Is Your Data Strategy Ready for Generative AI?

[this originally appeared on the Gradient Flow newsletter ]

The deeper I explore Generative AI (GenAI) and Large Language Models (LLMs), the more I confront the complexities of integrating custom data to improve their performance. Whether it’s through fine-tuning or building retrieval augmented generation (RAG) systems, the success of these applications hinges on our ability to harness the power of unstructured data locked away in various formats.

Despite the growing demand for tools to facilitate these new data requirements, it is unfortunate that investment in data tools has not kept pace with other areas. I can’t help but reminisce about the influx of data engineering tools and startups that emerged several years ago, focusing on building and orchestrating data pipelines for structured data destined for data warehouses and lakehouses . Where are the equivalents for unstructured data in the era of GenAI and LLMs?

While building GenAI and LLM prototypes might seem simple, achieving a robust and replicable AI practice demands a sophisticated approach to data management and engineering. The primary hurdle is to modify these established practices to accommodate the novel and diverse range of data types and sources employed by GenAI. Uncertain long-term costs and evolving regulations pertaining to GenAI add another layer of complexity.?

Fortunately, the landscape is undergoing a positive shift. We’re seeing more tools specifically built to handle the unstructured data that fuels custom GenAI and LLM applications.

Can GenAI reach its full potential? Let’s explore the data engineering and data management strategies crucial for its success.

Data Gathering and Ingestion

• Unstructured Data Processing: Two key components of a RAG application, chunking and embedding, rely on extracting data from various formats. Most enterprise data exists in unstructured formats like HTML, PDF, CSV, and images. Other data types, such as audio, pose unique challenges in terms of processing and analysis. Extracting and transforming this data into AI-friendly formats (e.g., JSON) is essential for utilizing it with vector databases and LLM (Large Language Model) frameworks. This capability enables AI applications to leverage the vast amount of unstructured data available, opening up possibilities for various use cases, such as financial report analysis, customer support, and document search.
• Data Source Aggregation and Integration: AI systems require data from various sources, such as databases, cloud storage, and APIs. Integrating these sources into AI workflows simplifies data pipeline creation and model training, allowing enterprises to leverage existing data assets and streamline AI development. For GenAI, tools must address the need for large volumes of varied, unstructured data across different domains, essential for peak performance and broad applicability.
• Optical Character Recognition (OCR): Many documents and images contain valuable textual information that is not readily accessible to AI systems. OCR technology extracts text from images and documents, converting unstructured visual data into machine-readable text. This is crucial for digitizing and making textual data from visuals accessible for AI applications. OCR enables various use cases, such as invoice processing, receipt data extraction, and building datasets for text recognition models, expanding the potential for AI to process and analyze visual information.

Data Preparation and Refinement

• Data Cleaning and Standardization: This process removes artifacts, inconsistencies, and errors from raw data, making it suitable for AI consumption. By clustering , standardizing and correcting text data, removing irrelevant sections, and merging duplicate information, LLMs can process the data more efficiently and accurately, reducing the effort required for data preparation. For visual data, such as images and video, the team at Visual Layer addresses these needs and provides a discovery and search interface.
• Data Enrichment and Transformation: Enhancing raw data with additional information, such as metadata, and transforming it into AI-optimized formats like JSON enables more accurate and context-aware AI applications. Generating summaries of text data also helps LLMs better understand and process the information. Incorporating additional structure on top of metadata, such as knowledge graphs , can yield even better GenAI applications. In this context, entity resolution – the process of linking mentions of real-world entities across data sources – is especially beneficial.
• Data Labeling and Annotation: Adding context or identifying elements within data points through labeling and annotation supports supervised learning, which is crucial for fine-tuning, training, and evaluating GenAI models. There are many startups and services to choose from for this purpose.
• Data Validation and Quality Assurance: Text data poses unique challenges compared to structured data, as it often contains ambiguities, context-dependent meanings, and diverse writing styles. Validating and cleaning text data requires advanced techniques like natural language processing, and entity recognition and resolution. Computer vision and audio data also present their own challenges, such as variations in lighting, angles, and background noise. Rigorous data validation and quality assurance processes are essential to mitigate biases, inconsistencies, and errors that could negatively impact the performance and fairness of Generative AI and LLM applications.

Data Utilization and AI Integration

• Data Retrieval and Search: Enables efficient querying and retrieval of relevant information from large datasets in real-time. For GenAI, data primarily comes from unstructured sources, and specific applications may require custom representations (embeddings). This functionality is crucial for AI applications such as question-answering, recommendations, and copilots.
• Data Visualization and Exploration: Offers tools to visualize data distributions, trends, and anomalies, which can be more challenging for unstructured data. These tools help gain insights, validate data quality, and make informed decisions during AI project development. Visual Layer provides an example of what an interface would look like for computer vision data.
• AI Model Integration and Serving: Seamlessly integrates data management with AI modeling and serving pipelines, allowing efficient data flow for model training, real-time inference, and incremental updates. This integration improves iteration speed and reliability for developing and deploying AI systems. In the case of GenAI, data of various modalities may come from different formats and sources, adding complexity to the integration process.

The Post-Modern Data Stack ???

GenAI and LLMs underscore the critical role of data work, often demanding domain expertise from individuals who may not possess coding proficiency. Consequently, data tools must evolve to cater to non-technical users, empowering them to contribute their valuable insights. Moreover, the iterative nature of data preparation for LLMs and GenAI necessitates tools that can be readily refined and enhanced to attain the desired level of accuracy. Once data extraction, classification, and other models are established, the ability to execute batch processing becomes essential for efficient and scalable data handling.

Companies with a well-established data practice have consistently held an advantage in machine learning and AI. To create a sustainable and scalable AI environment, particularly for GenAI, organizations must develop a robust data management and engineering infrastructure that can handle novel data types and sources, while navigating the uncertainties of long-term costs and the impact of regulations.

In the rapidly evolving field of GenAI, overcoming hurdles associated with new data types is crucial for success. This requires expertise in distributed computing tools like Ray and Spark, which enable teams to process and analyze large volumes of complex data efficiently. Those who enable teams to overcome their data challenges will be at the forefront of securing an organization’s GenAI budget.

The move to GenAI requires a renewed focus on data engineering and data management tools, specifically designed for the unique challenges posed by unstructured data. By investing in these tools and building a mature data infrastructure, businesses can unlock the full potential of GenAI and LLMs, paving the way for a new cohort of intelligent applications and services.

For more information, check out my slide presentation, "Data for GenAI and LLMs," here .

If you enjoyed this post please support our work by encouraging your friends and colleagues to subscribe to our newsletter:

gradientflow.substack.com