Generate with classic techniques, supporting libraries and AI a full overview for the user and GitLab the product (in the sense that the product has information for tasks, suggestions and so on) about a specific projects code base. LLM are also quite good at converting unstructured data to a pre-given JSON format for example. The idea is to have a container based scanner for repositories that gathers as much information as possible from a repository, saves the data to a database that can also run locally in IDE's and the server. Repository : https://gitlab.com/gitlab-org/code-creation/repository-x-ray POC Repo: https://gitlab.com/timzallmann/repo-xray-test ## Usage of the collected meta data * Code suggestions context for both code generation and code completion * Duo Chat * Provide possibility to ask questions about a repository ("where are error codes for the API defined?", etc.) * Code Tasks - Extra context for tasks around specifically selected code * Project Tasks - Complete project activities based on a local agent setup * Code Intelligence + Navigation - Together with the latest code search we should be able to provide with this full code navigation * Statistics - Clear insights around repositories, what they use and for example also insights on a organization level between different projects ## Information that can be gathered * Project Type * Languages used * Frameworks/Libraries used, Frameworks for specific tasks (like which frontend framework, testing framework is used, etc.) * Call Trees and Code graphs * Test coverage * Commit History + Statistics of project files * Linting warnings * Regex explanations * Actual usage data from production * Actual stack traces from review/production/etc * Graphs of deployed web sites * etc. # Workflow ### 1. Scanner indexes repository * Our pre-built scanner container can be pointed at any repository (like our security scanners, or just part of CI to start with) * Scans files, using tooling, LLM's, etc. to gather as much information as possible #### 1.1 Data collection Examples what and how data can be collected: * Using tools like tree sitter can help us to gather full trees of each code file (e.g. functions, classes) * Using LLM's to explain each file, function and also inspect code files for certain patterns (are specs, etc.) * Embeddings for content (descriptions + actual code) are created ### 3. Data stored in portable DB * Idea is to store data once in a portable database that is capable of storing embeddings * DB can be downloaded with the repository and be used directly in the IDE to provide capabilties locally * During POC tried both Vector DB's and Graph DB * Graph DB have the advantage to store relationships and LLM's are quite good to create Graph Queries to gather the needed data --- # Meta Data Usage ## IDE Local Download By having the possibility for IDE's to download this meta database we can enable use cases like better context for code suggestions, code navigation and more. Also extend use cases like create specs would have the knowledge where the connected file lives. ## Monolith Usage #### Code Suggestions context During the creation of the prompt for code generation or code tasks we can extend the context by finding information in the meta database through semantic search based on embeddings or also dynamig Graph DB queries to gather needed information. ### Repo View - Code Navigation / Intelligence / Meta information * This would allow us together with Zoekt to have full code navigation in the repo browser * We would be able to show a lot of meta information per file ### Project Analysis By having the meta data this would enable us to do repo wide analysis for tech debt removal for example or usage patterns, etc.