This Chapter describes the approach chosen in the research to acquire surgical knowledge “topdown”, i.e. from textbooks and interviews with the physicians, so that the robot can learn autonomously how to perform surgery. The Chapter is organized into five main Sections, describing the technological instruments that were developed, how these tools work and what performance was achieved. After the Introduction, Sect. 2 describes how Large Language Models (LLM) can be used to analyze medical text and their limitations. Section 3 addresses the shortcomings of standard LLM by developing an LLM specifically tuned to recognized procedural sentences in surgical text. Section 4 develops the second stage of the pipeline able to extract relational information based on the actions and the actors involved. Section 5 uses the tools developed in an example of extraction of detailed surgical actions from the description of procedures on robot assisted surgical textbooks and papers. Finally, Sect. 6 summarizes the work described and draws some conclusions.