Data pool search settings
Permissions from the source, passages per answer, release level and sleep mode — what you can set, what tunes itself, and how to check what gets found with "Try it"
What you can set
| Setting | Applies to | Where | Default | What it does |
|---|---|---|---|---|
| Permissions from the source | SharePoint and Google Drive sources in the cloud | Settings → Data sources, for the whole workspace | off | Everyone finds only the files they may open in SharePoint or Google Drive themselves |
| Passages per answer | Outpost | data source → Overview → Settings | 10 (1–100) | Upper limit of passages that go into an answer |
| Release level | Outpost | in the Outpost app; the page only shows it | Text only | What leaves the folder: text, or the original file as well |
| Glossary | Cloud | Settings → Data sources → Glossary button, for the whole workspace | empty | Words that mean the same; search looks for all of them, see Glossary |
| Authoritative / Outdated | Cloud | data source → Knowledge map or Files, per file | none | Slightly prefers a file or moves it back and labels it outdated, see File marks |
A cloud data source has no release level: whoever may use it gets the full text, the original file and images — except sources restricted earlier to Matching passages only or Text only, see Citations in answers. More: Release level.
For Cloud data sources you decide what is connected and who may use it. How it is searched MeinGPT tunes itself: keyword and meaning search always both run, images are read wherever the format allows, and rarely used sources go into sleep mode and come back out on their own. Answers always show their sources; this cannot be turned off.
Permissions from the source
Under Settings → Data sources you decide for the whole workspace whether search in SharePoint and Google Drive sources applies the permissions from the source:
- On: Everyone finds in a SharePoint or Google Drive source only the files they may open there themselves — admins too. Anyone who never signed in to MeinGPT with Microsoft finds nothing in SharePoint sources; anyone who has not signed in with Google finds nothing in Google Drive sources. If the permissions are not current, the source returns no results rather than showing too much. You can see their state on the data source under Settings → Permissions.
- Off: Everyone with access to a data source finds every file in it.
The setting is off by default. Because it changes who finds what, the page asks for confirmation before switching; every change is recorded in the audit log. The setting only appears in workspaces for which MeinGPT has enabled it. If you don't see it, everyone with access to a source finds every file in it; contact MeinGPT support if you'd like to use it. The exception: sources that already apply permissions from the source (including ones created after the setting was switched on) stay locked and return no results while the feature is not enabled for your workspace.
Note
Applies to SharePoint and Google Drive for now. For Confluence and other sources, permissions are not carried over yet: there, everyone with access to the data source finds everything in it.
Google Drive specifics. Shares with people and with the whole domain (when the domain may find the file without a link) are carried over, and so are shared drive members. Shares with a Google group and sharing links ("Anyone with the link", "Anyone in your organization with the link") give nobody access in MeinGPT for now; the data source shows how many such shares it holds. Someone who needs such a file can be shared on it personally. MeinGPT recognises a person's Google account from their Google sign-in to MeinGPT; that sign-in has to be repeated at least every 90 days.
Sleep mode
A Cloud data source that nobody has searched for 60 days goes into sleep mode. You can tell by the "Sleep mode" badge in the list of data sources. There is nothing you need to or can set for it.
Note
Being rolled out gradually. Sleep mode is not active in every environment yet. Until it is active in yours, no source newly falls asleep. A source that is already asleep stays in sleep mode until a search wakes it.
- Keyword search keeps working. Assistants and chats find everything that contains the same words; only meaning search pauses, and new files are not embedded.
- The next search rebuilds the source automatically — in a chat, in search (Cmd+K) or with Try it. It returns keyword matches right away and says that the source is being rebuilt; meaning search is back shortly after. While it is being rebuilt, the source shows "Being rebuilt …".
Nothing is deleted in sleep mode.
Passages per answer
This value sets how many passages from an Outpost data source may go into an answer at most. The range is 1 to 100, the default is 10. For Cloud data sources the assistant chooses the number per search; a fixed value would not be read there and is therefore not shown.
Start at 10. Raise it in small steps when answers stop short — for example in long operating manuals with tables — and check every change with Try it. More passages make the answer slower and take up space in the model's context window; at a high enough value this can lead to a "chat too large" error. More: Chat too large for the model.
The real lever: Try it
If you want to judge the quality of a data source, Try it tells you more than any setting. You find it in the data source on the Overview tab.
There you ask real questions of the source — along the same path assistants use. Every result shows the file and passage; under How was it found? you see whether it came from keyword search, meaning search or both, which rank it had, and which passage and page it comes from.
Search as a person. Instead of searching as yourself, you can search as a colleague. This checks whether that person may use the source and shows what they find in it. If that person has no access to the source, you get no results but the reason — for example that they are in none of the teams it is shared with. Every search as another person is recorded in the audit log, without the search text.
Note
Per-document permissions apply in two cases only: in SharePoint and Google Drive sources with Permissions from the source switched on, and in Outpost folders with Windows permissions. There, "Search as a person" shows the state of that person's permissions, not their results. In every other source, whoever may use the data source finds everything in it; "Search as a person" then checks access to the source.
An approach that works: collect five to ten questions your colleagues would actually ask and go through them in Try it. If a question finds no fitting document, the cause is almost always the scope of the source or a document that is not in the index — not a setting. Whether a particular file is searchable, and why not, is shown on the Files tab. How to improve the scope: Narrow the data scope.
What cannot be set
Some search values are fixed and deliberately not shown in the interface:
- How many passages go into an answer — for Cloud data sources. The assistant chooses this per search.
- How thoroughly a source is searched. When needed, the assistant runs several searches in a row and reads whole files itself; there is no separate switch for it.
- Whether sources are shown. Every answer drawn from a data source names its sources.
- Whether meaning search runs. It is on for every Cloud data source; only the assistant can narrow a single search to one method.
- Whether images are read. Always, wherever the format allows.
- When a source sleeps and wakes. See Sleep mode.
- The mix between meaning search and keyword search.
- The minimum relevance below which a result is dropped.
- A second relevance check (reranking). The platform does not re-sort results with a separate model.
- The choice of search method. The platform decides this automatically per request.
These values are tuned across collections; making them individually adjustable causes more misconfiguration than benefit.
Related pages
- Data sources & RAG: connecting data sources, indexing, sync and access control
- Data quality: how to prepare documents for search
- Knowledge search, glossary and file marks: where search comes back empty and how to help it
- Outpost: data sources in your own network