Hi guys,
I was experimenting with Glide’s native Ai tools, my idea was to extract data from a photo of an invoice with the Image to text tool, the idea is to extract the invoice number, the supplier, the amount and a description, it can be done by creating an Image to text column to extract only the information you need, but this in the future may affect the application because each column would be using 2 updates, in my case there would be 4 columns (8 updates), this would represent a problem in the future because according to the client’s requirements they may increase and this would mean greater consumption, I managed to find a way to use a single column of AI where with a prompt I can extract the necessary information and use it as a base, here is the step by step:
Prompt: Extract the following data from the invoice.
Return ONLY plain text in this exact format:
Numero de Factura:
Proveedor:
Monto:
Descripcion:
If a value is missing, leave it empty.
Do not add extra text.
Step 1: Create a Split Text type column configured with a line break.
Step 2: Create a column of Single Value Type where each line of the Split is Extracted. In my case it is configured from the beginning and I indicate which row I need to Isolate, and so on for each data, in my case there are 4 columns of Single Value for each data (invoice number, supplier, amount, description).
Step 3: Add a column of type Replace ALL, where I replace the title with an empty value in order to obtain only the value, example: In my single Value I have the Value: Supplier: Multitextil, in my Remplace All I replace the text “Supplier:” with an empty field and as a result I get “Multitextil”, so on according to the fields you need.
The prompt can be improved in order to directly obtain only the result without the need for titles, but the Ai tool can confuse the fields, the solution I have found so far is to clarify it as much as possible,
With this, it is possible for the user to add an expense record with only a photo of the invoice, thus improving the user experience with the lowest possible update cost(Only 2 Updates per image).
I attach an example.
If the way of extracting can be improved, you can collaborate.
