Imagine querying unstructured text data with the same ease and efficiency as a structured database. That's the promise of ELEET, a novel query execution engine designed to handle text and tables seamlessly. Currently, analyzing data spread across tables and text files requires complex extraction pipelines and manual transformations. ELEET changes this by introducing *learned multi-modal operators* (MMOps). These operators, powered by a specialized small language model (SLM), act like traditional database operators (joins, unions, etc.) but can directly process text. For example, a multi-modal join can link a patient table directly with their textual medical reports, allowing analysts to correlate patient age with diagnoses extracted *directly from the text*. Why not just use large language models (LLMs)? While LLMs like GPT-4 can technically convert text to tables, their massive size makes them computationally expensive for query processing. ELEET's smaller, purpose-built SLM dramatically speeds up queries—up to 575x faster than LLM-based approaches in experiments—without sacrificing accuracy. This efficiency comes from a novel model architecture that avoids the costly autoregressive decoding used in LLMs, opting for a faster, single-pass extraction method. The model is pre-trained on a massive new dataset of aligned Wikipedia text and Wikidata tables, teaching it the core skills needed for table extraction. While this research primarily focuses on online query execution, ELEET can also pre-compute extractions offline, like a materialized view. However, online execution offers advantages for constantly updated text collections and ad-hoc queries. This work opens exciting possibilities for querying unstructured data efficiently. Future research includes optimizing the query compiler and extending ELEET to handle other data modalities, such as images, bringing us closer to truly unified data analysis.