Knowledge Hub: connect your sources without losing their permissions

An agent is only as good as what it can read. The Knowledge Hub is the part of Scalt that connects your sources, organises their access rights and follows their updates. It is what Scalter and your agents read. Here is how it works, and why we built it around permissions rather than volume.

The cost of silos

In most companies, useful information is spread across a drive, a wiki, a CRM, a support tool and a few databases. Each tool has its own search, its own permissions and its own habits. To answer a simple question about a customer, you often have to open three of them, and know which one is authoritative.

The usual answer is to export the data into a new shared space. It creates one more silo, quickly out of date, where the original permissions have disappeared. That is exactly what we wanted to avoid. Every copy also becomes one more security question: who has access to it, and until when?

Connect rather than migrate

The Knowledge Hub plugs into your sources where they are: documents, databases and CRMs, APIs, cloud storage, live feeds and the web, for instance Notion, Drive, SharePoint, Confluence, HubSpot, Postgres, Snowflake or S3. You pick a connector, authenticate, and preview what will be read. Your teams keep working in their tools; nothing needs migrating.

To be precise: for search to be fast, Scalt keeps an index, that is, a working copy of the extracted text and of its representation for search. This index is synchronised with the source, versioned, and purged when the source removes a document. The source remains the reference; the index is only its reflection.

Checked before being indexed

Connecting a source does not mean indexing everything blindly. Every document is checked before it enters the index: provenance and format, extraction quality, personal data, hidden instructions, duplicates and contradictions. At the slightest doubt, it stays in quarantine and is never served to an agent without a review. We detail these checks in our article on enterprise AI and data.

Each source’s permissions, applied when reading

An agent must never read more than its user. The Knowledge Hub discovers the entities of each source, and permissions are granted by team or by person. When an agent or Scalter searches, those permissions are applied: a salesperson does not see another region’s contracts, and neither does the agent working for them.

Applying permissions when reading, rather than filtering an answer already written, makes all the difference for confidentiality: a document you have no access to never enters the agent’s context, and so cannot show through in its answer.

Granting permissions by team rather than person by person makes them easier to manage: someone who belongs to a team, and the agents working for them, read what that team can read. You set access once, in the Knowledge Hub, instead of maintaining a list of exceptions by hand in every tool.

Following updates and removals

Sources synchronise on a schedule, and every synchronisation is a version. You compare two versions, and roll back if a synchronisation went wrong, without breaking the agents that read the source. A document removed from a source is purged from the index within 24 hours, and retention periods are set source by source.

Finding the right information

Search combines meaning and words: vector search finds the passages close to the question, keyword search finds exact terms, such as a contract reference, and both are merged then reranked by relevance, in French as in English. On top, a graph links companies, contacts, documents and cases, so that agents follow relationships instead of guessing among excerpts.

Monitoring source health

A source that stops synchronising means an agent answering on old data. The Knowledge Hub shows the state of each source, the most frequent queries and the quality of retrieval. When you change the indexing method, you rerun the computation in one click.

A shared foundation

Scalter, search and all your agents read the same Knowledge Hub, with the same permissions. A source is connected once, checked once, and used everywhere. That is what avoids having one agent per tool, each with its own copy of the data and its own rules. When you add an agent, there is nothing to reconnect: it reads the sources already plugged in, with the permissions already set. And when you remove a source, it disappears for everyone at once.

Frequently asked questions