Knowledge Base
PickYour documents, organized so an AI can look things up in them and answer with the facts you wrote, not with guesses.
01 · Introduction
What a knowledge base is
A knowledge base is a collection of documents an AI can consult: policies, manuals, product sheets, past answers, meeting notes.
A model on its own only knows what it learned during training, which is general and out of date. Connected to a knowledge base it works like an open-book exam: when you ask a question, the system first finds the relevant pages, then the model answers from them and can cite them.
This technique is called RAG (retrieval-augmented generation). You will see the term everywhere; it just means "look it up first, then answer".
02 · When to use it
When you need one
- Support or sales questions that are answered somewhere in your documents, if only people could find them.
- Internal "how do we…" questions: expenses, onboarding, tools, processes.
- A product with a long manual or changing specifications.
- Anything where a wrong answer from memory is worse than "I could not find that".
For a dozen short documents, simply attaching them to a chat Project is enough. A knowledge base earns its keep from a few hundred pages upward.
03 · Setup
Building a small one
You can start without writing code:
- Collect the documents into one folder. Convert scans to text; delete duplicates and outdated versions.
- Choose a tool. Easiest: a Claude or ChatGPT Project (upload files, ask questions). Next step: a knowledge tool such as Notion AI or a no-code RAG service, which indexes a folder or a site. Full control: a vector database plus the API.
- Index. The tool splits documents into chunks and stores them for search.
- Test with ten real questions and check each answer against the source page.
- Decide what the assistant says when it finds nothing. "I could not find this in the documents" is the right answer; a guess is not.
04 · How to use it
Keeping it useful
- Give every document an owner and a review date; stale pages produce confident wrong answers.
- Log the questions people ask. The ones with no good answer tell you what to write next.
- Keep one source of truth per topic; three versions of the holiday policy means three different answers.
- Ask the assistant to quote and link its source so readers can check.
- Re-index after big changes; most tools do it automatically on a schedule.
05 · Pitfalls
Pitfalls
- Garbage in, garbage out. A messy document set gives messy answers; cleaning the documents is most of the work.
- Search finds similar text, not correct text. A page that looks relevant but is outdated will be used.
- Tables, scanned PDFs and slides often lose their structure when split into chunks; convert the important ones to clean text.
- Access control: if the knowledge base can read a confidential document, the assistant can repeat it to anyone who asks.
06 · Resources
Where to go next
The Context and Memory handbooks explain what the model actually sees and what it keeps between sessions.