llm-vscode

Generates code suggestions in real-time as developers type by sending the current file's prefix and suffix context (relative to cursor position) to a configurable LLM backend (Hugging Face Inference API, Ollama, OpenAI, or TGI). The extension automatically tokenizes input using the tokenizers library to fit within the model's context window, constructs a prompt with special tokens (start_token, end_token, middle_token), and renders completions as ghost-text overlays matching VS Code's native completion UI pattern. Supports multiple model backends without leaving the editor.

Detects whether generated code matches sequences in The Stack training dataset by performing a rapid first-pass Bloom filter lookup against a pre-built index, then optionally linking to stack.dataportraits.org for detailed attribution verification. The extension requires a minimum 50-character code sequence and sufficient surrounding context to perform matching. Triggered via the 'Cmd+Shift+A' keyboard shortcut or command palette. Uses probabilistic matching (Bloom filter) for speed, with acknowledged false positives.

Allows developers to select and switch between 4 different LLM backend types (Hugging Face Inference API, Ollama, OpenAI-compatible, Text Generation Inference) via VS Code settings without modifying code or restarting the extension. Each backend has configurable parameters: base URL, model ID, and custom request body JSON. The extension constructs HTTP POST requests with backend-specific URL patterns and forwards the configured requestBody to the selected endpoint. Supports automatic token counting to fit prompts within each model's context window.

Automatically measures and fits the code completion prompt within each model's context window by using the tokenizers library to count tokens in the prefix, suffix, and surrounding code. If the combined prompt exceeds the model's maximum context length, the extension truncates the prefix and/or suffix to fit. This ensures requests succeed without manual context management by the developer. Token counting happens per-request with computational overhead.

Exposes core extension functionality through VS Code's command palette (Cmd/Ctrl+Shift+P) and dedicated keyboard shortcuts. Documented commands include 'Llm: Login' for authentication and 'Llm: Code Attribution Check' (Cmd+Shift+A). The extension registers these commands with VS Code's command registry, making them discoverable and remappable. Additional commands exist but are not enumerated in available documentation.

Manages Hugging Face API authentication by automatically detecting tokens from the huggingface-cli cache on disk (if huggingface-cli was previously configured) or accepting manual token entry via the 'Llm: Login' command. Tokens are stored in VS Code's secure credential storage (mechanism not specified). The extension validates tokens before making API requests to the Hugging Face Inference API. Tokens can be obtained from hf.co/settings/token.

Exposes extension configuration through VS Code's standard settings UI (Cmd+, → filter 'Llm'). Developers can configure backend type, model ID, base URLs, request body parameters, and other options via a searchable settings panel. The full list of available configuration options is not enumerated in documentation. Settings are persisted in VS Code's configuration store and applied immediately or after extension reload.

Renders generated code completions as ghost-text overlays in the editor, matching VS Code's native code completion UI pattern. The extension inserts completions at the cursor position when accepted (typically via Tab or Enter key). Ghost-text appears in a dimmed color to distinguish it from actual code. The rendering is handled by VS Code's InlineCompletionItemProvider API (or similar completion API).