I am using TabbyML in our organization and noticed that the repository context injection is not providing relevant code pieces. The injected context often consists of unrelated code from different modules or repositories. How can I improve the repository indexing in TabbyML to ensure better context injection?
Felix Kleinsteuber
Asked on Apr 22, 2024
TabbyML's repository context searches all relevant snippets with keywords matched in a certain context window, which may lead to unrelated code pieces being injected.
To reduce the impact of verbose keywords like public
, TabbyML uses tf-idf scoring, but there may still be cases of bad matches.
TabbyML is working on incorporating import analysis through LSP in IDE/extension side to improve context injection without the need for repository context.
Future updates will include more utilities and tools for benchmarking repository context with internal codebases on a larger scale to enhance the effectiveness of context injection.