Merge duplicate customer records without a mistake you can't undo

The same customer shows up five different ways in my data, and I'm not sure the LLM's merge decisions are ones I can trust.

IntermediateRead 8 minDoing: an afternoon8 steps

▶ Start

What you'll have at the end

A merge pipeline that blocks by a shared key, routes unsure matches to a person, and can unmerge any bad call.

Before you start

You need

A customer or contact table where you suspect the same real business or person appears more than once, and a working call to an LLM that already returns JSON matching a schema you define.

Not covered

Building a dedicated record-linkage engine for a table too large to compare pairwise in code, and deciding which of two disagreeing field values should be shown once a merge is confirmed. This covers a table you can iterate over directly, and it stops once every contribution is captured rather than displayed.

Leans on

8 steps

Where to go next

Resubmit a rejected reply with the schema error attached

the comparison call itself fails schema validation instead of returning a clean same-or-different verdict, and you want the model to see the exact error before it tries again.

Checked 18 Aug 2026

Part of the Structured Output Generation with LLMs cookbook