Enterprise AI agents must continuously adapt to maintain accuracy, reduce latency, and remain aligned with user needs. We present a practical implementation of a data flywheel in NVInfo AI, NVIDIA’s Mixture-of-Experts (MoE) Knowledge Assistant serving over 30,000 employees. By operationalizing a MAPE-driven data flywheel, we built a closed-loop system that systematically addresses failures in retrieval-augmented generation (RAG) pipelines and enables continuous learning.Over a 3-month post-deployment period, we monitored feedback and collected 495 negative samples. Analysis revealed two major failure modes: routing errors (5.25%) and query rephrasal errors (3.2%). Using NVIDIA NeMo Microservices, we implemented targeted improvements through fine-tuning. For routing, we replaced a Llama 3.1 70B model with a fine-tuned 8B variant, achieving 96% accuracy, a 10× reduction in model size, and 70% latency improvement. For query rephrasal, fine-tuning yielded a 3.7% gain in accuracy and a 40% latency reduction.Our approach demonstrates how human-in-the-loop (HITL) feedback, when structured within a data flywheel, transforms enterprise AI agents into self-improving systems. Key learnings include approaches to ensure agent robustness despite limited user feedback, navigating privacy constraints, and executing staged rollouts in production. This work offers a repeatable blueprint for building robust, adaptive enterprise AI agents capable of learning from real-world usage at scale.

Retrieval-augmented generation (RAG) offers a robust solution for developing enterprise internal virtual assistants by leveraging domain-specific knowledge and utilizing information from frequently updated corporate document repositories. In this work, we introduce the Enterprise-Knowledge RAG (EKRAG) dataset to benchmark RAG for enterprise knowledge question-answering (QA) across a diverse range of corporate documents, such as product releases, technical blogs, and financial reports. Using EKRAG, we systematically evaluate various retrieval models and strategies tailored for corporate content. We propose novel embedding-model (EM)-as-judge and ranking-model (RM)-as-judge approaches to assess answer quality in the context of enterprise information. Combining these with the existing LLM-as-judge method, we then comprehensively evaluate the correctness, relevance, and faithfulness of generated answers to corporate queries. Our extensive experiments shed light on optimizing RAG pipelines for enterprise knowledge QA, providing valuable guidance for practitioners. This work contributes to enhancing information retrieval and question-answering capabilities in corporate environments that demand high degrees of factuality and context-awareness.

Existing multilingual video corpus moment retrieval (mVCMR) methods are mainly based on a two-stream structure. The visual stream utilizes the visual content in the video to estimate the query-visual similarity, and the subtitle stream exploits the query-subtitle similarity. The final query-video similarity ensembles similarities from two streams. In our work, we pro- pose a simple and effective strategy termed as Cross-lingual Cross-modal Consolidation (C3 ) to improve mVCMR accuracy. We adopt the ensemble similarity as the teacher to guide the training of each stream, leading to a more powerful ensemble similarity. Meanwhile, we use the teacher for a specific language to guide the student for another language to exploit the complementary knowledge across languages. Ex- tensive experiments on mTVR dataset demonstrate the effectiveness of our C3 method.

By exploiting the cross-modal attention, cross-BERT methods have achieved state-of-the-art accuracy in cross-modal retrieval. Nevertheless, the heavy text-image interactions in the cross-BERT model are prohibitively slow for large-scale retrieval. Late-interaction methods trade off retrieval accuracy and efficiency by exploiting cross-modal interaction only in the late stage, attaining a satisfactory retrieval speed. In this work, we propose an inflating and shrinking approach to further boost the efficiency and accuracy of late-interaction methods. The inflating operation plugs several codes in the input of the encoder to exploit the text-image interactions more thoroughly for higher retrieval accuracy. Then the shrinking operation gradually reduces the text-image interactions through knowledge distilling for higher efficiency. Through an inflating operation followed by a shrinking operation, both efficiency and accuracy of a late-interaction model are boosted. Systematic experiments on public benchmarks demonstrate the effectiveness of our inflating and shrinking approach.

Recent pretrained vision-language models have achieved impressive performance on cross-modal retrieval tasks in English. Their success, however, heavily depends on the availability of many annotated image-caption datasets for pretraining, where the texts are not necessarily in English. Although we can utilize machine translation (MT) tools to translate non-English text to English, the performance still largely relies on MT’s quality and may suffer from high latency problems in real-world applications. This paper proposes a new approach to learn cross-lingual cross-modal representations for matching images and their relevant captions in multiple languages. We seamlessly combine cross-lingual pretraining objectives and cross-modal pretraining objectives in a unified framework to learn image and text in a joint embedding space from available English image-caption data, monolingual and parallel corpus. We show that our approach achieves SOTA performance in retrieval tasks on two multimodal multilingual image caption benchmarks: Multi30k with German captions and MSCOCO with Japanese captions.