---
id: ch6-rag-technology
type: spine
title: "Inside the Personal Farmer: prompts, intents and pickup hubs"
readingTime: 2
standalone: true
core: false
teaser: "How a chat about dinner becomes a checkout and a pickup point, and why another PoliRural+ chatbot uses retrieval instead."
voice: thinker
parent: null
diagram: fig-field2fork-architecture
recallQ: "How does Field2Fork's 'Personal Farmer' turn a chat into an order, and which PoliRural+ open-call project documents retrieval-augmented generation instead?"
recallA: "A conversational middleware uses structured system prompting and intent routing over a PrestaShop catalogue, shows products as HTML cards, uses named-entity recognition to assign a local pickup hub, and hands over to checkout through a tokenised URL. Retrieval-augmented generation is documented for DASK's Kythera chatbot, not for Field2Fork."
highlights:
  - "The shop (PrestaShop) keeps products, prices and orders; the AI layer handles only the conversation"
  - "Intent routing decides what the user wants; HTML cards show real catalogue items in the chat"
  - "Named-entity recognition picks out a place to assign a pickup hub; a tokenised URL hands over to checkout"
  - "D6.3 documents RAG for DASK (Kythera), not for Field2Fork"
status: accepted
concept: field2fork-ai-hub
lang: en
lens: italy
depth: technical
formalism: none
visuality: balanced
lengthBand: standard
genre: explainer
state: edited
carriers: prose|table|diagram
---

D6.3 describes the Field2Fork AI Hub as a **containerised, hybrid e-commerce microservices architecture**: a PrestaShop back end with a separate conversational front end. The split matters more than any single model.

## Four moves from chat to order

1. **Structured system prompting.** The assistant works under a structured system prompt that frames the conversation around the producers' catalogue.
2. **Intent routing.** Each message is routed by what the user intends to do, so that a question about a product, a basket request or a pickup request reaches the right handler.
3. **Dynamic HTML cards.** Product suggestions appear as cards rendered from catalogue records, so the user sees catalogue items with their shop data rather than free-text descriptions.
4. **Named-entity recognition (NER) for logistics, then a tokenised checkout.** NER extracts the place a user mentions, and the zero-kilometre engine matches it to the nearest farm pickup hub. A tokenised checkout URL then hands the basket to the ordinary shop checkout.

In other words, **the shop remains the system of record**: products, prices, orders and payment stay in PrestaShop. The AI layer translates between how people talk ("something vegetarian for six, no nuts") and what the catalogue holds. The same catalogue also serves a conventional, search-optimised store, which is the second track for users who prefer not to chat.

## Retrieval is another option

The old version of this block described Field2Fork as a retrieval-augmented generation (RAG) system. D6.3 does not say that. Within the Develop call, RAG is documented for **DASK** on Kythera: its localised [JackDaw](#c/jackdaw) chatbot uses RAG over a knowledge base, with voice input for Greek-language farming advice. LocalLinks.AI uses AI differently again, for **semi-automated data extraction with confidence scoring** when it builds its producer database.

| Project | AI job | What stays authoritative |
|---|---|---|
| Field2Fork | Conversation to shopping basket | Shop catalogue and orders |
| DASK | Answering farming questions | Curated knowledge base (retrieved) |
| LocalLinks.AI | Extracting and validating supplier data | Human-validated records |

## The hidden cost

In all three cases someone has to curate the data. The cross-pilot policy brief for the Commission (D7.3, Annex 10) found that feeding rural, multilingual data into PoliRural+ AI systems "required months of manual work: annotation, formatting, and validation". It recommends a **"Data Curation Line" in EU AI project budgets**. A Personal Farmer is only as good as the catalogue the producers keep up to date.

**Sources:**
- D6.3 Comprehensive Report on Funded Activities and Impact Assessment, §3.2 Project results and achievements (Develop call: Field2Fork AI Hub, DASK, LocalLinks.AI). [PDF](https://www.poliruralplus.eu/docs/437/D6.3_Comprehensive_report.pdf)
- D7.3 Policy Brief No. 1, Annex 10 "AI-Enabled AKIS, FAIR Data and Carbon-Smart Food Systems", Execution challenges and Recommendation 1. [PDF](https://www.poliruralplus.eu/docs/349/D7.3_Policy_Brief_No.1.pdf)