How it works
- You add documents: a file, a pasted text, a web page or pages from Notion.
- Pawabase reads each one, splits it into passages and turns each passage into a vector with your embedding model.
- When someone searches, Pawabase finds the passages closest in meaning and those that match the words, and returns the best.
- An agent given the knowledge base searches it as one of its tools, and answers using what it found.
Create one
You need an embedding model first (see Providers and models). Then:1
Open AI → Knowledge → Create a knowledge base
Give it a name such as Help articles and say what is in it.
2
Choose the embedding model
Under Using which model? pick the embedding model. Keep the same model for the life of the base: its documents are indexed with it.
3
Choose who can search it
Under Who can search it? pick an access rule. By default only your servers can.
4
Add your first document
See Add documents.
Use it
Private documents for each user
A signed-in user can add documents that only they can find, for example their own notes or uploads. Their documents are kept out of everyone else’s searches and out of an agent’s answers for other people. Your servers can also add a document for one user.Keep it good
- One topic per document retrieves better than one giant file.
- Update, don’t duplicate. Give each document a
source(such as its URL or file name). Adding the same source again replaces the old version. - Try real questions in Try a search before giving the base to an agent, and fix gaps by adding or rewriting documents.
- Write documents as answers. “Refunds take 5–7 business days” retrieves better than a vague paragraph.
Add documents
Files, text, web pages and Notion.
Agents
Give a knowledge base to an assistant.