Large Language Models (LLMs) have demonstrated impressive ability in generation and reasoning tasks but struggle with handling up-to-date knowledge, leading to inaccuracies or hallucinations. Retrieval-Augmented Generation (RAG) mitigates this by retrieving and incorporating external knowledge into input prompts. In particular, due to LLMs' context window limitations and long-context hallucinations, only the most relevant "chunks" are retrieved. However, current RAG systems face three key challenges: (1) chunks are often retrieved independently without considering their relationships, such as redundancy and ordering; (2) the utility of chunks is non-monotonic, as adding more chunks can degrade quality; and (3) retrieval strategies fail to adapt to the unique characteristics of different queries. To overcome these challenges, we design a cost-constrained retrieval optimization framework for RAG. We adopt a Monte Carlo Tree Search (MCTS) based strategy to find the optimal chunk combination order, which considers the chunks' correlations. In addition, to address the non-monotonicity of chunk utility, instead of treating budget exhaustion as the termination condition, we design a utility computation strategy to identify the optimal chunk combination without necessarily exhausting the budget. Furthermore, we propose a configuration agent that predicts optimal configurations for each query domain, improving our framework's adaptability and efficiency. Experimental results demonstrate up to a 30

In this paper, we propose two techniques for near-duplicate image detection at high confidence and large scale. First, we show that entropy-based filtering eliminates ambiguous SIFT features that cause most of the false positives, and enables claiming near-duplicity with a single match of the retained high-quality features. Second, we show that graph cut can be used for query expansion with a duplicity graph computed offline to substantially improve search quality. Evaluation with web images show that when combined with sketch embedding [6], our methods achieve false positive rate orders of magnitude lower than the standard visual word approach. We demonstrate the proposed techniques with a large-scale image search engine which, using indexing data structure offline computed with a Hadoop cluster, is capable of serving more than 50 million web images with a single commodity server.

K-Nearest Neighbor Graph (K-NNG) construction is an important operation with many web related applications, including collaborative filtering, similarity search, and many others in data mining and machine learning. Existing methods for K-NNG construction either do not scale, or are specific to certain similarity measures. We present NN-Descent, a simple yet efficient algorithm for approximate K-NNG construction with arbitrary similarity measures. Our method is based on local search, has minimal space overhead and does not rely on any shared global index. Hence, it is especially suitable for large-scale applications where data structures need to be distributed over the network. We have shown with a variety of datasets and similarity measures that the proposed method typically converges to above 90% recall with each point comparing only to several percent of the whole dataset on average.

As data have been growing rapidly in data centers, deduplication storage systems continuously face challenges in providing the corresponding throughputs and capacities necessary to move backup data within backup and recovery window times. One approach is to build a cluster deduplication storage system with multiple deduplication storage system nodes. The goal is to achieve scalable throughput and capacity using extremely high-throughput (e.g. 1.5 GB/s) nodes, with a minimal loss of compression ratio. The key technical issue is to route data intelligently at an appropriate granularity. We present a cluster-based deduplication system that can deduplicate with high throughput, support deduplication ratios comparable to that of a single system, and maintain a low variation in the storage utilization of individual nodes. In experiments with dozens of nodes, we examine tradeoffs between stateless data routing approaches with low overhead and stateful approaches that have higher overhead but avoid imbalances that can adversely affect deduplication effectiveness for some datasets in large clusters. The stateless approach has been deployed in a two-node commercial system that achieves 3 GB/s for multi-stream deduplication throughput and currently scales to 5.6 PB of storage (assuming 20X total compression).

The volume of images and other non-text data is growing exponentially in today's digital universe. A popular way of extracting useful information from such data is to conduct content-based similarity search, e.g., to search for images with content similar to a query image. How to build data structures to support efficient similarity search on a large scale is an issue of increasing importance. The challenge is that feature-rich data are usually represented as high-dimensional feature vectors, and the curse of dimensionality dictates that as dimensionality grows, any search strategy examines an increasingly large portion of the dataset and eventually degenerates to brute-force scan. In this dissertation, we study several key issues to improve the accuracy and efficiency of high-dimensional similarity search. Specifically, we make the following contributions: First, we enhance Multi-Probe LSH, the state-of-the-art high-dimensional similarity search technique, by developing an accurate performance model that predicts search accuracy with an error rate of less than 5% using only a 1/10 sample of the whole dataset. We use this model to realize automatic performance tuning, which would otherwise be tedious, and to develop an adaptive query processing strategy to ensure the high search quality of each individual query. Second, we develop an efficient offline method for simultaneously finding the K nearest neighbors of every point in the dataset. Our method is 2 to 16 times faster than the state-of-the-art technique when evaluated with different datasets. Furthermore, we show how to use the offline computed nearest-neighbor information to double the speed of online similarity search. Third, we develop compact representations of high-dimensional feature vectors optimized for similarity search tasks. We present an asymmetric distance estimation framework to exploit the information in the uncompressed query data. We use that to further compress the indexed dataset, by 10% to 40%, without compromising search quality. Fourth, we develop a scheme to compactly represent sets of feature vectors, an increasingly popular data representation that is more accurate than single vectors, but also more expensive. Our method dramatically reduces the matching cost in both space and time. In an image classification task, our method achieves 25 times speedup over one of the best existing techniques. Finally, we demonstrate some of the proposed techniques by building a large-scale near-duplicate image search engine. Our system serves more than 50 million web images on a single commodity server and responds to queries within a few seconds.

The explosion of image data on the Internet has the potential to foster more sophisticated and robust models and algorithms to index, retrieve, organize and interact with images and multimedia data. But exactly how such data can be harnessed and organized remains a critical problem. We introduce here a new database called “ImageNet”, a large-scale ontology of images built upon the backbone of the WordNet structure. ImageNet aims to populate the majority of the 80,000 synsets of WordNet with an average of 500-1000 clean and full resolution images. This will result in tens of millions of annotated images organized by the semantic hierarchy of WordNet. This paper offers a detailed analysis of ImageNet in its current state: 12 subtrees with 5247 synsets and 3.2 million images in total. We show that ImageNet is much larger in scale and diversity and much more accurate than the current image datasets. Constructing such a large-scale database is a challenging task. We describe the data collection scheme with Amazon Mechanical Turk. Lastly, we illustrate the usefulness of ImageNet through three simple applications in object recognition, image classification and automatic object clustering. We hope that the scale, accuracy, diversity and hierarchical structure of ImageNet can offer unparalleled opportunities to researchers in the computer vision community and beyond.

As the commonly used representation of a feature-rich data object has evolved from a single feature vector to a set of feature vectors, a key challenge in building a content-based search engine for feature-rich data is to match feature-sets efficiently. Although substantial progress has been made during the past few years, existing approaches are still inefficient and inflexible for building a search engine for massive datasets. This paper presents a randomized algorithm to embed a set of features into a single high-dimensional vector to simplify the feature-set matching problem. The main idea is to project feature vectors into an auxiliary space using locality sensitive hashing and to represent a set of features as a histogram in the auxiliary space. A histogram is simply a high dimensional vector, and efficient similarity measures like L 1 and L 2 distances can be employed to approximate feature-set distance measures. We evaluated the proposed approach under three different task settings, i.e. content-based image search, image object recognition and near-duplicate video clip detection. The experimental results show that the proposed approach is indeed effective and flexible. It can achieve accuracy comparable to the feature-set matching methods, while requiring significantly less space and time. For object recognition with Caltech 101 dataset, our method runs 25 times faster to achieve the same precision as Pyramid Matching Kernel, the state-of-the-art feature-set matching method.

Efficient similarity search in high-dimensional spaces is important to content-based retrieval systems. Recent studies have shown that sketches can effectively approximate L 1 distance in high-dimensional spaces, and that filtering with sketches can speed up similarity search by an order of magnitude. It is a challenge to further reduce the size of sketches, which are already compact, without compromising accuracy of distance estimation. This paper presents an efficient sketch algorithm for similarity search with L 2 distances and a novel asymmetric distance estimation technique. Our new asymmetric estimator takes advantage of the original feature vector of the query to boost the distance estimation accuracy. We also apply this asymmetric method to existing sketches for cosine similarity and L 1 distance. Evaluations with datasets extracted from images and telephone records show that our L 2 sketch outperforms existing methods, and the asymmetric estimators consistently improve the accuracy of different sketch methods. To achieve the same search quality, asymmetric estimators can reduce the sketch size by 10% to 40%.

Although Locality-Sensitive Hashing (LSH) is a promising approach to similarity search in high-dimensional spaces, it has not been considered practical partly because its search quality is sensitive to several parameters that are quite data dependent. Previous research on LSH, though obtained interesting asymptotic results, provides little guidance on how these parameters should be chosen, and tuning parameters for a given dataset remains a tedious process. To address this problem, we present a statistical performance model of Multi-probe LSH, a state-of-the-art variance of LSH. Our model can accurately predict the average search quality and latency given a small sample dataset. Apart from automatic parameter tuning with the performance model, we also use the model to devise an adaptive LSH search algorithm to determine the probing parameter dynamically for each query. The adaptive probing method addresses the problem that even though the average performance is tuned for optimal, the variance of the performance is extremely high. We experimented with three different datasets including audio, images and 3D shapes to evaluate our methods. The results show the accuracy of the proposed model: the recall errors predicted are within 5% from the real values for most cases; the adaptive search method reduces the standard deviation of recall by about 50% over the existing method.

Sketches are compact data structures that can be used to estimate properties of the original data in building large-scale search engines and data analysis systems. Recent theoretical and experimental studies have shown that sketches constructed from feature vectors using randomized projections can effectively approximate L1 distance on the feature vectors with the Hamming distance on their sketches. Furthermore, such sketches can achieve good filtering accuracy while reducing the metadata space requirement and speeding up similarity searches by an order of magnitude. However, it is not clear how to choose the size of the sketches since it depends ondata type, dataset size, and desired filtering quality. In real systems designs, it is necessary to understand how to choose sketch size without the dataset, or at least without the whole datase. This paper presents an analytical model and experimental results to help system designers make such design decisions. We present arank-based filtering model that describes the relationship between sketch size and data set size based on the dataset distance distribution. Our experimental results with several datasets including images, audio, and 3D shapes show that the model yields good, conservative predictions. We show that the parameters of the model can be set with a small sample data set and the resulting model can make good predictions for a large dataset. We illustrate how to apply the approach with a concrete example.