hallucinateapi

HallucinateAPI is an API gateway that implements every GET and POST operation defined in an OpenAPI specification using a language model. Instead of writing business logic, each operation’s description serves as the instruction for the LLM, and the response schema constrains the model output to valid, structured JSON.

Key Features

OpenAPI-driven: Define your API with an OpenAPI 3.0 spec, and HallucinateAPI implements it automatically.
Multi-provider: Supports Google’s Gemini models via Vertex AI and any OpenAI-compatible Chat Completions API.
Schema-validated: Both requests and responses are validated against the OpenAPI schema for type safety.
Swagger UI included: Interactive API documentation served at the root path.
Simple deployment: Single binary with configuration via environment variables or CLI flags. Container-ready with Docker support.
Secure by design: Strict input allowlisting, request size limits, and prompt injection resistance.

How It Works

Load an OpenAPI specification defining your API endpoints.
HallucinateAPI validates the spec and registers routes for all GET and POST operations.
When a request arrives, inputs are validated and formatted into a prompt.
The prompt is sent to the configured LLM provider with the response schema as a structured output constraint.
The model response is validated and returned as the API response.

Quick Start

Using Gemini on Vertex AI:

hallucinate serve \
  --provider gemini \
  --openapi-path /path/to/your/spec.yaml \
  --gcp-project your-gcp-project \
  --gcp-location us-central1 \
  --model gemini-2.5-flash

Using OpenAI:

hallucinate serve \
  --provider openai \
  --openapi-path /path/to/your/spec.yaml \
  --model gpt-4o \
  --api-key your-api-key

Using a local OpenAI-compatible server (e.g., Ollama):

hallucinate serve \
  --provider openai \
  --openapi-path /path/to/your/spec.yaml \
  --model llama3 \
  --url http://localhost:11434/v1/chat/completions

Visit http://localhost:8080 for the Swagger UI.