Designing a Knowledge Base
Organise a knowledge base into owned collections, separating static, dynamic, internal, external and restricted knowledge.
A pile of documents is not a knowledge base
After ingestion you have thousands of clean documents. It is tempting to throw them all into one index and start searching. That works for a demo, and then problems appear. The student assistant quotes a staff-only circular. The 2023 leave policy outranks the 2026 one. A question about hostel fees returns the placement cell's fee for a mock test. Nobody knows who should fix a wrong answer, because nobody owns the document it came from.
A knowledge base is a collection of documents organised for answering questions: grouped sensibly, labelled consistently, owned by someone, and kept current. Designing it is less about technology and more about deciding how knowledge is organised, who it is for, and how it changes. A little design at this stage prevents many retrieval problems later.
Collections: grouping what belongs together
The first decision is how to split the knowledge into collections, groups of documents that are searched, managed or protected together. Collections usually follow one of three lines:
| Split by | Example collections | Good when |
|---|---|---|
| Department or owner | Academics, Hostel, Placements, IT help | Different teams maintain different documents |
| Audience | Student-facing, Staff-only | People must see different things |
| Document type | Policies, FAQs, Forms, Circulars | Types need different handling (e.g. forms vs rules) |
In practice, a knowledge base often combines these: collections by department, with an audience label on every document. Collections make it easy to search only where it makes sense (a hostel question does not need to search placement records), to give each team control over its own content, and to apply different rules, such as refreshing exam notices every hour but archived policies never.
Collections follow ownership, and a separate restricted collection keeps staff-only material apart.
Static and dynamic knowledge
Not all knowledge changes at the same speed, and the knowledge base should treat it differently.
Static knowledge changes rarely: academic regulations, the anti-ragging policy, how to apply for a bonafide certificate. It can be refreshed weekly or when someone edits it.
Dynamic knowledge changes often: exam timetables, hostel closure dates, placement drive schedules, today's mess menu. It needs frequent refreshes, clear dates on every document, and rules for retiring old versions, otherwise the assistant confidently answers with last month's timetable.
Some dynamic data should not be in the text knowledge base at all. A student's own attendance percentage lives in a database and is better answered by querying that database directly than by searching documents.
Internal and external knowledge
Most knowledge bases are internal: the organisation's own documents. Sometimes useful knowledge is external: a university's affiliation rules, a government scholarship scheme, an accreditation body's guidelines. External sources can be very helpful, but they bring questions of their own. Who checks they are current? Is the organisation allowed to copy them? Should answers clearly say "according to the university's rules" rather than "our rules"? Keeping external material in its own collection, labelled as external, makes these questions easier to manage.
Owners and update rules
Every collection needs an owner: a person or team responsible for what is in it, for keeping it current, and for fixing it when an answer is wrong. Without owners, a knowledge base slowly fills with outdated, duplicated and contradictory documents, and quality decays.
A simple way to capture all of this is a short design record for each collection. The program below shows one, and checks each collection against two basic rules.
Expected output:
Regulations OK
Exam notices OK
Hostel FIX: no owner
HR policies OK
Mess menu FIX: changes often but refreshed monthly
Each collection is a dictionary of design decisions. The loop applies two rules to it and gathers any failures in problems: an empty owner, and a dynamic collection whose refresh is only weekly or monthly.
Two collections need attention before launch: hostel content with nobody responsible for it, and a daily mess menu refreshed only once a month.
Design decisions in practice
Try placing documents into a sensible design below.
A good knowledge base design fits on one page: a list of collections, each with an owner, an audience, a refresh rule and a note on whether it is static or dynamic. Writing it down before building anything turns many future arguments ("why did the bot say that?") into simple questions with clear answers ("which collection did it come from, and who owns it?").