The shift from keyword to meaning
Keyword search asks: “Where do the exact words I typed appear?”
Search by meaning asks: “Where does the concept I described actually live in my sources?”
The difference matters more than it sounds. With keyword search, the word “renewal” finds documents that say “renewal.” With search by meaning, the phrase “what are our renewal terms” surfaces the right paragraphs — even if those paragraphs talk about “contract extensions” or “term continuations” without ever using the word “renewal.”
That shift unlocks a different kind of question.
Stop thinking about keywords. Start thinking about concepts. Ask the question you'd ask a knowledgeable colleague — not the search query you'd punch into a website.
What you can search across
The Data layer treats all your sources as one searchable knowledge base, scoped to what you have access to. That includes:
- Files you’ve uploaded directly
- Folders, spaces, and projects from connected apps
- Past conversations and saved outputs (if your workspace enables this)
- Generated artifacts you’ve stored
You can search across all of them or narrow to a specific scope.
How to ask great questions
A few patterns that consistently produce sharper answers:
1. Ask the question, don’t search for words
Weak (keyword-style): “renewal terms MSA”
Strong (question-style): “What are the renewal terms in our standard MSA, and how do they differ from what we agreed with Acme Corp?”
The strong version gives the AI the audience (you), the comparison (standard vs. Acme), and the scope (renewal terms). The keyword version is hoping you’ll get lucky.
2. Name the scope explicitly when you can
If you know which source or folder the answer lives in, say so:
“Looking only at our ‘Engineering Standards’ Confluence space, what are our principles for code review?”
The scope makes the search faster and the answer more reliable.
3. Combine sources in one question
The biggest unlock is asking questions that require multiple sources to answer well:
“Compare what our standard contract template says about indemnification with what we agreed in the attached Acme MSA.”
The AI reads both, identifies the differences, and returns a structured comparison with citations.
4. Ask for citations
Even if you trust the answer, asking for citations builds a verification habit:
“Summarize our policy on customer data retention — and cite the specific sections you’re drawing from.”
If a citation doesn’t check out, you learn fast.
Narrowing when results are too broad
If an answer feels generic, the issue is usually scope. Three ways to narrow:
By source
“Use only the attached customer-2025-Q3-renewal.pdf.”
By folder, space, or project
“Look only at our ‘Customer Success Playbook’ Confluence space.”
By date range or other metadata
“Only consider notes from the last 30 days.”
The narrower the scope, the sharper the answer.
A useful sequence
When a first answer is too generic:
- Ask for citations on the original answer.
- See which sources got cited.
- Re-ask the question, scoped to just those sources.
This usually gets you a much tighter answer in one or two follow-ups.
Broadening when results are too narrow
The opposite problem: a search returns one good answer but you know there’s more.
Open up the source scope
“In addition to the Q3 playbook, also reference the Q2 retrospective and the customer call notes from August.”
Ask for “all references”
“Find every reference in our connected sources to our migration approach — list each one with the source and a one-line summary.”
Ask for adjacent concepts
“What do we say about migration? Also include anything about transition, cutover, or switchover — the team uses these words interchangeably.”
When the answer feels wrong
Search by meaning is powerful but not infallible. A few signs to watch for and what to do:
| Symptom | Likely cause | Fix |
|---|---|---|
| Answer cites the wrong document | Source content overlaps; the AI grabbed the wrong source | Re-ask with explicit scope |
| Answer cites a real document but the quote isn’t in it | Paraphrasing drift | Ask “show me the exact quote” — verify against the source |
| Answer is confidently wrong | Source content is contradicted across multiple places | Ask “show all sources that mention this” — reconcile manually |
| Answer references a doc you don’t have access to | Connection scope is broader than expected | Check the connected scope; narrow if needed |
Citations don't equal correctness. A cited source is a reference, not a proof. Open the cited source for anything you'll act on — especially numbers, names, dates, and policy language.
Working with very large source sets
When your team’s Data area has hundreds of files and dozens of connections, search behavior changes — there’s more potential for the AI to draw from less-relevant sources.
A few habits that help:
1. Use the most specific scope you can
If you know the answer is in one space, scope to that space. Generic searches across a huge corpus are noisier.
2. Prefer “what does X say about Y” over “tell me about Y”
The specific source pointer disambiguates the search.
3. Maintain your sources
A bloated, stale Data area produces noisy answers. See Connecting sources for cleanup habits.
4. Pin source priorities (if available)
Some workspaces let you mark certain sources as “high authority” for certain topics. The AI prefers those when answering related questions.
Search as the start of a workflow
Search isn’t always the end of a task — often it’s the beginning.
A common pattern:
- Search to find the relevant sources.
- Read the cited material to make sure the framing is right.
- Run an agent or network using those sources to produce a deliverable.
That sequence — search for the right sources, then generate from them — is the most reliable way to produce grounded outputs at speed.
Where to go next
- Use cases — search-driven scenarios with real questions and outputs.
- Connecting sources — keep the sources you search across clean and current.
- VDF AI Networks — when search is the first stage of a larger workflow.