Abstract: A graph-based representation of a source code program is generated in a background process of an edit session of a software development tool. The graph is used to facilitate the construction of a context for a prompt to a large language model that answers a user's query regarding the source code program in the edit session. The graph contains nodes that represent functions, macros, and types of the program and edges that depict a usage or definitional relationship between two connected nodes. The edges are generated from internal data structures generated from compiler-related analyses performed on the source code program in a background process. The graph is traversed to generate a sequence of code directives that provide the model with a structure of the source code program that includes data from the internal data structures not apparent from or contained in the source code program.

Abstract: A virtual directory is created in a software development tool that lists the files having source code components (e.g., files, functions, methods, types, classes) of a codebase that relate to a user query about the codebase. The files of the codebase are partitioned into chunks with each chunk having a respective embedding. A search for the source code components relevant to the query is performed using an embedding of the query and the chunk embeddings representing the files of codebase. As a file of the virtual directory is edited, the chunk embeddings are updated and the virtual directory is updated with a reference to the revised file.

Abstract: A virtual directory is created in a software development tool that lists the files having source code components (e.g., files, functions, methods, types, classes) of a codebase that relate to a user query about the codebase. The files of the codebase are partitioned into chunks with each chunk having a respective embedding. A search for the source code components relevant to the query is performed using an embedding of the query and the chunk embeddings representing the files of codebase. As a file of the virtual directory is edited, the chunk embeddings are updated and the virtual directory is updated with a reference to the revised file.

Abstract: Some embodiments engineer a prompt for submission to a language model, such as a software development large language model. Some embodiments ascertain a relationship between code development information and potential context. Code development information includes static analysis results, project settings, development tool history or status data, and other software development data which augments training data previously embedded in the language model. Some embodiments compute a prompt inclusion score of the potential context, based on at least the relationship, and use the inclusion score to determine whether to include the potential context in the language model prompt. In some scenarios, an embodiment determines where to place the context in the prompt. Scoring is performed by a formula, statistical scoring model, or machine learning scoring model. Some embodiments reduce context inclusion false positives and false negatives that were based on the use of embedding similarity scores alone.

Abstract: Techniques are described herein that are capable of controlling and/or visualizing context of an artificial intelligence prompt. A user-generated artificial intelligence prompt is detected. In a first technique, a visual representation of contextual information, which includes context regarding the prompt, is generated. Based at least on detection of a user-generated instruction, presentation of the visual representation is triggered. In a second technique, a determination is made that an initial scope of contextual information, which includes context regarding the prompt, includes previous contextual information, which includes context regarding a previous user-generated prompt in a prompt chain that includes the prompt. The initial scope of the contextual information is automatically changed to provide a changed scope that does not include at least a portion of the previous contextual information.