One search box. Everyone sees only what they could already open.
Knowledge is spread across file shares, wikis, ticketing systems, and databases. Data.co indexes them inside your environment, carries each source's permissions forward, and filters every result by the person asking before anything is shown or summarized.
Permissions travel with the content.
Source ACLs are synced with the content. At query time, results are filtered by the user's identity and groups before the model sees them.
Enterprise search projects fail on permissions, not relevance.
The fastest way to end an internal search project is for someone to find the compensation spreadsheet. Systems that index with a privileged service account and filter loosely afterward tend to expose exactly the documents that were protected only by obscurity.
Data.co filters before ranking and before generation. A document that the user could not open in its source system is never retrieved for them, never summarized, and never cited. Classification labels add a second layer: even an overshared file can be excluded if its label says so.
What users get.
Hybrid retrieval
Keyword and vector search combined, so exact identifiers and natural-language questions both work.
Cited answers
Generated answers link to the passages they used, so readers can check them.
Freshness
Incremental sync picks up changes and permission updates on a schedule you set.
Scoped collections
Teams can search a single collection, such as HR policies or engineering runbooks.
Questions worth answering early.
- Which sources have reliable permissions today
- How group membership maps from your identity provider
- Which labels should exclude content regardless of ACLs
- How often permission changes need to propagate
- Who reviews oversharing discovered during indexing
- Whether answer text is retained in the audit log
An engineer searches for the on-call runbook.
A new engineer types "database failover runbook" into the search box on their first week.
- 1
Identity
The search app passes the engineer's SSO identity. Their groups come from your identity provider at query time, so yesterday's onboarding change is already in effect.
- 2
Candidate set
Hybrid search looks across the wiki, the ticket system, and the file share, but only over content whose synced ACLs include one of the engineer's groups.
- 3
Label check
A post-incident review in the same space carries a restricted label for a customer name; policy excludes it for this role, even though the space is open.
- 4
Answer
A local model summarizes the two runbooks that remain and cites the sections it used. Each citation opens the source in its own system.
- 5
Record
The audit log records who searched, which collections were queried, which items were filtered and why, and the model used.
A phased launch that surfaces problems early.
| Phase | Audience | What you learn |
|---|---|---|
| Index and report | Nobody searches yet | Which content is overshared, which sources have stale permissions |
| Pilot | One team, one or two sources | Relevance gaps, missing sources, group mapping issues |
| Department | A full department | Load, freshness targets, how people phrase questions |
| Broad launch | Everyone with access to the sources | Usage patterns that justify adding sources or collections |
Questions buyers ask
What if a source has overly broad permissions?
Data.co respects source permissions but can also restrict by label. Indexing often surfaces overshared content, which we recommend reviewing before broad launch.
How quickly do permission changes take effect?
Permission sync runs on a schedule you configure, and some connectors support event-driven updates. Group changes from your identity provider apply at query time.
Does the search index leave our environment?
No. Content, embeddings, and the index are stored in your infrastructure.
How is this different from enterprise RAG?
Knowledge search is the user-facing experience of finding and reading internal content. Enterprise RAG is the underlying pattern for grounding any model or agent in that content, including inside other applications.
How do we judge whether search is good enough to launch?
Build a short list of real questions with known correct sources from pilot users, and check results against it before each phase. Your team sets the bar; the evaluation set makes it repeatable when content or models change.
Do users need special search syntax?
No. Vector search matches meaning rather than exact words, so plain questions work, and keyword search still finds exact identifiers such as ticket numbers or part codes. Results depend on the embedding model chosen for your content.
What about sources without usable permissions?
Sources with no reliable ACLs can be indexed into a restricted collection with explicit group access, or left out until their owners fix permissions. Indexing them as open is the mistake to avoid.
See it against your own use case.
A 30-minute technical walkthrough, policy simulation on your scenarios, and a deployment recommendation for on-prem, private cloud, or hybrid.