AI farm record retrieval.

The questions it answers

Farm record retrieval answers the questions that otherwise mean a phone call to the manager, a login to three systems and a hunt through a notebook. When did we last put nitrogen on the river block? Which heifers were treated for lameness in March, and with what? Did the contractor record the product for the gorse spraying last spring?

The answers exist, spread across herd software such as LIC’s MINDA, paddock and livestock tools such as FarmIQ, spray and fertiliser diaries, vet notes, contractor invoices and paper. The assistant searches all of them at once and replies with the record, the source system and the date it was written. It is a retrieval system over the farm’s own data, not a farming adviser.

How record retrieval runs

Inputs

Herd records (calving, mating, treatments), paddock records, spray and fertiliser diaries, vet notes, NAIT movement confirmations, contractor invoices, resource consent documents, staff training records, and photos of paper notebooks.

Systems

A cloud document store, scheduled exports or saved reports from each farm system, a search index with embeddings, a language model, and a chat channel on the farm phone or computer.

Steps

Refresh each source on a schedule; extract text from PDFs and photos; tag every record with paddock, animal, product and date; on each question, retrieve matching records; answer with the source and record date; reply “not found” when nothing matches; log the question.

Human review

The farm manager checks the cited record before acting on anything about animal treatment, chemicals or stock movement. The owner reviews a monthly sample of answers. The vet confirms anything about withholding.

Outputs

Short answers linked to source records, “not found” replies with the closest matches, and a monthly list of gaps such as spray events with no product recorded.

Measures

Answers with the correct source in the monthly sample, share of “not found” replies, age of the newest record in each source, and minutes to answer a typical question.

Risks

Answers from stale exports, mixed-up paddock or animal names, unit errors, and staff information visible too widely.

Illustrative example: a 700-cow Canterbury dairy farm

Illustrative example, synthetic numbers: a 700-cow Canterbury dairy farm with a manager and three staff holds about 4,000 documents and exports going back 12 years. The owner, manager and staff ask roughly 65 record questions a month. Today each one takes about 10 minutes of searching, often by the manager.

Assumptions, at prices listed on 19 October 2026: indexing 4,000 documents at about 1,500 tokens each (6 million tokens) with OpenAI’s text-embedding-3-small at US$0.02 per million (OpenAI pricing). Each answer reads about 10,000 tokens of retrieved records and writes 300, on Claude Haiku 5.5 at US$0.10 per million input and US$0.50 per million output tokens (Anthropic pricing), about a tenth of a US cent per question.

NZ records worth indexing first

• Animal movements. OSPRI states that all movements of cattle and deer between NAIT locations must be recorded within 48 hours. NAIT stays the record of truth; the assistant can show whether a movement in the farm diary has a matching confirmation, so gaps surface inside that window rather than months later.
• Agrichemical and veterinary records. NZS 8409:2021 gives guidance on managing agrichemicals, including plant protection products and veterinary medicines. Spray and treatment records are the most common questions, so index them carefully: product, rate, area or animal, date and operator.
• Business records. Inland Revenue expects records to be kept for at least 7 tax years, and offshore or cloud storage needs IRD approval held by the farm or its provider. Treat the search index as a copy, not the archive.
• Staff information. Contact details, training records and incident notes are personal information under the Privacy Act 2020. Keep them in a restricted collection that only the owner and manager can search.

Decisions that stay on the farm

The assistant finds records. People decide what to do with them: animal health and treatment choices with the vet, whether milk or stock can go, what to spray and when, and what to file in NAIT. A record that looks complete can still be wrong, so the manager reads the source before acting. Our guide to building a searchable business knowledge base covers how to decide what goes in first.

Where farm record search goes wrong

• Stale exports. An export from the herd system that last ran three weeks ago gives a confident, outdated answer. Show the refresh date of every source next to the answer, and alert when a source stops updating.
• Paddock names. “Paddock 14”, “Back 14” and “the one by the cutting” are the same place. Build a short alias list with the manager in week one, add to it whenever a search misses, and retest using the method in our guide to RAG retrieval evaluation.
• Animal identifiers. Management numbers, ear tag numbers and names in vet notes refer to the same cow in different ways. Map them once and search on all of them.
• Units. A spray rate recorded as litres per hectare in one diary and millilitres per 100 litres in another can be compared wrongly. Keep units with every figure and never let the model convert them silently.
• Treatment date versus clear date. Questions about treated animals can blur when the treatment happened and when it ends. Return the treatment record as written and leave the second calculation to people.

Retrieval sits alongside equipment maintenance logs and compliance evidence collection, which both depend on the same well-tagged records. The agriculture sector guide shows how the three fit together.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.