Retrieval-augmented generation (RAG) pipelines have become the de-facto approach for building AI assistants with external knowledge. Given a user query, RAG pipelines retrieve (R) information from external sources, before invoking a Large Language Model (LLM), augmented (A) with this information, to generate (G) responses. However, developing effective RAG pipelines is challenging because retrieval and generation components—often chained in varying orders—are intertwined, making it hard to identify which component(s) cause errors in the output. Developers often need to answer “what-if” questions—e.g., what if chunk sizes were larger or retrieval used embeddings versus keywords—but such experimentation requires hours of re-processing. We present raggy, a developer tool that enables rapid “what-if” analysis by combining a Python library of composable RAG primitives with an interactive debugging interface. We contribute the design and implementation of raggy, insights into expert debugging patterns through a qualitative study with 12 engineers, and design implications for RAG tools.

Querying and extracting value from unstructured document collection remains a considerable challenge. While Large Language Models (LLMs) have made remarkable progress in document understanding, they fail to give high accuracy results for analytical queries on documents, and additionally incur high costs. While Retrieval-Augmented Generation (RAG) can reduce costs, accuracy degrades further. Our key insight is that documents in a collection often follow similar templates that impart a common semantic structure. We therefore introduce Zendb, a document analytics system that leverages this semantic structure, coupled with LLMs, to answer ad-hoc SQL queries on document collections. Zendb efficiently extracts semantic hierarchical structures from such templatized documents and introduces a novel query engine that leverages these structures for accurate and cost-effective query execution. Extensive experiments on three real-world document collections demonstrate ZENDB's benefits, achieving up to 31× cost savings compared to LLM-based baselines, while maintaining or improving accuracy, and surpassing RAG-based baselines by up to 61% in precision and 81% in recall, at a marginally higher cost.

Reya Vir, Shreya Shankar, Harrison Chase, William Hinthorn, Aditya Parameswaran. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

intent is hard to communicat  Prompts need to be detailed and dataspecific  LLM behavior varies across document  Requires fine-grained decomposition A B C Docs & Outputs condition discomfort_level symptoms

Large Language Models (LLMs) are being increasingly used as a building block in data systems to process large text datasets. To do so, LLM model providers offer multiple LLMs with different sizes, spanning various cost-quality trade-offs when processing text at scale. Top-of-the-line LLMs (e.g., GPT-4o, Claude Sonnet) operate with high accuracy but are prohibitively expensive when processing many records. To avoid high costs, more affordable but lower quality LLMs (e.g., GPT-4o-mini, Claude Haiku) can be used to process records, but we need to ensure that the overall accuracy does not deviate substantially from that of the top-of-the-line LLMs. The model cascade framework provides a blueprint to manage this trade-off, by using the confidence of LLMs in their output (e.g., log-probabilities) to decide on which records to use the affordable LLM. However, existing solutions following this framework provide only marginal cost savings and weak theoretical guarantees because of poor estimation of the quality of the affordable LLM's outputs. We present BARGAIN, a method that judiciously uses affordable LLMs in data processing to significantly reduce cost while providing strong theoretical guarantees on the solution quality. BARGAIN employs a novel adaptive sampling strategy and statistical estimation procedure that uses data and task characteristics and builds on recent statistical tools to make accurate estimations with tight theoretical guarantees. Variants of BARGAIN can support guarantees on accuracy, precision, or recall of the output. Experimental results across 8 real-world datasets show that BARGAIN reduces cost, on average, by up to 86% more than state-of-the-art, while providing stronger theoretical guarantees on accuracy of output, with similar gains when guaranteeing a desired level of precision or recall.

Analyzing unstructured data has been a persistent challenge in data processing. Recent proposals offer declarative frameworks for LLM-powered processing of unstructured data, but they typically execute user-specified operations as-is in a single LLM call—focusing on cost rather than accuracy. This is problematic for complex tasks, where even well-prompted LLMs can miss relevant information. For instance, reliably extracting all instances of a specific clause from legal documents often requires decomposing the task, the data, or both. We present DocETL, a system that optimizes complex document processing pipelines, while accounting for LLM shortcomings. DocETL offers a declarative interface for users to deine such pipelines and uses an agent-based approach to automatically optimize them, leveraging novel agent-based rewrites (that we call rewrite directives ), as well as an optimization and evaluation framework. We introduce (i) logical rewriting of pipelines, tailored for LLM-based tasks, (ii) an agent-guided plan evaluation mechanism, and (iii) an optimization algorithm that efficiently finds promising plans, considering the latencies of LLM execution. Across four real-world document processing tasks, DocETL improves accuracy by 21–80% over strong baselines. DocETL is open-source at docetl.org and, as of March 2025, has over 1.7k GitHub stars across diverse domains.

Due to the cumbersome nature of human evaluation and limitations of code-based evaluation, Large Language Models (LLMs) are increasingly being used to assist humans in evaluating LLM outputs. Yet LLM-generated evaluators simply inherit all the problems of the LLMs they evaluate, requiring further human validation. We present a mixed-initiative approach to “validate the validators”—aligning LLM-generated evaluation functions (be it prompts or code) with human requirements. Our interface, EvalGen, provides automated assistance to users in generating evaluation criteria and implementing assertions. While generating candidate implementations (Python functions, LLM grader prompts), EvalGen asks humans to grade a subset of LLM outputs; this feedback is used to select implementations that better align with user grades. A qualitative study finds overall support for EvalGen but underscores the subjectivity and iterative nature of alignment. In particular, we identify a phenomenon we dub criteria drift: users need criteria to grade outputs, but grading outputs helps users define criteria. What is more, some criteria appear dependent on the specific LLM outputs observed (rather than independent and definable a priori), raising serious questions for approaches that assume the independence of evaluation from observation of model outputs. We present our interface and implementation details, a comparison of our algorithm with a baseline approach, and implications for the design of future LLM evaluation assistants.

Organizations rely on machine learning engineers (MLEs) to deploy models and maintain ML pipelines in production. Due to models' extensive reliance on fresh data, the operationalization of machine learning, or MLOps, requires MLEs to have proficiency in data science and engineering. When considered holistically, the job seems staggering---how do MLEs do MLOps, and what are their unaddressed challenges? To address these questions, we conducted semi-structured ethnographic interviews with 18 MLEs working on various applications, including chatbots, autonomous vehicles, and finance. We find that MLEs engage in a workflow of (i) data preparation, (ii) experimentation, (iii) evaluation throughout a multi-staged deployment, and (iv) continual monitoring and response. Throughout this workflow, MLEs collaborate extensively with data scientists, product stakeholders, and one another, supplementing routine verbal exchanges with communication tools ranging from Slack to organization-wide ticketing and reporting systems. We introduce the 3Vs of MLOps: velocity, visibility, and versioning --- three virtues of successful ML deployments that MLEs learn to balance and grow as they mature. Finally, we discuss design implications and opportunities for future work.

Large language models (LLMs) are being increasingly deployed as part of pipelines that repeatedly process or generate data of some sort. However, a common barrier to deployment are the frequent and often unpredictable errors that plague LLMs. Acknowledging the inevitability of these errors, we propose data quality assertions to identify when LLMs may be making mistakes. We present spade, a method for automatically synthesizing data quality assertions that identify bad LLM outputs. We make the observation that developers often identify data quality issues during prototyping prior to deployment, and attempt to address them by adding instructions to the LLM prompt over time. spade therefore analyzes histories of prompt versions over time to create candidate assertion functions and then selects a minimal set that fulfills both coverage and accuracy requirements. In testing across nine different real-world LLM pipelines, spade efficiently reduces the number of assertions by 14% and decreases false failures by 21% when compared to simpler baselines. spade has been deployed as an offering within LangSmith, LangChain's LLM pipeline hub, and has been used to generate data quality assertions for over 2000 pipelines across a spectrum of industries.

Machine learning (ML) models in production pipelines are frequently retrained on the latest partitions of large, continually- growing datasets. Due to engineering bugs, partitions in such datasets almost always have some corrupted features; thus, it's critical to find data issues and block retraining before downstream ML accuracy decreases. However, current ML data validation methods are difficult to operationalize: they yield too many false positive alerts, require manual tuning, or are infeasible at scale. In this pa- per, we present an automatic, precise, and scalable data validation system for ML pipelines, employing a simple idea that we call a Partition Summarization (PS) approach to data validation: each timestamp-based partition of data is summarized with data quality metrics, and summaries are compared to detect corrupted partitions. We demonstrate how to adapt PS for any data validation method in a robust manner and evaluate several adaptations-which by themselves provide limited precision. Finally, we present gate, our data validation method that leverages these adaptations, giving a 2.1× average improvement in precision over the baseline from prior work on a case study within our large tech company.

Computational notebooks are commonly used for iterative workflows, such as in exploratory data analysis. This process lends itself to the accumulation of old code and hidden state, making it hard for users to reason about the lineage of, e.g., plots depicting insights or trained machine learning models. One way to reason about code used to generate various notebook data artifacts is to compute a program slice , but traditional static approaches to slicing can be both inaccurate (failing to contain relevant code for artifacts) and conservative (containing unnecessary code for an artifacts). We present nbslicer, a dynamic slicer optimized for the notebook setting whose instrumentation for resolving dynamic data dependencies is both bolt-on (and therefore portable) and switchable (allowing it to be selectively disabled in order to reduce instrumentation overhead). We demonstrate Nbslicer's ability to construct small and accurate backward slices (i.e., historical cell dependencies) and forward slices (i.e., cells affected by the "rerun" of an earlier cell), thereby improving reproducibility in notebooks and enabling faster reactive re-execution, respectively. Comparing nbslicer with a static slicer on 374 real notebook sessions, we found that nbslicer filters out far more superfluous program statements while maintaining slice correctness, giving slices that are, on average, 66% and 54% smaller for backward and forward slices, respectively.

Software organizations are increasingly incorporating machine learning (ML) into their product offerings, driving a need for new data management tools. Many of these tools facilitate the initial development of ML applications, but sustaining these applications post-deployment is difficult due to lack of real-time feedback (i.e., labels) for predictions and silent failures that could occur at any component of the ML pipeline (e.g., data distribution shift or anomalous features). We propose a new type of data management system that offers end-to-end observability , or visibility into complex system behavior, for deployed ML pipelines through assisted (1) detection, (2) diagnosis, and (3) reaction to ML-related bugs. We describe new research challenges and suggest preliminary solution ideas in all three aspects. Finally, we introduce an example architecture for a "bolt-on" ML observability system, or one that wraps around existing tools in the stack.

Convex relaxations have emerged as a promising approach for verifying properties of neural networks, but widely used using Linear Programming (LP) relaxations only provide meaningful certificates when networks are specifically trained to facilitate verification. This precludes many important applications which involve \emph{verification-agnostic} networks that are not trained specifically to promote verifiability. On the other hand, semidefinite programming (SDP) relaxations have shown success on verification-agnostic networks, such as adversarially trained image classifiers without additional regularization, but do not currently scale beyond small networks due to poor time and space asymptotics. In this work, we propose a first-order dual SDP algorithm that provides (1) any-time bounds (2) requires memory only linear in the total number of network activations and (3) has per-iteration complexity that scales linearly with the complexity of a forward/backward pass through the network. By exploiting iterative eigenvector methods, we express all solver operations in terms of forward and backward passes through the network, enabling efficient use of hardware optimized for deep learning. This allows us to dramatically improve the magnitude of $\ell_\infty$ perturbations for which we can verify robustness verification-agnostic networks ($1\% \to 88\%$ on MNIST, $6\%\to 40\%$ on CIFAR-10). We also demonstrate tight verification for a quadratic stability specification for the decoder of a variational autoencoder.

Machine learning models are vulnerable to adversarial examples: small changes to images can cause computer vision models to make mistakes such as identifying a school bus as an ostrich. However, it is still an open question whether humans are prone to similar mistakes. Here, we address this question by leveraging recent techniques that transfer adversarial examples from computer vision models with known parameters and architecture to other models with unknown parameters and architecture, and by matching the initial processing of the human visual system. We find that adversarial examples that strongly transfer across computer vision models influence the classifications made by time-limited human observers.