Engineering note
How we built the Planta Livre agent pipeline
How we built Plant AI Agent for Planta Livre: GPT-4.1 multilingual extraction, semantic matching over a 10,560+ plant catalogue, and a real-time budgeting UI processing 150+ line items per request.
In the plant industry, procurement still runs on email. Requests arrive as free-form messages, PDF price lists, spreadsheets, and photos — and someone has to turn all of that into an accurate, priced budget. For Planta Livre we built Plant AI Agent, an intelligent procurement and budgeting system that automates that workflow end to end.
The problem
A single procurement request can contain more than 150 line items. Staff extracted plant names, quantities, and specifications by hand — from Portuguese and English text, across every format a customer might send — and then matched each item against a catalogue of over 10,560 plants. Informal descriptions, synonyms, and regional naming variations made that matching the slowest, most error-prone step of the whole process.
Understanding requests with GPT-4.1
The first stage of the pipeline is multilingual natural-language understanding (PT/EN) built on OpenAI GPT-4.1. It extracts structured line items — plant name, quantity, specifications — from unstructured emails, documents, and chat messages. Ingestion is multi-format by design: emails, PDFs, Excel, DOCX, and images all feed the same pipeline, unifying every procurement channel into a single system.
Semantic product matching
Extraction is only half the problem — “buxinho” still has to become the right catalogue product. We embed both requests and catalogue entries with Sentence-Transformers and rank candidates by cosine similarity in a vector database covering the full 10,560+ plant catalogue. The matcher is robust to informal descriptions, synonyms, and naming variations, and returns results in under a second per item.
From matches to budgets
Matches land in a web interface built with Next.js 15, React 19, and TypeScript on a FastAPI backend: real-time budget tables, drag-and-drop interactions, and confidence-based indicators so staff review the uncertain matches first instead of re-checking everything. Finished budgets export directly into the procurement workflow.
Results
- Significant reduction in budget creation time compared to the manual workflow
- 150+ line items processed per request
- Sub-second product matching across the full catalogue
- Fewer errors in procurement quotes
Architecture notes
The system combines GPT-4.1, Sentence-Transformers, and a vector database, wired together over the MCP Protocol so that models and input channels can be swapped without rebuilding the pipeline. The pattern — LLM extraction, semantic matching, and a human-review UI on top — generalizes to procurement in any catalogue-heavy industry.