GenAIScript

Executes LLM queries using JavaScript template literal syntax (backtick-delimited prompts with $` markers) embedded directly in GenAIScript files. The runtime parses these template expressions, sends them to configured LLM providers (OpenAI, Anthropic, or local models), and returns structured or unstructured responses that can be assigned to variables for downstream processing. This approach enables prompt composition as first-class JavaScript expressions rather than string concatenation.

Automatically extracts and parses content from diverse file formats (PDF, DOCX, CSV, plain text) using specialized parsers accessible via the `parsers.*` API. Files are matched using glob patterns or explicit file arrays, parsed into structured or text representations, and made available to LLM prompts via the `env.files` context. The runtime handles encoding detection, format-specific extraction (e.g., PDF text layers, DOCX metadata), and error handling for malformed files.

Executes scripts with automatic file discovery and filtering based on glob patterns or explicit file lists. The runtime matches files against patterns, loads their content, and makes them available to the script via `env.files`. This enables batch processing of files with consistent logic without manual file enumeration.

Formats script execution results for display or export, supporting multiple output formats (plain text, JSON, structured logs). Results can be written to stdout, files, or returned as structured data for downstream processing. The runtime handles serialization of complex data types and provides options for formatting output for human readability or machine parsing.

Defines JSON schemas (using JSON Schema or Zod syntax) to validate and repair LLM-generated outputs. The runtime enforces schema constraints, attempts to repair malformed data (e.g., fixing JSON syntax errors or missing fields), and provides structured output that matches the schema definition. Schemas are defined inline in scripts using `defSchema()` and can be referenced in prompts to guide LLM output format.

Performs semantic similarity search across project files using embeddings and vector retrieval. The `retrieval.vectorSearch()` API accepts a query string, embeds it using a configured embedding model, and returns the most similar files or file chunks ranked by cosine similarity. This enables context-aware file selection for LLM prompts without explicit file enumeration, supporting use cases like 'find similar code' or 'retrieve relevant documentation'.

Enables prompts to invoke other prompts via the `runPrompt()` function, allowing multi-stage LLM workflows where outputs from one prompt feed into subsequent prompts. Each nested prompt has its own context (files, variables, schema), and results are returned as structured data that can be processed or passed to downstream prompts. This pattern supports complex reasoning chains, iterative refinement, and modular prompt reuse.

Executes GenAIScript scripts from the command line using `npx genaiscript run`, enabling automation outside VS Code and integration with CI/CD pipelines, cron jobs, or shell scripts. The CLI accepts script paths, environment variables, and input parameters, executes the script in a headless runtime, and outputs results to stdout or files. This decouples script development (in VS Code) from script execution (in automation contexts).

Provides access to execution context via the `env.*` namespace, including matched files (`env.files`), environment variables, and script parameters. Variables are scoped to the current script execution and can be used in prompts, file matching, and conditional logic. This enables scripts to adapt behavior based on runtime context without hardcoding values.

Provides a native VS Code editor environment for writing, editing, and debugging GenAIScript files with syntax highlighting, autocomplete, and inline execution. Scripts are edited as `.genaiscript` or `.js` files with GenAIScript syntax, and can be executed directly from the editor via command palette or keybindings. The editor provides real-time feedback on script validity and execution results.

Allows configuration of LLM providers (OpenAI, Anthropic, local models) and model selection within scripts or global settings. Configuration may be specified via environment variables, VS Code settings, or inline script parameters. The runtime abstracts provider-specific API differences, enabling scripts to work with multiple LLM backends without code changes.

Supports variable interpolation within prompt templates using JavaScript template literal syntax, allowing dynamic prompt construction based on runtime data. Variables are substituted into prompts before LLM execution, enabling parameterized prompts that adapt to different inputs. Interpolation supports expressions, function calls, and complex data structures.