WhereNext.ai

Three fan-out/fan-in sub-orchestrators

• Guide info runs eight generations concurrently via Task.all: introduction, travel logistics, climate and seasons, accommodation, food and drink, attractions and activities, other experiences, and partnership resolution.
• Itinerary generates a skeleton first, then fans out one task per month, then persists the result to Strapi.
• Images generates and uploads in parallel, with generated URLs cached in Azure SQL keyed by image type, provider and variant.
• Every activity is invoked through callActivityWithRetry with a 5-second first retry and 3 attempts, so a transient model or API error retries in place rather than discarding a half-built guide.

Chunked itinerary generation with a retry boundary that matches the failure unit

• Itineraries are generated in week-day chunks rather than in one shot, so the retry boundary lines up with the unit that actually fails — a broken chunk is regenerated on its own instead of throwing away a finished itinerary.
• Chunking, schema shape and persistence are each covered by their own tests (itinerary-chunked, itinerary-schema, week-day-itinerary-data, strapi-itinerary).

An anti-repetition ledger carried between chunks

• Piecewise generation creates a real problem: each chunk has no idea what the previous ones recommended. Each chunk is therefore handed a compact ledger of what earlier chunks already used, so a seven-day guide doesn't suggest the same market four times.

prose-length-safety.ts — one source of truth for prose length

• The limits the post-generation shortener enforces and the maxCharacters guidance baked into the OpenAI JSON schema descriptions both read from a single PROSE_LIMITS object, so the two can't silently drift apart — a nasty and hard-to-spot class of bug, designed out rather than tested for.
• All prose passes through prepareProseForPersistence before it reaches Strapi, giving a single choke point for length safety.

An AI country fallback for contested geography

• Some upstream location lookups return no country division at all — disputed or contested territories. Rather than writing empty country data into Strapi and breaking webhooks and partnership targeting downstream, the pipeline resolves a best-guess country and backfills it during setup.

Structured-output discipline against early, unreliable models

• Built starting in late 2023, when getting dependable JSON out of a model was genuinely hard. Prompts are assembled from composable helpers (prompt-helpers.ts, buildPromptText) and paired with explicit JSON schemas using additionalProperties: false, minItems/maxItems and required fields, so malformed output is caught rather than persisted.
• Prompt construction also adapts to trip mode — a long-stay guide asks materially different FAQ questions (visas, monthly rentals, eSIMs, coworking) than a week-long trip.