Background: Creation and reuse of reliable clinical code sets could accelerate the use of EHR data for research. To support that vision, there is an imperative need for methodologically. driven, transparent and automatic approaches to create error-free clinical code sets. Objectives: Propose and evaluate an automatic, generalizable, and knowledge-based approach that uses as starting point a correct and complete knowledge base of ingredients (e.g., the US Drug Enforcement Administration Controlled Substance repository list includes fentanyl as an opioid) to create medication code sets (e.g., Abstral is an opioid medication with fentanyl as ingredient). Methods: Algorithms were written to convert lists of ingredients into medication code sets, where all the medications are codified in the RxNorm terminology, are active medications and have at least one ingredient from the ingredient list. Generalizability and accuracy of the methods was demonstrated by applying them to the discovery of opioid and anti-depressant medications. Results: Errors (39 (1.73%) and 13 (6.28%)), obsolete drugs (172 (7.61%) and 0 (0%)) and missing medications (1,587 (41.26%) and 1,456 (87.55%)) were found in publicly available opioid and antidepressant medication code sets, respectively. Conclusion: The proposed knowledge-based algorithms to discover correct, complete, and up to date ingredient-based medication code sets proved to be accurate and reusable. The resulting algorithms and code sets have been made publicly available for others to use.