Knowledge

What the assistant knows before it calls anything.

Knowledge is what the business has published: uploads, crawled pages, FAQs, the company profile, and hand-written cards. Retrieval is hybrid — vector search plus Postgres full-text, fused by reciprocal rank fusion — and runs once per turn before the model sees anything.

Knowledge is not your data

Never push your live records into knowledge. Prices, stock and order status belong in tools, not in an index.

Indexed text is a snapshot. A price embedded on Tuesday is quoted on Friday with total confidence and no way for the model to know it has changed. The assistant is explicitly told that knowledge is background and that live figures come from tools.

Use knowledge for what is semantically true and slow-moving:

Belongs in knowledge	Belongs in a tool
Delivery and returns policy	Where this order is
Care instructions, size guides	This design’s price
Who the business is, opening hours	Whether it is in stock
”How do I…” explanations	This buyer’s cart

Searching from a turn

An assistant with knowledge sources gets search_knowledge, which it can call mid-conversation like any other tool. You do not need to expose knowledge yourself.

The context budget

Everything competes for one budget: system prompt, knowledge, memory, conversation history, and your tool results. When it overflows, the platform drops in a fixed order — memory first, then knowledge, then it trims the summary, then the oldest messages. The system prompt and the incoming message are never dropped.

Long tool results consume that budget. A 4,000-character catalogue dump can push out the knowledge the assistant needed to answer the next question. Return what was asked for.