Cloud AI finders

DocView Capture can hand a document to the AI services of Microsoft, Amazon and Google, or to a language model, and use what they read as finder values. Five finders do this:

Every finder gives each value it reads as an output (for example Azure.InvoiceTotal). Line items become the output Items, which fills a table field.

Your data: these finders send the pages (Comprehend: the text) to the provider, so the data leaves your premises. Each provider is off until a site administrator switches it on under System › AI connections. Check that the provider's terms and region suit your data before you switch it on.

Set up the connections

Go to Administration › System › AI connections (site administrators). Each provider has a card:

1. Enter the keys (and the region, project or address).
2. Switch the provider On.
3. Click Save, then Test. The test sends no document.

Keys are encrypted and never shown again. To change a key, type the new one; empty boxes keep the saved key.

Azure AI Document Intelligence

1. In the Azure portal, create a Document Intelligence resource in a region near you.
2. Open it, go to Keys and Endpoint, and copy the endpoint and KEY 1 into the Azure card.

The Azure OCR engine uses the same endpoint and key.

AWS (Textract, Comprehend, Bedrock)

1. In IAM, create a user (for example docview-ai) with access keys for an application outside AWS.
2. Give it a policy that allows these actions: textract:AnalyzeDocument, textract:AnalyzeExpense, textract:AnalyzeID, comprehend:DetectEntities, comprehend:DetectPiiEntities, comprehend:DetectDominantLanguage, bedrock:InvokeModel.
3. Choose the region and enter the access key ID and secret access key.

When DocView runs on AWS (EC2, ECS) with a role, leave the keys empty and the role is used.

For Bedrock models, request access to the models you want (for example Claude or Amazon Nova) under Bedrock › Model access, in the same region.

Google Document AI

1. In the Google Cloud console, enable the Document AI API.
2. Create a service account with the role Document AI API User. Add a JSON key and paste the file's contents into the Google card.
3. Choose the location (eu or us). The project is taken from the key file unless you enter one.
4. Under Document AI, create the processors you need (Form parser, Invoice parser, Expense parser, ID parser or Custom extractor) in the same location. Each finder needs the ID of its processor.

AI models (Claude, OpenAI, Gemini and others)

Enter the keys of the services you use:

AWS Bedrock uses the AWS card's keys. DocView OmniRead needs no key; its vision model must be allowed under System › DocView OmniRead.

Try it in two minutes

Administration › Batch Setups › New › System features has a ready-made template for each finder:

• Azure AI Document Intelligence
• AWS Textract
• Google Document AI
• Ask an AI model
• Find personal data (AWS Comprehend)

Each comes with sample invoices. Create one, switch its provider on, import the samples, and open a document. The invoice number, date, supplier, total and line items are filled by the service.

Add a finder

On a document class's Finders tab, click New finder and choose the method. How to use fills in a working example.

Azure AI Document Intelligence

• Model: the kind of document. With a prebuilt model, leave the Values tab empty to get every value of the model, or list only the ones you want: Supplier | VendorName. Values inside other values use a path, such as Employee.Name.
• Your own custom model: choose it and enter its model ID from Document Intelligence Studio.
• Layout model: gives the key-value pairs of any document and its first table. With Ask for the values by name, Azure also looks for each value on the Values tab by its name (query fields).

AWS Textract

Choose What to read:

AWS Comprehend

• Choose Personal data or Names, organisations, places, dates.
• List the types you need: Email | EMAIL, NiNumber | UK_NATIONAL_INSURANCE_NUMBER. Leave the list empty for the usual ones.
• Every match gives all of them, separated by ; .
• Personal data can be found in English and Spanish text.

Google Document AI

• Choose the processor type and paste the Processor ID.
• The invoice, expense and identity parsers give their values without setup.
• For the form parser and custom extractors, list the values: PolicyNumber | Policy number, or the custom extractor's field name.

Ask an AI model

1. Choose the model provider and, optionally, the model. Empty uses the provider's default.
2. On the Questions tab, write one line per value: Name | question | optional pattern.
3. Say exactly how to answer, for example Answer as DD/MM/YYYY.
4. Optionally list the line items columns (Description | the item or service). The model then lists every line item as Items.

All questions go to the model in one request. Extra instructions (Advanced tab) apply to every question. Also send the text DocView read helps with small print but costs more.

Line items into a table field

Give the table field the finder value Items (for example Azure.Items). Each table column takes the service's column with the same name or label, or a common synonym:

Any column left over takes the service's column in the same position.

Tips

• Cloud finders take a few seconds per page and cost money per page. Use them for the values your own finders cannot read, or as a fallback finder.
• Run test on a sample shows exactly what the service returned for each value, with its confidence.
• Lowest acceptable confidence leaves a value empty when the service is unsure, so a person checks it.
• If a finder says its service is switched off or not set up, ask a site administrator to complete System › AI connections.

Related pages

• Finders and zones
• Ask AI
• OCR engines and languages
• DocView OmniRead