How it works

From an agency file to a search result

An authority loads its own notices. A person maps the columns, by hand or from a suggested sample, then saves the map. Search reads those saved records.

FAQ

What is the difference between a notice and a recall?

A notice is the published announcement. A recall is the safety action that notice is about. Search, the ledger, and the detail page show notices. When we say recall, we mean stop using the product, return it, or get it fixed.

Notice

The published announcement

Title, date, product, hazard, and remedy as the agency wrote them. Each row you search or count is one notice. A later update from the same agency is a later notice.

Recall

The safety action

The notice is how the agency tells the public about that action. One notice describes one recall as it was published that day.

Agencies upload their data

  1. 01

    Sign in

    The agency desk is the only place a source is added.

  2. 02

    Name the source

    Country, public site, and whether notices arrive as a file or a dated API.

  3. 03

    Map the columns

    Record id and title are required. Everything else can stay blank.

  4. 04

    Load notices

    A file is a batch of new and changed rows. A dated API pulls on its own schedule.

Open the agency desk

Two ways to map columns

Both end at the same saved map. The upload step then tells you to rename the unique-id column and the title column to those saved names.

Static

A person writes the paths

  1. 1

    Open the field map

    Each shared field has a box for that source’s column name.

  2. 2

    Type a path, or upload a map file

    A path looks like Subject, or Hazards[].Name when the file is nested.

  3. 3

    Save

    No model is called. The saved names are what a later upload must use.

Suggested

A model proposes paths from the file

  1. 1

    Pick a notice file

    The file stays in the browser. Notices are not loaded yet.

  2. 2

    Read headers and a few cells

    Up to 60 column names. Up to three non-empty cells from the first 20 rows. Each cell is cut at 120 characters.

  3. 3

    Ask AI Model

    One prompt: the shared fields, plus each header with those sample cells. The answer must be the header, not the cell text.

  4. 4

    Try the next AI model only if that call fails

    If the first call fails, ask the next AI model. The first usable reply wins.

  5. 5

    A person confirms

    Unknown fields and invented column names are dropped. Confirm and save is the only step that writes the map.

How search works

One query. Each tab runs only when you open it. Text does not call a model. Similar and Photos compare embeddings that were stored when the notice was indexed.

  1. Text

    Word match

    The query is matched against the published fields in the database. No model call.

  2. Similar

    Meaning

    The query is embedded and compared with the vector stored for each notice.

  3. Photos

    Pictures

    A photo, or the words, are compared with vectors of the published product photos. This tab runs only when you open it.

Search published records