Bloop apps

Enables fast pattern-matching searches across codebases using regular expressions and literal text queries, powered by Tantivy (a Rust-based full-text search engine). The system pre-indexes code files into an inverted index structure, allowing sub-millisecond regex matching across millions of lines of code without scanning the entire repository on each query. Supports complex regex patterns with syntax highlighting of matches.

Enables developers to search code using natural language queries by converting both code and queries into dense vector embeddings stored in Qdrant (a vector database). The system computes semantic similarity between the query embedding and indexed code embeddings, returning contextually relevant code snippets even when exact keyword matches don't exist. Uses embedding models to capture code intent and functionality semantically rather than syntactically.

Maintains conversation history and context across multiple user queries, allowing developers to ask follow-up questions about code without re-specifying context. The system stores previous search results, code snippets, and LLM responses in memory, and includes them in subsequent prompts to maintain coherent conversations. Supports conversation branching and context pruning to manage token limits.

Implements the core search, indexing, and AI functionality in Rust, providing high performance and memory safety. The backend uses async/await patterns (tokio runtime) for concurrent request handling, allowing multiple search queries and indexing operations to proceed simultaneously without blocking. Includes optimized data structures for fast index lookups and memory-efficient storage of large codebases.

Automatically detects changes in local and remote repositories and re-indexes only modified files rather than the entire codebase. The system tracks file modification timestamps and git commit hashes to identify deltas, then updates both the Tantivy lexical index and Qdrant semantic index incrementally. Supports continuous indexing in the background without blocking user searches.

Manages indexing and searching across multiple repositories simultaneously, supporting both local file system repositories and remote GitHub repositories. The system maintains separate index instances per repository, handles repository cloning/syncing, and provides unified search across selected repositories. Supports adding/removing repositories dynamically without restarting the application.

Processes natural language questions about code by combining search results with LLM reasoning to generate contextual explanations. The system retrieves relevant code snippets using semantic search, constructs a context window with the code and question, and sends this to an LLM (OpenAI, Anthropic, or local models) to generate explanations. Supports follow-up questions and maintains conversation context across multiple queries.

Generates code patches and new features by combining semantic search with LLM code generation, using the indexed codebase as context to ensure consistency with existing code style and patterns. The system retrieves similar code sections, analyzes code style (indentation, naming conventions, patterns), and instructs the LLM to generate patches that match the codebase's conventions. Supports generating patches for bug fixes, feature additions, and refactoring.

Packages the Bloop backend and React frontend into a native desktop application using Tauri, enabling offline-first operation and system-level integration (file system access, native menus, keyboard shortcuts). The Tauri wrapper communicates with the embedded Rust backend via IPC, eliminating the need for external servers and allowing the application to run entirely locally. Supports Windows, macOS, and Linux with native look-and-feel.

Provides a responsive web-based user interface built with React that displays search results with syntax highlighting, code navigation, and interactive exploration. The UI includes real-time search result updates, collapsible code context, file tree navigation, and keyboard shortcuts for power users. Supports multiple themes (light/dark) and internationalization for global audiences.

Analyzes code in 40+ programming languages using language-specific tokenization rules to understand code structure (functions, classes, imports, etc.) rather than treating code as plain text. The system uses language parsers to identify semantic boundaries (function definitions, class declarations, import statements) and indexes these as separate tokens, enabling more precise search and better semantic understanding. Supports custom tokenization rules for domain-specific languages.

Abstracts LLM provider selection, allowing developers to choose between OpenAI, Anthropic, local models, or other providers through configuration. The system implements a unified interface for LLM calls, handling provider-specific API differences, token counting, and error handling. Supports switching providers without code changes and includes fallback mechanisms for provider outages.