Voyant Travel Assistant
Builds trust, then answers fast.
Production-ready travel assistant featuring Transformers.js NLP, Amadeus flight search, Vectara RAG for policy documents, intelligent web crawling with Playwright, and comprehensive anti-hallucination safeguards.
██╗ ██╗ ██████╗ ██╗ ██╗ █████╗ ███╗ ██╗████████╗
██║ ██║██╔═══██╗╚██╗ ██╔╝██╔══██╗████╗ ██║╚══██╔══╝
██║ ██║██║ ██║ ╚████╔╝ ███████║██╔██╗ ██║ ██║
╚██╗ ██╔╝██║ ██║ ╚██╔╝ ██╔══██║██║╚██╗██║ ██║
╚████╔╝ ╚██████╔╝ ██║ ██║ ██║██║ ╚████║ ██║
╚═══╝ ╚═════╝ ╚═╝ ╚═╝ ╚═╝╚═╝ ╚═══╝ ╚═╝
✈️ VOYANT Travel Agent CLI
───────────────────────────────────────
Ask travel questions. I answer in concise English.
You can ask in any language; I will reply in English.
───────────────────────────────────────
What I can help with:
• Weather — "Weather in London this week?"
• Packing — "What to pack for Tokyo in March?"
• Attractions — "Kid‑friendly things to do in SF in late Aug"
• Destinations — "Where to go from Tel Aviv in August?"
• Flights — "Flights from NYC to London March 15" (live search via Amadeus)
• Policies (RAG) — "United baggage allowance", "Marriott cancellation"
and visas — "Do I need a visa for Japan with a US passport?"
• Web search on consent — events, flights, live info
───────────────────────────────────────
How to ask:
• Include city and month/dates when possible.
• I may ask to use web search or deep research — reply "yes" to proceed.
• I avoid prices/budgeting; I can still suggest options.
───────────────────────────────────────
Commands:
/why Show how I got my answer (sources, reasoning, fact-checking)
exit Quit
───────────────────────────────────────
Note: The web interface offers limited functionality (e.g., deep research is disabled) due to free tier resource constraints.
Please use the CLI for the complete feature set.
Please use the CLI for the complete feature set.
Key Features
Anti‑Hallucination
Self‑evaluates to prevent hallucinations, requiring citations and secondary verification.
Web Insights
Crawlee/Playwright crawling with anti-bot stealth for policy extraction. Brave/Tavily search with fallback when internal data insufficient.
Smart Routing
Transformers multilingual NLP with zero-shot routing. 20-40% LLM reduction via Regex→Transformers→LLM cascade. Natural multi-turn support.
Multi-Source Data Integration
Connects to Open‑Meteo, REST Countries, OpenTripMap, and Amadeus for verified travel info.
Enterprise-grade RAG
Advanced retrieval-augmented generation with Vectara integration for searching and retrieving information from internal knowledge bases, policy documents, and enterprise data sources.
Clean Architecture
CLI/REST API with TypeScript validation, 685+ tests, circuit breakers, rate limiting, and Prometheus observability.
Quick Start
# Clone repository and enter root directory
git clone https://github.com/chernistry/voyant/ && cd voyant/root
# Install dependencies
npm install
# Start CLI
npm run cli
# Or start HTTP server
npm run dev
# Run tests
npm test
git clone https://github.com/chernistry/voyant/ && cd voyant/root
# Install dependencies
npm install
# Start CLI
npm run cli
# Or start HTTP server
npm run dev
# Run tests
npm test
Live Conversation Demo
Demo Notice: This page shows a real conversation transcript from the actual assistant. The conversation content is authentic and was recorded during live testing using the MistralAI Mistral-Nemo model. Only the presentation format is static - the actual assistant runs via a CLI and simple REST API.
CLI
npm run cliREST
POST /chat, GET /healthz, GET /metrics
Details and examples in README.md.
⚡ Update: Functionality has been significantly expanded since this conversation was recorded. For the full experience with latest features including advanced RAG, policy search, and enhanced routing, please use the CLI version instead of just reading the transcript.
⚡ Update: Functionality has been significantly expanded since this conversation was recorded. For the full experience with latest features including advanced RAG, policy search, and enhanced routing, please use the CLI version instead of just reading the transcript.
Notes on Prompt Engineering
- Format Priming: Enforced JSON/bullet-point schemas ensure structured, parseable outputs while preventing hallucinations.
- Few-Shot Prompting: Curated examples in router.md and cot.md train accurate intent classification and slot extraction.
- Chain-of-Thought: Structured reasoning in cot.md and verify.md decomposes analysis into verifiable steps.
- Guardrails: blend.md requires fact-grounding + verify.md provides secondary validation of claims.
- Query Optimization: Rewrites user search queries to reduce noise.
Other Perks
- Anti-hallucination receipts with self-check and sources.
- Resilient data blend: APIs → search fallback when providers fail.
- Smart routing + memory for natural multi-turn conversations.
- Optional /metrics endpoint (Prometheus or JSON) and structured logs (PII‑redacted by default).
- Security-conscious: host allowlist, timeouts, retries with jitter, and circuit-like backoff.
- E2E transcripts and golden tests for the core scenarios.
Priority Roadmap
These are the features we would implement next if we continued development beyond this demo.
Deliberate tradeoffs (assignment scope):
- Linear graph today: prioritized debuggability and clear guardrails; interfaces are stateless and typed to drop in parallel branches and rollbacks.
- Prod wrappers: avoided Docker/CI to keep local start ≤5 min; packaging and minimal CI are listed below.
- check_circleRouter Optimization: COMPLETED
Implemented Transformers.js with 'Xenova/bert-base-multilingual-cased-ner-hrl' for multilingual NER, language detection with langdetect library, enhanced entity extraction (LOCATION, DATE, MONEY, DURATION), and Transformers > LLM > Rules cascade routing with 20-40% latency reduction. - check_circleWeb Insights: COMPLETED
Implemented Crawlee (Cheerio/Playwright) for comprehensive web page content extraction and AI summarization when users request complex queries, providing detailed analysis beyond basic search snippets with source citations and receipts. Cheerio handles the fast/simple pages, Playwright handles the complex ones that need JavaScript rendering. - check_circlePolicy RAG & Anchored Citations: COMPLETED
Implemented Vectara RAG integration with PolicyAgent for airlines, hotels, and visas. Seeded 6 policy documents with paragraph-level citations including URL + snippet in receipts. Full cascade routing (Transformers > LLM > Rules) with robust error handling and Brave search fallbacks. - check_circleFlight Search Integration: COMPLETED
Implemented Amadeus API integration for live flight search with complete itineraries, pricing, connection details, and IATA code conversion. Features natural language date parsing and comprehensive flight information display. - check_circleCircuit Breaker & Rate Limiting: COMPLETED
Implemented opossum circuit breaker with per-host configuration and Bottleneck rate limiter. Features configurable timeouts, error thresholds, and comprehensive metrics/stats collection with /metrics endpoint integration. - check_circlePolicy Browser Mode v1: COMPLETED
Headless Playwright navigation to official policy pages with screenshot+DOM evidence, verbatim citations, and content hashing for chain of custody. - constructionIRROPS & Partial-Leg Changes: FOUNDATION READY
Foundational components implemented: PNR schemas, constraint validators (MCT/fare/carrier), option ranking algorithms, and disruption classification. Ready for production workflow integration and policy rule engine connection. - Hotels & Stays (Amadeus): City + dates + guests search with policy compliance filtering, loyalty integration, and unified itinerary view.
- Profile & Policy Guardrail Engine: YAML-based rules engine for corporate travel policy with IN-POLICY/OUT-OF-POLICY badges and exception workflows.
- Unified Itinerary & Smart Notifications: Merged flights+hotels+POIs with ICS export and proactive nudges (check-in, gate changes, cancellation windows).
- Trip Risk & Resilience Score: Predictive disruption scoring with Plan B recommendations based on MCT buffers, weather patterns, and carrier reliability.
- Enterprise Observability & Packaging: P95/P99 metrics, OpenAPI documentation, Docker/Lambda packaging, and cost guardrails with budget controls.