Which documents to send to Dri and how to write them

Where to find it

In your account menu, under Dri's knowledge, Documents tab, Upload document button. That's where you choose the file, the Title (optional), the Type and the Source of each document.

Which file formats Dri's knowledge base accepts

The Upload document window accepts:

• PDF: with text or scanned;
• Word (.docx);
• spreadsheet (.xlsx);
• text (.csv, .txt, .md) and HTML;
• image (PNG, JPEG, WEBP).

Each file can be up to 20 MB, and the organization can have up to 200 documents in the knowledge base. Beyond that, the screen rejects the upload and says why.

An .md or .txt file can also become a Help Center page, like text written on the screen: open the document and use Publish to the Help Center. It must be saved in UTF-8 and have up to 100 thousand characters. PDF, Word, spreadsheet and image Dri reads, but they don't become pages.

You can also drag the file onto the Documents list: the Upload document window opens with it already chosen.

The .md file that comes already classified

An .md or .txt file can bring, at the top, a header with its classification. That way you don't choose anything in the window: drag, click Upload, and the document goes in classified (and published, if the header asks for it).

---
titulo: Orders and ERP integration
tipo: perguntas-frequentes
origem: curado
quem_pode_ler: equipe-de-vendas
publicar: sim
---

## Do you send the order to the ERP?

Yes. ...

• titulo: the document's title.
• tipo: organizacional, politica-comercial, ficha-tecnica, perguntas-frequentes or outro.
• origem: coletado or curado.
• quem_pode_ler: todos, equipe-de-vendas or administradores.
• publicar: sim or nao (publishes to the Help Center on upload). With sim (yes), the quem_pode_ler line is required.

All lines are optional; whatever is missing follows the window's choice. What the header says outweighs the window, and the window shows what the header has already decided. The template is also in the window itself, under View header template.

To change who can read or the publication of a document already uploaded, fix the header and drag the same file again: identical text doesn't become another document, and what the header says about quem_pode_ler and publicar is applied to what already exists. The file with a header must be saved in UTF-8: otherwise, the header can't be read and the upload is rejected (in Notepad, use Save as and choose the UTF-8 encoding). A key or a value the screen doesn't know makes the upload be rejected, with the line of the error: a half-read header would upload the document for people who shouldn't read it. The header doesn't go into the text: Dri doesn't read it and the page doesn't show it.

Scanned PDFs and images are read by vision: the text that appears on the page is transcribed as it is, not interpreted. ⚠️ In a very long scanned PDF, only the first 30 pages are transcribed: if the important material is at the end, split the file or send the version with text (exported from Word, not scanned).

Upload several .md files at once, in a .zip

To upload many texts at once, put the .md or .txt files in a .zip and upload the .zip through the same Upload document window (or drag it onto the Documents list). Each file inside the .zip goes in as a separate document, with the classification from its own header.

• Up to 50 files per .zip, and up to 20 MB in total.
• Only .md and .txt files go in. PDF, Word, spreadsheet and image inside the .zip appear as ignored: upload those one at a time, through the window.
• A file without a header uses the Type and Source chosen in the window, and the title is the text's first # Título (or the file name).
• A rejected file (one with a header error, for example) doesn't stop the others from going in.

After the upload, the window shows what happened to each file: Added (and whether it was published to the Help Center), Already in the knowledge base, Rejected or Ignored, with the reason.

Sending the same .zip again is safe: text that was already in the knowledge base doesn't become a repeated document, and what the header says about quem_pode_ler and publicar is applied again. To change only who reads or the publication, fix the header and upload the whole .zip again.

Fixing the text is different: the file with the changed text goes in as a new document, and the old one stays in the knowledge base (and published, if it was). When another document with the same title already exists, the report line warns you: remove the old one from the Documents list.

How to organize the folder before compressing:

1. One folder per set (for example, the company or the subject), with one file per subject: one about orders and ERP, another about commercial terms, another about customer registration. A short, single-subject file answers better than a huge file.
2. Every file with a complete header, saying who can read it. That way nothing depends on what is chosen in the window.
3. A simple file name, without accents, that says the subject (pedidos-e-erp.md). The name appears in the upload report.
4. Before publishing to all, check that the text has nothing the buyer shouldn't read: negotiated price, discount, customer credit, personal contact.

Formats the knowledge base doesn't read yet, and what to do

The screen rejects them right away, with the recipe for what to do:

Uploading the same file twice doesn't duplicate anything: the screen warns "This file is already in the knowledge base: nothing changed."

What kind of document is worth sending

Dri's knowledge base is for what is durable and specific to your company: what a customer asks where the answer doesn't change from one week to the next:

• freight policy, delivery times by region, minimum order;
• branches, addresses, hours, who handles what;
• payment terms, warranty, exchanges and returns;
• discount policies with no deadline, written as Curated;
• product specifications and compatibility;
• frequently asked questions the team already answers off the top of their heads.

What's not worth sending to Dri's knowledge base

• Product price, stock and time-limited promotions. The reading separates these passages on purpose, into the Volatile drawer: if stored, they'd become answers with yesterday's number. Dri never gives a product price or stock from the knowledge base: that comes from the catalog, the price tables and the ERP.
• The product catalog. Product registration comes in through the catalog import, not here. See Import the catalog by spreadsheet.
• Advertising material without facts. Slogans and institutional text become Noise and answer no question at all.

How to choose the document's Type

The Type changes how the text is read and where Dri looks:

When a customer asks a technical question, Dri looks in the technical sheets and the frequently asked questions. When the question is commercial or about the company (payment terms, freight, payment, branch), she looks in the commercial policies, the organizational documents and the frequently asked questions.

⚠️ A document of type Other stays out of both searches and is only consulted when the question doesn't clearly fit either of them. Always prefer a specific type. FAQ goes into both searches.

Collected or Curated: which Source to choose

• Collected (the default): material that already existed: supplier catalog, manufacturer manual, sales presentation. The reading is stricter: what it can't classify goes to Review.
• Curated: text your company wrote for this purpose. The reading is more permissive (when in doubt, it becomes a fact) and it gets preference when it contradicts a collected document.

In a Curated document, a commercial policy with no deadline goes to Fact, even with a percentage or amount: "orders placed through the site get a 10% general discount", minimum order, payment condition. Dri states the rule as it is written. With a deadline ("10% until Friday"), the passage goes to Volatile. In a Collected document the reading is stricter, and every discount goes to Volatile.

Use Curated only for what was actually written for Dri. If everything is curated, all weights become equal and no contradiction resolves itself. See Knowledge base conflicts: when two documents disagree.

How to write a document so Dri answers well

1. One subject per section, with a heading. Use Word's headings (Heading 1, Heading 2) or # in an .md file. The heading above each passage is what tells the reading what it's about: the same address under "Showroom" and under "Rio Branch" are different things.
2. Complete sentences, with the subject. "Freight is free for orders over R$ 800 in the Southeast state capitals" answers on its own; "over 800, free" doesn't.
3. Condition and exception together. "We deliver in 3 business days, except on regional holidays". Splitting the two onto different pages makes Dri answer only half.
4. One value per subject. If the delivery time changed, update the document and remove the old one; two documents with different times become a conflict.
5. In a technical sheet, write the product code exactly as in the catalog. When the code matches a product in some store of the organization, the specification starts enriching that product; a different or approximate code stays only as general knowledge.
6. In a spreadsheet, the first row of each tab is the header. The tab name becomes the block's title, and each table is read whole: without a header, a row like "Salvador | BA | 3 days" means nothing.

Common questions

Can I send the same PDF to several stores? No need: the knowledge base belongs to the whole organization, and Dri uses the same document for all stores.

I sent a document with prices. Will it get in the way? No. The price passages go to Volatile and stay out; the rest of the document goes in normally.

I fixed the file. How do I update it? Upload the new version and remove the old one. The reprocess button rereads the stored original file, not a new file. See After uploading: the processing and the details of a document.