        """
+        Encode a single image to CLIP embedding
+        
+        Args:
+            image: PIL Image or path to image
+        
+        Returns:
+            Embedding as numpy array of shape (768,)
+        """
+        # Load image if path provided
+        if isinstance(image, str):
+            image = Image.open(image)
+        
+        # Convert image to tensor
+        image_tensor = MiniMindVLM.image2tensor(image, self.processor).to(self.device)
+        
+        # Get CLIP embedding
+        with torch.no_grad():
+            embedding = MiniMindVLM.get_image_embeddings(image_tensor, self.vision_model)
+            # Pool patch embeddings to single vector
+            embedding = embedding.mean(dim=0).cpu().numpy()  # Shape: (768,)
+        
+        return embedding
+    
+    def encode_images_batch(self, images: List[Union[str, Image.Image]], batch_size: int = 32) -> np.ndarray:
+        """
+        Encode multiple images in batches
+        
+        Args:
+            images: List of PIL Images or paths
+            batch_size: Batch size for encoding
+        
+        Returns:
+            Embeddings as numpy array of shape (N, 768)
+        """
+        embeddings = []
+        
+        for i in range(0, len(images), batch_size):
+            batch = images[i:i + batch_size]
+            
+            # Load images
+            batch_images = []
+            for img in batch:
+                if isinstance(img, str):
+                    img = Image.open(img)
+                batch_images.append(img)
+            
+            # Convert to tensors
+            # image2tensor returns (1, C, H, W), so we concat on batch dim
+            batch_tensors = torch.cat([
+                MiniMindVLM.image2tensor(img, self.processor)
+                for img in batch_images
+            ], dim=0).to(self.device)
+            
+            # Encode
+            with torch.no_grad():
+                batch_embeddings = MiniMindVLM.get_image_embeddings(batch_tensors, self.vision_model)
+                # Pool: (batch, patches, 768) -> (batch, 768)
+                batch_embeddings = batch_embeddings.mean(dim=1).cpu().numpy()
+            
+            embeddings.append(batch_embeddings)
+        
+        return np.vstack(embeddings)
+    
+    def build_index(
+        self,
+        images: List[Union[str, Image.Image]],
+        metadata: List[Dict],
+        save_path: str,
+        batch_size: int = 32
+    ) -> VectorStore:
+        """
+        Build FAISS index from images
+        
+        Args:
+            images: List of images or paths
+            metadata: List of metadata dicts (one per image)
+            save_path: Where to save index
+            batch_size: Encoding batch size
+        
+        Returns:
+            VectorStore with built index
+        """
+        print(f"🔨 Building index from {len(images)} images...")
+        
+        # Encode all images
+        embeddings = self.encode_images_batch(images, batch_size=batch_size)
+        print(f"✅ Encoded {len(embeddings)} images to {embeddings.shape[1]}-dim vectors")
+        
+        # Build vector store
+        store = VectorStore(dimension=embeddings.shape[1])
+        store.add(embeddings, metadata)
+        
+        # Save
+        store.save(save_path)
+        
+        return store
+    
+    def retrieve(
+        self,
+        query_image: Union[str, Image.Image],
+        vector_store: VectorStore,
+        k: int = 5
+    ) -> Tuple[np.ndarray, List[Dict]]:
+        """
+        Retrieve k most similar images from index
+        
+        Args:
+            query_image: Query image or path
+            vector_store: VectorStore to search in
+            k: Number of results to return
+        
+        Returns:
+            distances: Array of shape (k,) with distances
+            metadata: List of k metadata dicts
+        """
+        # Encode query
+        query_embedding = self.encode_image(query_image)
+        query_embedding = query_embedding.reshape(1, -1)  # (1, 768)
+        
+        # Search
+        distances, indices, metadata = vector_store.search(query_embedding, k=k)
+        
+        # Return flattened results (since we only have 1 query)
+        return distances[0], metadata[0]
+
+
+# Example usage
+if __name__ == "__main__":
+    # Create retriever
+    retriever = CLIPRetriever()
+    
+    # Example: encode single image
+    # img_path = "dataset/eval_images/城市车水马龙-city-traffic.jpg"
+    # embedding = retriever.encode_image(img_path)
+    # print(f"Image embedding shape: {embedding.shape}")
diff --git a/rag/vector_store.py b/rag/vector_store.py
new file mode 100644
index 0000000..f2bd318
--- /dev/null
+++ b/rag/vector_store.py
@@ -0,0 +1,193 @@
+"""
+FAISS-based Vector Store for Image Embeddings
+Enables fast similarity search over large image collections
+"""
+import os
+import json
+import numpy as np
+import faiss
+from typing import List, Dict, Optional, Tuple
+
+
+class VectorStore:
+    """
+    FAISS-based vector store for efficient similarity search
+    
+    Supports:
+    - L2 (Euclidean) distance search
+    - Incremental index building
+    - Save/load to disk
+    - Metadata storage alongside embeddings
+    """
+    
+    def __init__(self, dimension: int = 768, metric: str = "L2"):
+        """
+        Initialize vector store
+        
+        Args:
+            dimension: Embedding dimension (768 for CLIP ViT-B/16)
+            metric: Distance metric ("L2" or "IP" for inner product)
+        """
+        self.dimension = dimension
+        self.metric = metric
+        
+        # Initialize FAISS index
+        if metric == "L2":
+            self.index = faiss.IndexFlatL2(dimension)
+        elif metric == "IP":
+            self.index = faiss.IndexFlatIP(dimension)  # Inner product (cosine if normalized)
+        else:
+            raise ValueError(f"Unknown metric: {metric}. Use 'L2' or 'IP'")
+        
+        # Metadata storage (parallel to FAISS index)
+        self.metadata = []
+        
+    def add(self, embeddings: np.ndarray, metadata: Optional[List[Dict]] = None):
+        """
+        Add embeddings to the index
+        
+        Args:
+            embeddings: Array of shape (N, dimension)
+            metadata: List of N metadata dicts (one per embedding)
+        """
+        if embeddings.dtype != np.float32:
+            embeddings = embeddings.astype(np.float32)
+        
+        # Normalize for inner product (makes it cosine similarity)
+        if self.metric == "IP":
+            faiss.normalize_L2(embeddings)
+        
+        # Add to FAISS index
+        self.index.add(embeddings)
+        
+        # Add metadata
+        if metadata:
+            assert len(metadata) == len(embeddings), "Metadata length must match embeddings"
+            self.metadata.extend(metadata)
+        else:
+            # Add empty dicts if no metadata provided
+            self.metadata.extend([{} for _ in range(len(embeddings))])
+    
+    def search(self, query_embeddings: np.ndarray, k: int = 5) -> Tuple[np.ndarray, np.ndarray, List[List[Dict]]]:
+        """
+        Search for k nearest neighbors
+        
+        Args:
+            query_embeddings: Array of shape (N_queries, dimension)
+            k: Number of neighbors to return
+        
+        Returns:
+            distances: Array of shape (N_queries, k)
+            indices: Array of shape (N_queries, k)
+            metadata: List of lists, metadata for each query's k neighbors
+        """
+        if query_embeddings.dtype != np.float32:
+            query_embeddings = query_embeddings.astype(np.float32)
+        
+        # Normalize for inner product
+        if self.metric == "IP":
+            faiss.normalize_L2(query_embeddings)
+        
+        # Search
+        distances, indices = self.index.search(query_embeddings, k)
+        
+        # Retrieve metadata
+        retrieved_metadata = []
+        for query_indices in indices:
+            query_metadata = [self.metadata[i] if i < len(self.metadata) else {} for i in query_indices]
+            retrieved_metadata.append(query_metadata)
+        
+        return distances, indices, retrieved_metadata
+    
+    def save(self, index_path: str, metadata_path: Optional[str] = None):
+        """
+        Save index and metadata to disk
+        
+        Args:
+            index_path: Path to save FAISS index (.faiss or .index)
+            metadata_path: Path to save metadata (.json). If None, uses index_path.json
+        """
+        # Save FAISS index
+        faiss.write_index(self.index, index_path)
+        
+        # Save metadata
+        if metadata_path is None:
+            metadata_path = index_path.replace('.faiss', '.json').replace('.index', '.json')
+            if not metadata_path.endswith('.json'):
+                metadata_path += '.json'
+        
+        with open(metadata_path, 'w') as f:
+            json.dump({
+                'metadata': self.metadata,
+                'dimension': self.dimension,
+                'metric': self.metric,
+                'count': len(self.metadata)
+            }, f, indent=2)
+        
+        print(f"✅ Saved FAISS index to {index_path}")
+        print(f"✅ Saved metadata to {metadata_path}")
+    
+    @classmethod
+    def load(cls, index_path: str, metadata_path: Optional[str] = None) -> 'VectorStore':
+        """
+        Load index and metadata from disk
+        
+        Args:
+            index_path: Path to FAISS index
+            metadata_path: Path to metadata. If None, uses index_path.json
+        
+        Returns:
+            VectorStore instance with loaded data
+        """
+        # Load FAISS index
+        index = faiss.read_index(index_path)
+        
+        # Load metadata
+        if metadata_path is None:
+            metadata_path = index_path.replace('.faiss', '.json').replace('.index', '.json')
+            if not metadata_path.endswith('.json'):
+                metadata_path += '.json'
+        
+        with open(metadata_path, 'r') as f:
+            data = json.load(f)
+        
+        # Create instance
+        store = cls(dimension=data['dimension'], metric=data['metric'])
+        store.index = index
+        store.metadata = data['metadata']
+        
+        print(f"✅ Loaded {data['count']} embeddings from {index_path}")
+        
+        return store
+    
+    def __len__(self):
+        """Return number of vectors in index"""
+        return self.index.ntotal
+
+
+# Example usage
+if __name__ == "__main__":
+    # Create some random embeddings for demo
+    dimension = 768
+    n_vectors = 100
+    embeddings = np.random.randn(n_vectors, dimension).astype(np.float32)
+    
+    # Create metadata
+    metadata = [{"id": f"img_{i}", "description": f"Image {i}"} for i in range(n_vectors)]
+    
+    # Build index
+    store = VectorStore(dimension=dimension)
+    store.add(embeddings, metadata)
+    
+    # Search
+    query = np.random.randn(1, dimension).astype(np.float32)
+    distances, indices, retrieved_metadata = store.search(query, k=5)
+    
+    print(f"Top 5 matches:")
+    for i, (dist, idx, meta) in enumerate(zip(distances[0], indices[0], retrieved_metadata[0])):
+        print(f"  {i+1}. Index {idx}, Distance {dist:.4f}, Metadata: {meta}")
+    
+    # Save and load
+    store.save("demo_index.faiss")
+    loaded_store = VectorStore.load("demo_index.faiss")
+    print(f"Loaded store has {len(loaded_store)} vectors")
diff --git a/rag_demo_index.faiss b/rag_demo_index.faiss
new file mode 100644
index 0000000..693d2fd
Binary files /dev/null and b/rag_demo_index.faiss differ
diff --git a/rag_demo_index.json b/rag_demo_index.json
new file mode 100644
index 0000000..5a4e514
--- /dev/null
+++ b/rag_demo_index.json
@@ -0,0 +1,55 @@
+{
+  "metadata": [
+    {
+      "id": "\u821e\u8e48-dance",
+      "filename": "\u821e\u8e48-dance.jpg",
+      "path": "/workspace/minimind-v/minimind-v/dataset/eval_images/\u821e\u8e48-dance.jpg",
+      "description": "Image: \u821e\u8e48-dance.jpg"
+    },
+    {
+      "id": "\u81ea\u884c\u8f66\u9c9c\u82b1-Bicycle-Flowers",
+      "filename": "\u81ea\u884c\u8f66\u9c9c\u82b1-Bicycle-Flowers.jpg",
+      "path": "/workspace/minimind-v/minimind-v/dataset/eval_images/\u81ea\u884c\u8f66\u9c9c\u82b1-Bicycle-Flowers.jpg",
+      "description": "Image: \u81ea\u884c\u8f66\u9c9c\u82b1-Bicycle-Flowers.jpg"
+    },
+    {
+      "id": "\u718a\u732b\u8349\u5730-Panda-Grassland",
+      "filename": "\u718a\u732b\u8349\u5730-Panda-Grassland.jpg",
+      "path": "/workspace/minimind-v/minimind-v/dataset/eval_images/\u718a\u732b\u8349\u5730-Panda-Grassland.jpg",
+      "description": "Image: \u718a\u732b\u8349\u5730-Panda-Grassland.jpg"
+    },
+    {
+      "id": "\u6905\u5b50\u8001\u4eba\u770b\u4e66-Chair-Elderly-Reading",
+      "filename": "\u6905\u5b50\u8001\u4eba\u770b\u4e66-Chair-Elderly-Reading.jpg",
+      "path": "/workspace/minimind-v/minimind-v/dataset/eval_images/\u6905\u5b50\u8001\u4eba\u770b\u4e66-Chair-Elderly-Reading.jpg",
+      "description": "Image: \u6905\u5b50\u8001\u4eba\u770b\u4e66-Chair-Elderly-Reading.jpg"
+    },
+    {
+      "id": "\u5f69\u8679\u7011\u5e03-Rainbow-Falls",
+      "filename": "\u5f69\u8679\u7011\u5e03-Rainbow-Falls.jpg",
+      "path": "/workspace/minimind-v/minimind-v/dataset/eval_images/\u5f69\u8679\u7011\u5e03-Rainbow-Falls.jpg",
+      "description": "Image: \u5f69\u8679\u7011\u5e03-Rainbow-Falls.jpg"
+    },
+    {
+      "id": "\u5c0f\u72d7\u7f8e\u5973\u6d77\u8fb9-Dog-Woman-Sea",
+      "filename": "\u5c0f\u72d7\u7f8e\u5973\u6d77\u8fb9-Dog-Woman-Sea.jpg",
+      "path": "/workspace/minimind-v/minimind-v/dataset/eval_images/\u5c0f\u72d7\u7f8e\u5973\u6d77\u8fb9-Dog-Woman-Sea.jpg",
+      "description": "Image: \u5c0f\u72d7\u7f8e\u5973\u6d77\u8fb9-Dog-Woman-Sea.jpg"
+    },
+    {
+      "id": "\u592a\u7a7a\u5b87\u822a\u5458-Astronaut-Space",
+      "filename": "\u592a\u7a7a\u5b87\u822a\u5458-Astronaut-Space.jpg",
+      "path": "/workspace/minimind-v/minimind-v/dataset/eval_images/\u592a\u7a7a\u5b87\u822a\u5458-Astronaut-Space.jpg",
+      "description": "Image: \u592a\u7a7a\u5b87\u822a\u5458-Astronaut-Space.jpg"
+    },
+    {
+      "id": "\u57ce\u5e02\u8f66\u6c34\u9a6c\u9f99-city-traffic",
+      "filename": "\u57ce\u5e02\u8f66\u6c34\u9a6c\u9f99-city-traffic.jpg",
+      "path": "/workspace/minimind-v/minimind-v/dataset/eval_images/\u57ce\u5e02\u8f66\u6c34\u9a6c\u9f99-city-traffic.jpg",
+      "description": "Image: \u57ce\u5e02\u8f66\u6c34\u9a6c\u9f99-city-traffic.jpg"
+    }
+  ],
+  "dimension": 768,
+  "metric": "L2",
+  "count": 8
+}
\ No newline at end of file
diff --git a/scripts/build_knowledge_base.py b/scripts/build_knowledge_base.py
new file mode 100755
index 0000000..d3af62e
--- /dev/null
+++ b/scripts/build_knowledge_base.py
@@ -0,0 +1,129 @@
+#!/usr/bin/env python3
+"""
+Build RAG Knowledge Base from Images
+Creates FAISS index from image directory for retrieval-augmented generation
+"""
+import os
+import sys
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
+
+import argparse
+import json
+from pathlib import Path
+from tqdm import tqdm
+from PIL import Image
+from rag.retriever import CLIPRetriever
+
+
+def build_knowledge_base(
+    images_dir: str,
+    output_path: str,
+    description_file: str = None,
+    batch_size: int = 32
+):
+    """
+    Build knowledge base from image directory
+    
+    Args:
+        images_dir: Directory containing images
+        output_path: Where to save FAISS index
+        description_file: Optional JSON file with descriptions per image
+        batch_size: Batch size for encoding
+    """
+    print(f"\n{'='*60}")
+    print("🔨 Building Knowledge Base for RAG")
+    print(f"{'='*60}\n")
+    
+    # Find all images
+    image_extensions = {'.jpg', '.jpeg', '.png', '.bmp', '.webp'}
+    images_dir = Path(images_dir)
+    image_files = [
+        f for f in images_dir.iterdir()
+        if f.suffix.lower() in image_extensions
+    ]
+    
+    print(f"📂 Found {len(image_files)} images in {images_dir}")
+    
+    # Load descriptions if provided
+    descriptions = {}
+    if description_file and os.path.exists(description_file):
+        with open(description_file, 'r') as f:
+            descriptions = json.load(f)
+        print(f"📝 Loaded descriptions for {len(descriptions)} images")
+    
+    # Create metadata
+    metadata = []
+    image_paths = []
+    
+    for img_file in tqdm(image_files, desc="Preparing metadata"):
+        img_name = img_file.name
+        
+        # Create metadata entry
+        meta = {
+            'id': img_file.stem,
+            'filename': img_name,
+            'path': str(img_file.absolute()),
+            'description': descriptions.get(img_name, f"Image: {img_name}"),
+        }
+        
+        metadata.append(meta)
+        image_paths.append(str(img_file))
+    
+    # Create retriever
+    print("\n🔧 Initializing CLIP retriever...")
+    retriever = CLIPRetriever()
+    
+    # Build index
+    print(f"\n🚀 Encoding images in batches of {batch_size}...")
+    vector_store = retriever.build_index(
+        images=image_paths,
+        metadata=metadata,
+        save_path=output_path,
+        batch_size=batch_size
+    )
+    
+    print(f"\n{'='*60}")
+    print(f"✅ Knowledge Base Built Successfully!")
+    print(f"{'='*60}")
+    print(f"  Total images: {len(image_paths)}")
+    print(f"  Index file: {output_path}")
+    print(f"  Metadata file: {output_path.replace('.faiss', '.json')}")
+    print(f"  Embedding dimension: {vector_store.dimension}")
+    print(f"{'='*60}\n")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Build RAG knowledge base from images")
+    parser.add_argument(
+        "--images_dir",
+        type=str,
+        required=True,
+        help="Directory containing images"
+    )
+    parser.add_argument(
+        "--output",
+        type=str,
+        required=True,
+        help="Output path for FAISS index (.faiss)"
+    )
+    parser.add_argument(
+        "--descriptions",
+        type=str,
+        default=None,
+        help="Optional JSON file with image descriptions"
+    )
+    parser.add_argument(
+        "--batch_size",
+        type=int,
+        default=32,
+        help="Batch size for encoding"
+    )
+    
+    args = parser.parse_args()
+    
+    build_knowledge_base(
+        images_dir=args.images_dir,
+        output_path=args.output,
+        description_file=args.descriptions,
+        batch_size=args.batch_size
+    )
diff --git a/scripts/demo_rag.py b/scripts/demo_rag.py
new file mode 100755
index 0000000..8700e09
--- /dev/null
+++ b/scripts/demo_rag.py
@@ -0,0 +1,100 @@
+#!/usr/bin/env python3
+"""
+RAG System Demo
+Demonstrates retrieval-augmented generation with MiniMindVLM
+"""
+import os
+import sys
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
+
+import argparse
+from PIL import Image
+from rag.retriever import CLIPRetriever
+from rag.vector_store import VectorStore
+
+
+def demo_retrieval(query_image_path: str, index_path: str, k: int = 3):
+    """
+    Demo RAG retrieval
+    
+    Args:
+        query_image_path: Path to query image
+        index_path: Path to FAISS index
+        k: Number of results to retrieve
+    """
+    print(f"\n{'='*70}")
+    print("🔍 RAG Retrieval Demo")
+    print(f"{'='*70}\n")
+    
+    # Load query image
+    query_image = Image.open(query_image_path)
+    print(f"📷 Query image: {query_image_path}")
+    print(f"   Size: {query_image.size}")
+    print()
+    
+    # Load vector store
+    print(f"📂 Loading knowledge base from {index_path}...")
+    vector_store = VectorStore.load(index_path)
+    print()
+    
+    # Create retriever
+    print("🔧 Initializing CLIP retriever...")
+    retriever = CLIPRetriever()
+    print()
+    
+    # Retrieve similar images
+    print(f"🔍 Retrieving top-{k} similar images...")
+    distances, metadata = retriever.retrieve(query_image, vector_store, k=k)
+    print()
+    
+    # Display results
+    print(f"{'='*70}")
+    print(f"📊 Top-{k} Most Similar Images:")
+    print(f"{'='*70}\n")
+    
+    for i, (dist, meta) in enumerate(zip(distances, metadata), 1):
+        print(f"{i}. {meta['filename']}")
+        print(f"   Distance: {dist:.4f}")
+        print(f"   Description: {meta.get('description', 'No description')}")
+        print(f"   Path: {meta.get('path', 'Unknown')}")
+        print()
+    
+    print(f"{'='*70}\n")
+    
+    # Recommendation
+    if distances[0] < 5.0:
+        print("✅ Retrieved very similar images (distance < 5.0)")
+    elif distances[0] < 10.0:
+        print("⚠️  Retrieved moderately similar images (distance < 10.0)")
+    else:
+        print("❌ No very similar images found (distance >= 10.0)")
+    
+    print(f"\n{'='*70}\n")
+    
+    return distances, metadata
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Demo RAG retrieval")
+    parser.add_argument(
+        "--query",
+        type=str,
+        required=True,
+        help="Path to query image"
+    )
+    parser.add_argument(
+        "--index",
+        type=str,
+        default="rag_demo_index.faiss",
+        help="Path to FAISS index"
+    )
+    parser.add_argument(
+        "--k",
+        type=int,
+        default=3,
+        help="Number of results to retrieve"
+    )
+    
+    args = parser.parse_args()
+    
+    demo_retrieval(args.query, args.index, args.k)
diff --git a/trainer/lora_utils.py b/trainer/lora_utils.py
new file mode 100644
index 0000000..734c633
--- /dev/null
+++ b/trainer/lora_utils.py
@@ -0,0 +1,144 @@
+"""
+LoRA (Low-Rank Adaptation) utilities for MiniMindVLM
+Enables parameter-efficient fine-tuning with <2% trainable parameters
+"""
+import torch
+from peft import LoraConfig, get_peft_model, TaskType, PeftModel
+from typing import Optional
+
+
+def get_lora_config(
+    r: int = 16,
+    lora_alpha: int = 32,
+    lora_dropout: float = 0.05,
+    target_modules: Optional[list] = None,
+    bias: str = "none",
+) -> LoraConfig:
+    """
+    Get LoRA configuration for MiniMindVLM
+
+    Args:
+        r: Rank of LoRA update matrices (higher = more parameters)
+        lora_alpha: Scaling factor for LoRA updates
+        lora_dropout: Dropout rate for LoRA layers
+        target_modules: Which modules to apply LoRA to. If None, uses default for MiniMind
+        bias: How to handle bias parameters ("none", "all", "lora_only")
+
+    Returns:
+        LoraConfig object ready to apply to model
+
+    Typical memory savings:
+        - r=8:  ~0.5% trainable params, 99.5% frozen
+        - r=16: ~1.0% trainable params, 99.0% frozen  (recommended)
+        - r=32: ~2.0% trainable params, 98.0% frozen
+    """
+
+    # Default target modules for MiniMind architecture
+    # These are the attention projection layers
+    if target_modules is None:
+        target_modules = [
+            "c_attn",    # Combined QKV projection in attention
+            "c_proj",    # Output projection in attention
+            "c_fc",      # First linear in MLP (optional, but helps)
+        ]
+
+    config = LoraConfig(
+        task_type=TaskType.CAUSAL_LM,
+        inference_mode=False,
+        r=r,
+        lora_alpha=lora_alpha,
+        lora_dropout=lora_dropout,
+        target_modules=target_modules,
+        bias=bias,
+    )
+
+    return config
+
+
+def apply_lora(model, lora_config: LoraConfig, verbose: bool = True):
+    """
+    Apply LoRA adapters to a MiniMindVLM model
+
+    Args:
+        model: MiniMindVLM model instance
+        lora_config: LoRA configuration from get_lora_config()
+        verbose: Print trainable parameter statistics
+
+    Returns:
+        Model with LoRA adapters applied
+    """
+    # Freeze vision encoder (it's already frozen in MiniMindVLM, but be explicit)
+    if hasattr(model, 'vision_encoder') and model.vision_encoder is not None:
+        for param in model.vision_encoder.parameters():
+            param.requires_grad = False
+
+    # Apply LoRA
+    model = get_peft_model(model, lora_config)
+
+    if verbose:
+        model.print_trainable_parameters()
+
+        # Calculate memory savings
+        trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
+        total = sum(p.numel() for p in model.parameters())
+
+        print(f"\n📊 LoRA Statistics:")
+        print(f"  Total parameters: {total / 1e6:.2f}M")
+        print(f"  Trainable parameters: {trainable / 1e6:.2f}M ({trainable / total * 100:.2f}%)")
+        print(f"  Frozen parameters: {(total - trainable) / 1e6:.2f}M ({(total - trainable) / total * 100:.2f}%)")
+        print(f"  Memory savings: ~{100 - (trainable / total * 100):.1f}%")
+
+    return model
+
+
+def load_lora_model(base_model, lora_adapter_path: str):
+    """
+    Load a model with LoRA adapters
+
+    Args:
+        base_model: Base MiniMindVLM model (without LoRA)
+        lora_adapter_path: Path to saved LoRA adapters
+
+    Returns:
+        Model with LoRA adapters loaded
+    """
+    model = PeftModel.from_pretrained(base_model, lora_adapter_path)
+    return model
+
+
+def merge_lora_weights(model, save_path: Optional[str] = None):
+    """
+    Merge LoRA weights into base model for deployment
+    After merging, the model is a standard model without adapters
+
+    Args:
+        model: Model with LoRA adapters
+        save_path: If provided, save merged model to this path
+
+    Returns:
+        Merged model (standard model, no adapters)
+    """
+    # Merge LoRA weights into base model
+    model = model.merge_and_unload()
+
+    if save_path:
+        model.save_pretrained(save_path)
+        print(f"✅ Merged model saved to {save_path}")
+
+    return model
+
+
+# Example usage:
+if __name__ == "__main__":
+    from model.model_vlm import MiniMindVLM, VLMConfig
+
+    # Create model
+    config = VLMConfig()
+    model = MiniMindVLM(config)
+
+    # Apply LoRA
+    lora_config = get_lora_config(r=16)
+    model = apply_lora(model, lora_config)
+
+    # Now model is ready for training with LoRA
+    # Only ~1% of parameters will be updated during training
diff --git a/trainer/train_sft_vlm_lora.py b/trainer/train_sft_vlm_lora.py
new file mode 100644
index 0000000..57707e7
--- /dev/null
+++ b/trainer/train_sft_vlm_lora.py
@@ -0,0 +1,232 @@
+"""
+SFT VLM Training with LoRA Support
+Extension of train_sft_vlm.py with LoRA for parameter-efficient fine-tuning
+"""
+import os
+import sys
+
+__package__ = "trainer"
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
+
+import argparse
+import time
+import warnings
+import torch
+import torch.distributed as dist
+from contextlib import nullcontext
+from torch import optim, nn
+from torch.nn.parallel import DistributedDataParallel
+from torch.utils.data import DataLoader, DistributedSampler
+from transformers import AutoTokenizer
+from model.model_vlm import MiniMindVLM, VLMConfig
+from dataset.lm_dataset import VLMDataset
+from trainer.trainer_utils import get_lr, Logger, is_main_process, init_distributed_mode, setup_seed, init_vlm_model, vlm_checkpoint, SkipBatchSampler
+from trainer.lora_utils import get_lora_config, apply_lora
+
+warnings.filterwarnings('ignore')
+
+
+def train_epoch(epoch, loader, iters, start_step=0, wandb=None):
+    start_time = time.time()
+    for step, (input_ids, labels, pixel_values) in enumerate(loader, start=start_step + 1):
+        input_ids = input_ids.to(args.device)
+        labels = labels.to(args.device)
+        pixel_values = pixel_values.to(args.device)
+        lr = get_lr(epoch * iters + step, args.epochs * iters, args.learning_rate)
+        for param_group in optimizer.param_groups:
+            param_group['lr'] = lr
+
+        with autocast_ctx:
+            res = model(input_ids, labels=labels, pixel_values=pixel_values)
+            loss = res.loss + res.aux_loss
+            loss = loss / args.accumulation_steps
+
+        scaler.scale(loss).backward()
+
+        if (step + 1) % args.accumulation_steps == 0:
+            scaler.unscale_(optimizer)
+            torch.nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
+
+            scaler.step(optimizer)
+            scaler.update()
+
+            optimizer.zero_grad(set_to_none=True)
+
+        if step % args.log_interval == 0 or step == iters - 1:
+            spend_time = time.time() - start_time
+            logger.log(
+                'Epoch:[{}/{}]({}/{}) loss:{:.3f} lr:{:.8f} epoch_time:{}min:'.format(
+                    epoch + 1,
+                    args.epochs,
+                    step,
+                    iters,
+                    loss.item() * args.accumulation_steps,
+                    optimizer.param_groups[-1]['lr'],
+                    spend_time / (step + 1) * iters // 60 - spend_time // 60))
+
+            if (wandb is not None) and is_main_process():
+                wandb.log({"loss": loss.item() * args.accumulation_steps,
+                           "lr": optimizer.param_groups[-1]['lr'],
+                           "epoch": epoch + 1,
+                           "step": step})
+
+        if step % args.save_interval == 0 or step == iters - 1:
+            if is_main_process():
+                vlm_checkpoint(model.module if hasattr(model, 'module') else model,
+                             epoch,
+                             step,
+                             args,
+                             suffix=f"_lora" if args.use_lora else "")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="MiniMind VLM SFT with LoRA")
+    parser.add_argument("--out_dir", type=str, default="out", help="Output directory")
+    parser.add_argument("--epochs", type=int, default=1, help="Number of epochs")
+    parser.add_argument("--batch_size", type=int, default=64, help="Batch size")
+    parser.add_argument("--learning_rate", type=float, default=1e-4, help="Learning rate")
+    parser.add_argument("--device", type=str, default="cuda:0" if torch.cuda.is_available() else "cpu")
+    parser.add_argument("--dtype", type=str, default="bfloat16" if torch.cuda.is_bf16_supported() else "float16")
+    parser.add_argument("--use_wandb", action="store_true", help="Use Weights & Biases")
+    parser.add_argument("--num_workers", type=int, default=0, help="Data loader workers")
+    parser.add_argument("--accumulation_steps", type=int, default=1, help="Gradient accumulation")
+    parser.add_argument("--grad_clip", type=float, default=1.0, help="Gradient clipping")
+    parser.add_argument("--warmup_iters", type=int, default=0, help="Warmup iterations")
+    parser.add_argument("--log_interval", type=int, default=100, help="Log interval")
+    parser.add_argument("--save_interval", type=int, default=1000, help="Save interval")
+    parser.add_argument("--local_rank", type=int, default=-1, help="DDP local rank")
+    parser.add_argument("--use_moe", action="store_true", help="Use Mixture of Experts")
+    parser.add_argument("--resume_step", type=int, default=0, help="Resume from step")
+    parser.add_argument("--hidden_size", type=int, default=512, help="Hidden size")
+    parser.add_argument("--num_heads", type=int, default=16, help="Number of attention heads")
+    parser.add_argument("--num_layers", type=int, default=8, help="Number of layers")
+
+    # LoRA-specific arguments
+    parser.add_argument("--use_lora", action="store_true", help="Use LoRA for parameter-efficient training")
+    parser.add_argument("--lora_r", type=int, default=16, help="LoRA rank")
+    parser.add_argument("--lora_alpha", type=int, default=32, help="LoRA alpha")
+    parser.add_argument("--lora_dropout", type=float, default=0.05, help="LoRA dropout")
+
+    args = parser.parse_args()
+
+    # Setup
+    init_distributed_mode(args)
+    setup_seed(42)
+    logger = Logger(args.out_dir, is_main_process())
+
+    # Load tokenizer and model
+    tokenizer = AutoTokenizer.from_pretrained('./model')
+
+    # Create VLM config
+    lm_config = VLMConfig(
+        hidden_size=args.hidden_size,
+        num_attention_heads=args.num_heads,
+        num_hidden_layers=args.num_layers,
+        use_moe=args.use_moe,
+    )
+
+    # Initialize model
+    model = init_vlm_model(lm_config, args.device)
+
+    # Apply LoRA if requested
+    if args.use_lora:
+        print("\n" + "=" * 50)
+        print("🚀 Applying LoRA for Parameter-Efficient Training")
+        print("=" * 50)
+
+        lora_config = get_lora_config(
+            r=args.lora_r,
+            lora_alpha=args.lora_alpha,
+            lora_dropout=args.lora_dropout,
+        )
+
+        model = apply_lora(model, lora_config, verbose=True)
+        print("=" * 50 + "\n")
+
+    # DDP if needed
+    if args.ddp:
+        model = DistributedDataParallel(model, device_ids=[args.local_rank])
+
+    # Optimizer
+    optimizer = optim.Adam(
+        filter(lambda p: p.requires_grad, model.parameters()),  # Only train LoRA params if using LoRA
+        lr=args.learning_rate
+    )
+
+    # AMP scaler
+    scaler = torch.cuda.amp.GradScaler(enabled=(args.dtype in ['float16', 'bfloat16']))
+    autocast_ctx = (
+        torch.cuda.amp.autocast(dtype=torch.bfloat16)
+        if args.dtype == 'bfloat16'
+        else torch.cuda.amp.autocast(dtype=torch.float16)
+        if args.dtype == 'float16'
+        else nullcontext()
+    )
+
+    # Load dataset
+    dataset = VLMDataset(
+        parquet_path='./dataset/sft_vlm.parquet',
+        tokenizer=tokenizer,
+        preprocess=model.processor if not args.ddp else model.module.processor,
+        max_length=512
+    )
+
+    # Data loader
+    if args.ddp:
+        sampler = DistributedSampler(dataset)
+        loader = DataLoader(
+            dataset,
+            batch_size=args.batch_size,
+            sampler=sampler,
+            num_workers=args.num_workers,
+            pin_memory=True
+        )
+    else:
+        loader = DataLoader(
+            dataset,
+            batch_size=args.batch_size,
+            shuffle=True,
+            num_workers=args.num_workers,
+            pin_memory=True
+        )
+
+    # WandB
+    wandb = None
+    if args.use_wandb and is_main_process():
+        import wandb as wandb_lib
+        wandb = wandb_lib
+        wandb.init(
+            project="drivemind-v",
+            name=f"sft_lora_r{args.lora_r}" if args.use_lora else "sft_full",
+            config=vars(args)
+        )
+
+    # Training loop
+    iters = len(loader)
+    logger.log(f"Starting SFT training {'with LoRA' if args.use_lora else 'full fine-tuning'}...")
+    logger.log(f"Total iterations: {iters} per epoch")
+
+    for epoch in range(args.epochs):
+        train_epoch(epoch, loader, iters, start_step=0, wandb=wandb)
+
+    # Save final model
+    if is_main_process():
+        if args.use_lora:
+            # Save LoRA adapters (tiny file ~2-5MB)
+            save_path = f"{args.out_dir}/lora_adapters_sft_vlm_{args.hidden_size}"
+            model_to_save = model.module if hasattr(model, 'module') else model
+            model_to_save.save_pretrained(save_path)
+            logger.log(f"✅ LoRA adapters saved to {save_path}")
+        else:
+            vlm_checkpoint(
+                model.module if hasattr(model, 'module') else model,
+                args.epochs - 1,
+                iters - 1,
+                args,
+                suffix="_final"
+            )
+
+    if wandb:
+        wandb.finish()
+
+    logger.log("Training complete!")