AIForge

Transforms natural language task descriptions into executable Python code through LLM generation, implementing a 'Code is Agent' philosophy where generated code directly manipulates the execution environment. The system uses multi-turn LLM interactions with configurable providers (OpenAI, DeepSeek, OpenRouter, Ollama) to synthesize task-appropriate code that runs in an isolated Python sandbox with pre-installed common libraries, enabling self-correction through iterative feedback loops when execution fails.

Provides a unified interface (AIForgeLLMManager) for seamless switching between multiple LLM providers including OpenAI, DeepSeek, OpenRouter, and local Ollama deployments. Implements lazy-loading to instantiate provider clients only when needed, reducing memory overhead and startup time. Each provider is abstracted behind a common interface, allowing runtime provider selection and fallback strategies without code changes.

Maintains execution state (variables, imported modules, defined functions) across multiple code generation and execution cycles within a single session, allowing subsequent generated code to reference and build upon results from previous executions. The system preserves the Python interpreter state between runs, enabling multi-step workflows where each step depends on outputs from previous steps without requiring explicit state passing or serialization.

Captures comprehensive execution logs including LLM prompts, generated code, execution output, error tracebacks, and timing information, storing them in structured format for debugging and auditing. The system provides detailed visibility into each step of the task execution pipeline, enabling developers to understand why code was generated a certain way and why execution succeeded or failed, with optional log export for external analysis.

Implements a hierarchical caching system with three tiers: (1) AiForgeCodeCache—basic SQLite-backed storage with metadata indexing, (2) EnhancedAiForgeCodeCache—semantic analysis and success rate tracking to prioritize high-confidence cached solutions, (3) TemplateBasedCodeCache—pattern matching with parameter extraction for reusable code templates. The system prioritizes execution of previously successful code modules over LLM generation, significantly reducing API calls and latency by matching incoming tasks against cached solutions before invoking the LLM.

Provides AIForgeRunner—a sandboxed Python execution environment that runs generated code with pre-installed common libraries (numpy, pandas, requests, etc.), real-time result feedback, detailed logging, and configurable error retry mechanisms. The environment maintains state persistence across multiple executions within a session, tracks execution errors, and supports automatic retry with up to N configurable rounds, allowing the LLM to receive feedback and self-correct failed code generation attempts.

Orchestrates end-to-end task execution through AIForgeCore, which coordinates natural language input → LLM code generation → sandbox execution → error feedback → iterative refinement cycles. The system manages task state, tracks execution history, and implements a feedback loop where execution errors are analyzed and passed back to the LLM to generate corrected code, enabling complex multi-step workflows to complete autonomously without manual intervention.

Provides AIForgeConfig system supporting four initialization modes: (1) Quick Start—direct API key initialization, (2) Provider-Specific—explicit provider and model selection, (3) Configuration File—TOML-based declarative configuration, (4) Configuration Wizard—interactive setup assistant. The system abstracts provider credentials, model selection, cache settings, and execution parameters into a unified configuration object, enabling flexible deployment across different environments (local development, Docker, cloud platforms) without code changes.

Provides a CLI interface enabling users to submit tasks directly from the terminal, specifying task descriptions, LLM provider/model, cache preferences, and execution parameters through command-line arguments or interactive prompts. The CLI integrates with AIForgeCore to execute tasks and display results, logs, and execution history in a terminal-friendly format, enabling scripting and automation of AIForge workflows.

Provides a terminal-based GUI using the Textual framework, enabling interactive task submission, real-time execution monitoring, result visualization, and configuration management through a mouse/keyboard-driven interface. The GUI displays execution logs, code generation steps, and results in organized panels, allowing non-technical users to interact with AIForge without command-line knowledge while maintaining full feature access.

Provides a web-based interface built on FastAPI and Uvicorn, exposing AIForge functionality through REST API endpoints and an optional web dashboard. The backend handles task submission, execution management, result retrieval, and configuration through HTTP endpoints, enabling remote access, integration with web applications, and deployment as a microservice. The web interface supports real-time execution updates through WebSocket or polling, allowing users to monitor task progress from any browser.

Automatically detects which libraries are available in the execution environment and adapts generated code to use available alternatives when preferred libraries are missing. The system maintains a library compatibility matrix and instructs the LLM to generate code using available libraries (e.g., using matplotlib instead of plotly if plotly is unavailable), ensuring generated code executes successfully without requiring manual library installation.