Grounded garden consultant for a regional grower
A consultant that diagnoses from symptoms, follows a climate-band calendar, and recommends only products in the grower's assortment — never an invented dose.
What changed. Seasonal questions meet a catalog-bound answer or a human; named products and doses that are not in the base do not reach the user.
Stack. symptom retrieval · grounded generation · dosage guard · messenger bot
Context
A regional grower and garden supplier sells chemistry, seed, stock, and tools to a seasonal audience. The buyer does not arrive asking for a fungicide SKU. The buyer arrives with dying cucumbers and a phone full of generic search results. A human agronomist could close that gap — one person, daytime hours, no chance of covering a spring peak. Free assistants answer “in general” and walk the visitor to whoever ranks. Invented doses on products children eat are not a copy problem. They are a harm and a brand event.
Task
Put a consultant on the grower’s own assortment: diagnose from a symptom description, say what to do this week in the customer’s climate band, and name a product that is actually on the shelf. Unknowns escalate. The bot is a sales instrument bound to inventory, not a public encyclopedia. Voice: a careful neighbour, not a lecture in Latin.
What we built
A reusable circuit with a knowledge base of diseases, pests, practice, varieties, and a calendar — each record tied to the client’s SKUs. The user describes the problem in their own words; retrieval is semantic, not a keyword lottery. Generation is allowed only from retrieved records. An output guard checks that every named product, dose, and interval exists in the base. Failures retry or refuse and pass to a specialist. If the model is down, the bot collects a contact instead of going mute.
The same kernel serves more than one instance. Channels are Telegram and WhatsApp. Dialogues and leads stay with the grower. The calendar is keyed to climate band and crop, not to a named province. Assortment and timing are meant to be refreshed before each season so the bot does not recommend last year’s pack.
What changed
Spring peak questions can be met at any hour without hiring a night agronomist. A recommendation is a SKU the warehouse can pick, with a dose the label already owns. What people ask — crop, symptom, season — becomes a feed for assortment and content, not a lost search elsewhere.
What we would do differently
Ship one crop and one climate band first, with a hard empty-set refusal, then widen the calendar. The temptation is to sound complete on day one. A garden consultant that covers “everything” with a thin base will invent the missing row. Completeness is a later season, not a launch claim.
Related service: AI consultant and chatbots