Marchay Platform

Appearance

Home/Projects/Application/Ai Chat Research

AI Chat and Research Interface ​

The primary interface for AI-powered research and discovery.

Epic: NLP Pipeline ​

Plan: Build a keyword-based intent classifier with LLM fallback and regex-based parameter extraction to parse advisor research requests.

Architectural Context: The NLP pipeline is the front door for all AI research. Keyword-based intent detection classifies requests into domains (dining, hotels, flights, experiences, transportation, DMCs) with an LLM fallback for ambiguous queries. Regex extraction pulls structured parameters (location, dates, party size, cuisine) from natural language. A data source selector routes to the appropriate partner APIs.

Tasks ​

  • Implement keyword-based intent detection (dining, hotels, flights, experiences, transportation, DMCs)
  • Build LLM fallback for keyword matching failures
  • Implement regex-based parameter extraction (location, dates, party size, cuisine)
  • Build data source selector (Resy Direct, Resy Parse, Yelp, OpenTable, Tock, Toast Tab, TheFork)

Epic: Search Execution ​

Plan: Integrate with the Orchestrator via POST /api/v1/search and display results as cards with source badges, availability, and pricing.

Architectural Context: Search execution sends parsed queries to the Orchestrator, which fans out to Research Agents. Results are returned as structured cards with source badges, time-slot availability, and pricing. Each result includes a hand-off button for booking.

Tasks ​

  • Implement POST /api/v1/search integration with Orchestrator
  • Build result display as cards with source badges, time-slot availability, pricing
  • Implement hand-off button for each result

Epic: Chat History and Persistence ​

Plan: Build client-side session storage with IndexedDB, auto-save, and session restoration on browser reopen.

Architectural Context: Chat history is stored client-side in IndexedDB for fast access and offline capability. Auto-save with 2-second debounce prevents data loss without excessive writes. Sessions are grouped as RECENT (today) and EARLIER (older) in the sidebar. Session restoration enables advisors to pick up where they left off.

Tasks ​

  • Build IndexedDB session storage with full restoration
  • Implement auto-save with 2-second debounce
  • Build RECENT/EARLIER grouping in ChatHistorySidebar
  • Implement session restoration on browser reopen

Epic: Context Attachment ​

Plan: Build a ContextSearchPopover for real-time Salesforce record search, enabling advisors to attach member/trip context to AI searches.

Architectural Context: Context attachment enriches AI searches with CRM data. The ContextSearchPopover searches Salesforce records in real-time, allowing advisors to attach a member or trip as context. This context is injected into Orchestrator requests, producing more relevant and personalized results.

Tasks ​

  • Build ContextSearchPopover for real-time Salesforce record search
  • Implement member/trip context attachment to AI searches
  • Build context injection into Orchestrator requests

Epic: Preset Routines ​

Plan: Implement 6 preset routines on the AI landing page with an animated AI button featuring lava blob CSS and sparkle effects.

Architectural Context: Preset routines provide quick-start workflows for common advisor tasks: Search Dining, Book Flight, Draft Itinerary, Find Hotels, Research Destination, Compose Email. The animated AI button serves as the primary call-to-action, drawing attention to the AI capabilities.

Tasks ​

  • Implement 6 preset routines on AI landing page: Search Dining, Book Flight, Draft Itinerary, Find Hotels, Research Destination, Compose Email
  • Build animated AI button with lava blob CSS and sparkle effects

Marchay Platform Documentation

© 2026 Marchay