Create a knowledge base.
**Write `description` — it is what routes questions to this base.** Before each answer the
platform reads every attached base's `description` and decides which ones this question needs;
a base whose description is vague or blank gets searched when it should not, or missed when it
should. One line, in the words a visitor would use, naming the subject **and the boundary**:
"Country-by-country medical device registration requirements — not company or pricing
information." It is also the line shown in the console list, but that is the lesser job.
**Write `instructions` at creation time — do not leave it blank.** It is injected next to this
KB's excerpts whenever they are retrieved. It does NOT affect recall (recall is vector search +
max_distance); it governs how the model USES what was retrieved. Derive it from the expected
usage, one line each:
- Scope: what it covers / does not, and what to do when out of scope
("Covers residential mortgages only; for car or personal loans, say so and hand off").
- Authority: where it ranks ("Current company policy; overrides industry norms").
- Usage rules specific to this content ("Any quoted rate must state its effective date").
The generic "answer from the excerpts; say when not covered" is built into the platform — never
repeat it here. Blank is acceptable only for generic reference material with no special rules.
Worked examples: search_agent4_docs("knowledge base instructions examples").
max_distance is the relevance cutoff (blank = global default 0.6); tighten for regulatory
content, and calibrate against real queries rather than guessing.
top_k is how many passages this base contributes per turn (1-10, blank = 3). Set it from the
**shape of the documents**, not from how important the base feels: if a typical document is
five passages long, three of them is a third of a record, and the model answers from a
fragment. Measured on an 8,500-record regulation base whose median document is 5 passages —
at k=3 the right record was retrieved 36/40 times but only 4 answers had enough to go on; at
k=6, same hit rate, 12. Going to 10 added 3 more and 77% more material. More is not free:
material in the prompt suppresses tool calls, dose-dependently.
