Knowledge base

​

Introduction

The knowledge base holds documents the agent draws answers from during a call. The platform splits each document into passages, and during the call it looks up the ones that match the caller’s question. That is why a document can be long: only the passages needed go into the prompt. This is not the same as the Knowledge base box in the template wizard, whose text goes into the prompt whole and is therefore limited to 2 MB.

​

Uploading documents

1. Open Files / Knowledge in the menu and go to the Upload New tab.
2. Drag files onto the box or click Choose files. You can pick several files at once.
3. Each file gets its own progress bar in the list. Two upload at a time; the rest wait in the queue.

Save older Word files (.doc) as DOCX first. A file in another format or over 100 MB shows up in the list with an error straight away and is not sent. If an upload fails (a dropped connection, say), click the retry icon next to the file. Files go straight to file storage, without passing through the application server, so large documents upload faster. Where your network does not allow that, the upload automatically goes through the application server instead. There is nothing to set up.

​

Processing

After upload, the document is processed in the background: converted to text, split into passages and indexed. This takes from a few seconds to a few minutes. You can follow the status in the All Files tab:

• Pending / Processing: the document is queued or in progress,
• Completed: the agent can use it,
• Failed: for example the file is damaged, or it duplicates a document already in your knowledge base (the content is compared, not the name).

​

How a document is split

A passage holds up to about 480 tokens, counted with the same tokenizer the search model uses. That is roughly a page of text. The platform takes care not to cut apart what belongs together:

• A table (a price list, say) goes into one passage when it fits. A longer one is split between rows, and every piece gets the table header, so the row “Oil change | 150 PLN” never loses its price column.
• A list and the paragraphs under one heading are merged as long as they fit into one passage.
• Consecutive passages of the same section overlap by about 15%: the end of one is repeated at the start of the next, so an item at the boundary is whole in at least one of them.

For each question the agent gets the 6 best-matching passages. A 40-item price list fits into two or three passages, so the agent sees all of it.

​

Attaching documents to an agent

Pick ready documents in the Knowledge tab of the agent editor. The agent only uses the documents attached to it.