Personal Second Brain

Most recent

See 344 more →

RestGPT - Connecting Large Language Models with Real-World RESTful APIs

2 min read

TL;DR

RestGPT connects LLMs to real-world RESTful APIs through a coarse-to-fine planning loop — planner, API selector, caller, and response executor — achieving 79% task success on complex multi-step benchmarks without fine-tuning.

Summary

Generic tool-use frameworks fail on real-world REST APIs because they lack structured planning and response parsing. RestGPT addresses this with a four-component architecture: a high-level planner decomposes goals into sub-tasks, an API selector identifies the correct endpoint from an OpenAPI specification, a caller constructs and executes the HTTP request, and a response executor maps raw API results back to task context. Evaluated on RestBench (TMDB and Spotify), it achieves 79% on complex multi-step tasks without any fine-tuning.

Key contributions

  1. Proposes a coarse-to-fine planning loop that decomposes high-level goals into concrete API call sequences.
  2. Introduces a dedicated response executor module that translates raw API output into task-relevant context — a component generic frameworks omit.
  3. Demonstrates that OpenAPI specifications provide sufficient context for LLMs to construct valid HTTP calls without fine-tuning.
  4. Establishes RestBench as a benchmark for evaluating LLM agents on authentic real-world web services (TMDB, Spotify).

When to cite

  1. When designing an LLM agent architecture for interacting with REST APIs and needing a reference decomposition.
  2. When arguing that separating planning, selection, and execution produces cleaner API integration than monolithic prompting.
  3. When evaluating LLM agents on multi-step real-world API tasks rather than synthetic tool-use benchmarks.
  4. When justifying the need for a dedicated response parsing layer in API-connected LLM systems.

LLM GPT API

Graph View