Knowledge base The curated documents an office AI is allowed to answer from.
Short answer. A knowledge base is a searchable collection of an organisation’s reference material: FAQs, manuals, procedures, policies and troubleshooting guides. Atlassian describes it as a centralised library that helps people find answers without raising a ticket. In AI, the term means the document set a model retrieves passages from before it answers, a method called retrieval-augmented generation. OpenAI’s file search, for example, lets a model retrieve from a knowledge base of uploaded files using semantic and keyword search. For a clinic, the knowledge base might hold the fee schedule, consent forms and front-desk protocol, so the assistant answers from those pages and cites them.
01
What goes in
Content
Good for
Watch out for
Policies and procedures
Staff questions about leave, expenses and escalation
Old versions left in the same folder
Fee schedules and price lists
Quoting consistently
Prices change, so keep one current copy
Client or matter files
Drafting and summaries
Permissions and confidentiality between teams
Past proposals and reports
Reusing wording and structure
Outdated claims copied forward
Meeting minutes and decisions
What did we agree in March?
Unapproved drafts mixed in with final minutes
02
How an AI uses it
The files are indexed so that they can be searched by meaning and by keyword. When a question arrives, the system retrieves the best passages, gives them to the model and asks it to answer from them and to cite them. Our retrieval-augmented generation page explains the method, and OpenAI’s file search documentation describes the same pattern with vector stores.
A knowledge base does not change the model. It changes what the model can see at the moment it answers. Removing a document therefore removes it from later answers, provided the index built from it is purged or rebuilt too.
03
Permissions and bad content
The first risk is showing someone what they should not see. OWASP’s guidance on vector and embedding weaknesses warns that inadequate or misaligned access controls can let a model retrieve and disclose personal or proprietary data, and recommends fine-grained, permission-aware stores. Microsoft’s Copilot Studio documentation shows the pattern: when a knowledge source uses the agent user’s authentication, the agent surfaces only content that specific user can access.
The second risk is bad content. OWASP also lists data poisoning, where manipulated or unverified documents change the answers, and recommends accepting data only from trusted, verified sources and auditing the knowledge base regularly. Old, wrong or hostile text in the folder becomes the assistant’s confident answer.
04
Keeping it healthy
Name an owner for each collection and a review date for each document.
Keep one current version of each policy. Move old ones out of the searchable set.
Tag documents by who may see them, then test with two accounts: one allowed, one not.
Remove a document properly: delete the file, then check that the index and any copies built from it are gone.
Log which passages were retrieved for each answer, so a wrong answer can be traced to its source.
Refresh the index after bulk changes.
05
Where ANVE stands
ANVE answers questions over the office’s own documents and shows the source passage next to each answer, which runs today in the development build. Private GPTs are assistants built on the office’s own material. Full document drafting with design templates is not on ANVE yet, and ANVE is pre-release.
06
Questions people ask
Is a knowledge base the same as a vector database?
No. A knowledge base is the content and its upkeep. A vector database is one technology used to index it so that passages can be found by meaning.
How big can a knowledge base be?
It depends on the product and the hardware. Check the stated limits, and test retrieval on your own documents rather than relying on a headline number.
Who should maintain it?
A named person for each collection, often the author of a document or the team that owns the policy. Without an owner a knowledge base decays.
Can we remove a document later?
Yes. Delete it and purge the index built from it. See our question page on deleting a document from the AI’s memory.
07
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
