An answer is only as useful as the context behind it.
Why connecting documents is only the beginning—and why source, ownership and access belong in the same conversation.
Useful knowledge needs a source, an owner and a boundary.
Knowledge for employees and AI teammates is in development. This note describes the design direction, not live availability.
The difficult question is “which version?”
Imagine a colleague asking how to handle a customer exception. There may be a policy document, an old discussion and a newer decision recorded in a meeting. Finding text that mentions the exception is easy compared with deciding which source should guide the response.
For a support colleague, the useful result is an answer they can stand behind: the current policy, the reason it applies and the person to ask about an exception. A search result that merely resembles the question leaves that verification work with the colleague.
For a business team, the cost of getting this wrong is straightforward: someone acts on an outdated answer, or spends time verifying an answer that came without a source. The product has moved the effort around instead of reducing it.
Take this with youBefore adding more documents, decide who maintains the answer and which source should be trusted.
From a service policy to a supported answer
Find the source
Start with a maintained policy or document. Know who owns it and when it was updated.
Under the hood
Preserve document identity and source metadata during ingestion so an answer can point back to the material it used.
Organise context
Group useful material by the work it supports. An accounts role and a support role need different context.
Under the hood
Collections provide a way to organise documents and their relationship to the roles that use them.
Check access
Ask which information the person and role should be able to use before preparing an answer.
Under the hood
Permission checks belong in the retrieval path. Similarity ranking cannot make an authorisation decision.
Ground the answer
Prepare an answer linked to its source. Flag missing or conflicting information for review.
Under the hood
Retrieval supplies context to generation. Evaluate source relevance and answer grounding separately; neither is guaranteed by a fluent response.
Give documents a place in the workflow.
In UseGlic, sources, document versions and collections have distinct roles. For the support-policy example, that means preserving the connection to the source and organizing the material around the questions a support colleague actually needs to answer.
A changed policy also needs a deliberate transition. If an old policy and a new draft both exist, the newest text is not necessarily the approved answer. Source version, publication state and ownership give the team ways to manage that distinction.
AI teammates can use source context to prepare answers that help someone check the underlying material. A source reference is useful because it lets the reader inspect the policy, notice a conflict or ask the right owner for a correction.
Under the hood +
A retrieval pipeline separates ingestion, document organisation, authorisation, ranking and generation. Embeddings help find related passages; they do not replace source metadata or access checks. Test retrieval with realistic questions, including questions that should have no answer.
“I need confirmation” can be a useful answer.
A knowledge system should have a sensible response when its sources disagree. For example, it can identify the conflict and ask the policy owner to confirm the current rule. Silently choosing the most convenient passage produces confidence without a reason to trust it.
Treat knowledge as maintained operational material. Someone needs to own the source, decide what belongs in a collection and review changes. The AI layer cannot create that ownership on its own.
For a first rollout, choose a small set of frequently asked questions and maintained documents. Include an outdated document, a restricted document and a question with no supported answer in your evaluation. These cases reveal more than a demonstration where every question has an obvious match.
Take this with youMeasure whether the answer is supported and appropriate, alongside whether it sounds helpful.
How do we improve the quality of AI answers?
Start with maintained documents and a named source owner. Organise sources around real questions, keep versions clear and route missing or conflicting information to the person responsible.