AI Travel Assistant — Intelligent Contract on GenLayer

Repository: https://github.com/UnityNodes/ai_travel_assistant

Overview

An AI-powered Intelligent Contract on GenLayer that helps users plan trips by combining live web data with non-deterministic AI reasoning and on-chain consensus.

Each trip request triggers real web fetches for destination data and currency exchange rates, feeds that context into an LLM prompt, and lets GenLayer validators reach consensus on the AI-generated recommendations through the equivalence principle.

GenLayer Capabilities Used

Capability	How it is used
gl.nondet.web.get() — Web Fetch	Fetches live destination info from the REST Countries API and current USD exchange rates from open.er-api.com. Each validator fetches independently.
gl.nondet.exec_prompt() — Non-deterministic AI	Generates 3 personalized travel options (economy / standard / premium) using the fetched real-world data as grounding context.
gl.eq_principle.prompt_comparative() — Comparative Consensus	Every validator independently runs web fetches + AI generation, then outputs are compared semantically to reach consensus.

Why these capabilities matter

• Web Fetch grounds the AI in real data — actual currencies, regions, languages, and live exchange rates — instead of relying on stale training data.
• Non-deterministic execution means each validator produces its own AI output independently, preventing single-point-of-failure answers.
• Comparative equivalence principle compares validator outputs semantically ("do they cover similar budget tiers and match user preferences?") rather than requiring exact string equality, which is essential for AI-generated text.

Features

• Request a personalized trip with budget, destination, dates, and preferences.
• Two independent web fetches per request inject real-world context into AI reasoning.
• Non-deterministic validators independently generate options; consensus via comparative principle.
• On-chain storage of user profiles and full trip history.
• View functions for querying history, profile, and aggregate stats.

Repository Structure

contracts/
  travel_assistant.py    # Intelligent Contract (Python)
deploy/
  deployScript.ts        # Deploys the contract via genlayer-js
  run.ts                 # End-to-end demo: deploy + request trip + read state
scripts/
  init_validators.py     # (Optional) Recreate validators with Gemini config
package.json

Prerequisites

• Node.js 18+ (tested with Node 23)
• npm
• Docker Desktop (for GenLayer Localnet)
• GenLayer CLI (npx genlayer)
• GenLayer Localnet or Studio access

Install

npm install

Start GenLayer Localnet

npx genlayer init

Follow the interactive prompts to select an LLM provider (e.g., Heurist AI) and enter your API key. This starts all Docker containers and configures validators.

If you need a clean restart:

npx genlayer init
# or manually:
npx genlayer up --reset-db --reset-validators --numValidators 5

Deploy and Run the Demo

npx tsx deploy/run.ts 1500 "France" "2026-05-01 to 2026-05-07" "Cheap flights, cultural tours"

Expected output:

• Transaction status: FINALIZED
• Consensus result: MAJORITY_AGREE
• AI-generated travel recommendation stored in on-chain history
• User profile updated with preferences

How It Works (Step by Step)

1. User calls request_trip(budget, destination, dates, preferences).
2. Inside the equivalence principle block, each validator independently:
   • Web Fetch 1: calls gl.nondet.web.get() to fetch destination data (capital, currencies, languages, region) from the REST Countries API.
   • Web Fetch 2: calls gl.nondet.web.get() to fetch live USD exchange rates from open.er-api.com.
   • AI Prompt: calls gl.nondet.exec_prompt() with all fetched data injected into the prompt context, generating 3 travel options (economy / standard / premium).
3. gl.eq_principle.prompt_comparative() compares all validator outputs and reaches consensus: outputs are equivalent if they cover similar budget tiers, prices are in a comparable range, and the best option matches user preferences.
4. The winning recommendation is parsed and stored as a TravelOption in the user's on-chain history.

Contract API

Write Methods

Method	Description
request_trip(budget, destination, dates, preferences)	Runs web fetch + AI analysis and stores the result
update_preferences(preferences)	Updates the caller's stored preferences

View Methods

Method	Returns
get_history(user)	JSON array of the user's past trip recommendations
get_history_len(user)	Number of trips in the user's history
get_profile(user)	The user's stored preferences string
get_all_histories()	Map of all addresses to their trip count

Storage

• profiles: TreeMap[Address, str] — user preference strings
• histories: TreeMap[Address, DynArray[TravelOption]] — trip history per user

Each TravelOption stores: destination, description, price, duration, rating, is_booked.

Troubleshooting

• NO_MAJORITY: try a different LLM model (DeepSeek V3 via Heurist works well), reduce prompt size, or increase validator count.
• type is not supported: i32: use int instead of i32 for contract method parameters.
• Web fetch errors: verify the REST Countries and exchange rate APIs are reachable from the GenLayer node (docker exec genlayer-jsonrpc-1 curl -I https://restcountries.com).
• running contract failed: check that the AI returns valid JSON; the contract strips markdown wrappers automatically.
• Receipt delays: increase retries and interval in the client script.
• Validator LLM errors: check docker logs genlayer-jsonrpc-1 for provider errors; some models have low max_tokens defaults.