Knowledge Hub settings

Nine settings that change how the Knowledge Hub finds and answers from your documents. Every one is off until you turn it on, and each is a switch, a dropdown, or a sentence in your own words. No technical knowledge needed.

Before you start

Most of these live in one place. In the app, open Knowledge Hub in the sidebar, then Settings. You will see three tabs across the top:

• Answers - how questions are answered.
• Reading documents - what happens to a file as it comes in.
• Models & storage - which AI models are used and where the index lives.

Three settings are per knowledge base rather than workspace-wide. For those, open Library, click into a knowledge base, and look for the Tuning and Fields to pull out cards.

Changing any of these needs the Knowledge Base admin permission. Without it you can read the settings but the controls are greyed out.

1. Result ranking

Settings › Answers. Uses AI credits on every question.

Searching is good at finding passages that share words with a question. It is not good at telling which of those actually answers it. Turn this on and AI re-reads the best passages before replying, then puts the useful one first.

Steps

1. Open Knowledge Hub → Settings and stay on the Answers tab.
2. Find the Result ranking card and turn the switch on.
3. Two more controls appear. Leave Passages to re-read at 24 to begin with. The range is 10 to 60 - more passages catch answers buried deeper and cost more per question.
4. Set Model to the smallest one in the list. Ranking is a simple judgement, and a small model does it just as well for a fraction of the cost.
5. Click Save.

What changes. Someone asks "how much holiday do I get?". Without ranking, the top result might be "Holiday cover must be arranged with your line manager two weeks in advance" - the same words, the wrong answer. With ranking, the entitlement clause comes first.

2. Follow-up questions

Settings › Answers. Uses AI credits on every question.

People do not repeat themselves. They ask something, then a short follow-up that only makes sense in context. Searched literally, that follow-up finds nothing.

Steps

1. On the Answers tab, find Follow-up questions and turn it on.
2. Pick a small, fast Model. Rewriting one line is a light job.
3. Click Save, then test it in chat with a two-part conversation.

What changes. Ask "how much holiday do full-time staff get?", then "what about part-timers?". The second question is searched as "How much holiday do part-time staff get?" instead of as that fragment. It also searches the more formal wording your documents are likely to use, so a casual question can still find a formally written policy.

3. Passage context

Settings › Reading documents. Uses AI credits once per document, at upload.

Long documents are cut into passages, and the Knowledge Hub searches passages rather than whole documents. A passage taken from the middle of a contract often makes no sense on its own. This adds one line to each passage saying where it came from.

Steps

1. Open Knowledge Hub → Settings and click the Reading documents tab.
2. Find Passage context and turn it on.
3. Pick a small Model.
4. Click Save.
5. Upload a document. This only applies to documents added after you switch it on.

What changes. A passage reading "This must be renewed every 24 months." becomes "From Section 7 of the Acme maintenance contract, about certification renewal. This must be renewed every 24 months." The first is meaningless alone; the second can be found and understood.

4. Glossary

Knowledge Hub › Glossary. Free - no AI credits.

Your people and your documents use different words for the same thing. Someone types "CTO"; the handbook says "chief technology officer". Nothing knows those are the same at your company until you say so.

Steps

1. Open Knowledge Hub → Glossary and click Add term.
2. In Term, type what people ask - for example CTO.
3. In Also search for, type what your documents actually say, one phrase per line - for example chief technology officer.
4. Add a Note if you want a reminder for whoever maintains the list. It has no effect on search.
5. Click Save.

Good rows to start with:

• PTO → paid time off, annual leave
• T&Cs → terms and conditions
• the portal → customer self-service portal

Matching ignores capitals and only matches whole words, so CTO matches "cto" but not "DIRECTOR" or "CTOs". A workspace can hold up to 500 terms, each with up to 10 phrases.

5. What is in here

A knowledge base › Tuning card. Free - no AI credits.

Documents are cut into passages, and the right size depends on what the document is. Cut a contract at the wrong point and a clause is split in half. Cut a transcript into contract-sized pieces and every answer drags in three unrelated speakers.

Steps

1. Open Knowledge Hub → Library and click into a knowledge base.
2. Find the Tuning card.
3. Under What is in here, pick what this knowledge base mostly holds.
4. Click Save, then upload documents.

The choices, and what each does:

• Contracts and agreements - small pieces, so one clause tends to be one passage.
• Policies and handbooks - medium, so a whole section stays together.
• Manuals and procedures - medium with more overlap, so a sequence of steps is not broken up.
• Transcripts and call notes - shorter, suited to back-and-forth speech.
• Questions and answers - smallest, so one question-and-answer pair is one passage.
• Mixed documents - the default, and the right choice when a knowledge base holds a bit of everything.

6. Search balance

A knowledge base › Tuning card. Free - no AI credits.

Two searches run at once on every question. One matches the exact letters you typed. The other matches what you meant, which is how it knows "car" and "vehicle" are related. This setting decides which carries more weight.

Steps

1. In the same Tuning card, find Search balance.
2. Choose Favour exact wording for knowledge bases full of part numbers, product codes, invoice references, or names.
3. Choose Favour meaning for ordinary writing - policies, handbooks, guidance notes.
4. Leave it on Balanced if you are unsure. That is the default and it is a sensible one.
5. Click Save.

Why it matters. Searching a parts catalogue for "QX-940", meaning-matching returns QX-930 and QX-950 too, because as words they are nearly identical. Exact wording returns the right one. The other way round, someone asking "can I work from home?" will never match a policy headed "Remote working arrangements" on exact letters alone - only meaning finds it.

7. Document dates

A document's details window. Free - no AI credits.

Every company has three copies of the same policy and only one of them is current. Without this, the Hub will quote the 2023 version with a confident citation. Give a document an expiry date and it stops being used to answer questions, while staying in your library for the record.

Steps

1. Open Knowledge Hub → Library and click into a knowledge base.
2. Find the document in the list, open its row actions and choose Details.
3. Scroll to Dates. Set any of the three:
   • In use from - the Hub ignores the document until this date.
   • Expires - after this date it stops being used in answers.
   • Review by - a reminder date for whoever owns the document.
4. Click Save dates.

What changes. A price list valid for 2026 gets a start of 1 January and an expiry of 31 December. It stays silent before and after, without anyone having to remember to remove it. An expired document shows a clear note in its details window saying it is not being used.

8. Saved answers

Knowledge Hub › Saved answers. Free, and it saves credits.

Some questions get asked constantly and have exactly one right answer. Searching documents for those is slower, costs credits, and can still get the wording slightly wrong. Write the answer once instead.

Steps

1. Open Knowledge Hub → Saved answers and click Add answer.
2. In Question, write it the way people actually ask it.
3. In Answer, write exactly what should be said. It is shown word for word.
4. Leave Use this answer now off while you get the wording right. Nothing goes live until you turn it on.
5. Turn it on and click Save.

For example, the question "What is our refund window?" with the answer "Thirty days from the delivery date. After thirty days, refunds need manager approval." Each row shows how many times it has been used, so you can see which are earning their place.

Questions nobody could answer

The second tab on that page lists questions people asked that your documents could not answer, with the most-asked at the top. Each row has a Write an answer action that carries the question straight into a new saved answer, and a Mark handled action to hide one without writing anything.

Turning it on: open Settings → Answers, find Unanswered questions, turn it on and save.

9. Fields to pull out

A knowledge base › Fields to pull out card. Uses AI credits once per document.

Name the details that every document in a knowledge base has, and AI reads each new document for them. This is what turns a pile of contracts into something you can see at a glance.

Steps

1. Open a knowledge base and find the Fields to pull out card.
2. Click Add field.
3. Type the field name as you would say it - for example Counterparty.
4. Choose the kind: Text, Date, or Number.
5. Repeat for each detail, up to twelve, then click Save.
6. Upload documents. This applies to new ones only.

A contracts knowledge base might use Counterparty (text), Effective date (date), Renewal date (date), and Annual value (number). To see what was found, open a document's Details window and look for Details read from this document.

What each one costs

Everything on this page is off by default, so nothing spends credits until you decide it should.

• Never costs anything - Glossary, What is in here, Search balance, Document dates.
• Saves credits - Saved answers, because a matched answer skips the search entirely.
• Costs a little on every question - Result ranking, Follow-up questions.
• Costs once per document at upload - Passage context, Fields to pull out.

For the four that use AI, you choose the model. Picking the smallest one in the list is almost always right: these are simple jobs and the price difference between models is large.

Where to start

If you only do two things, do the two that cost nothing and fix the most common complaints:

1. The Glossary - an afternoon's work, free, and it fixes "why can't it find that?" more often than anything else here.
2. Saved answers - write your ten most-asked answers by hand. Free, instant, always correct.

Then turn on Result ranking and Follow-up questions. Leave Passage context and Fields until you have documents worth re-uploading, since they only apply to what you add afterwards.