Trampoline-AI/predict-rlm
Production focused Self-harnessed LM runtime (RLM) that allows the LM to call its sub-lm with DSPy signatures. Define your inputs, outputs, and tools — the model handles its own control flow. Get fully interpretable trajectories and performance that scales directly with model improvements. Without context rot.
What it solves
Many LLM workflows are too complex for a single prompt but too unpredictable for a rigid, fixed chain of commands. This project provides a runtime environment that allows models to explore, reason, and adapt by writing and executing Python code in a sandbox to solve tasks where the final outcome is known, but the path to get there is not.
How it works
You define the inputs, outputs, tools, and operating procedures using DSPy signatures. The model then operates within a sandboxed REPL, where it can inspect files, maintain state, branch its logic, call focused sub-models, and use tools. It iteratively writes and executes Python code to gather evidence until it reaches the final result, recording every step in a detailed, inspectable trace.
Who it’s for
It is designed for developers building complex, agentic workflows such as codebase analysis, document redaction and extraction, log investigation, financial modeling, and multi-file synthesis. It is not intended for simple chat completions or one-shot classification tasks.
Highlights
- Adaptive Execution: Models can branch, retry, and verify their own work within a defined procedure.
- Sandboxed Environment: Uses a secure REPL (with optional Docker support) to execute Python code safely.
- Context Management: Keeps large inputs as file paths and metadata, avoiding "context rot" by only loading data when needed.
- Observability: Generates detailed, structured traces of every code execution, tool call, and sub-model invocation for easy debugging.
- Multimodal Support: Native ability to process images and documents through sub-model calls.
- Skill Bundling: Reusable modules for domain-specific tasks like spreadsheet manipulation or PDF processing.
Related
- Project
- Project
- Project
- Project
- Project