--- title: "Knowledge search, glossary and file marks" description: "See where searches in your data sources come back empty, which files answers rely on and which sources nobody needs — and help search along with a glossary and file marks" canonical_url: "https://meingpt.com/en/docs/integrations/knowledge-insights-glossary" language: en --- # Knowledge search, glossary and file marks Three tools show admins how the data sources are searched, and help search along: - **Knowledge search** in the [AI Cockpit](/en/docs/admin-guide/ai-cockpit): how often people searched, where searches come back empty, which files answers rely on and which sources nobody needs. - **Glossary** on **Settings → Data sources**: which words mean the same in your documents. - **Authoritative** or **Outdated** on single files, right on a data source's **Knowledge map**. ## Knowledge search in the AI Cockpit The **Knowledge search** tab in the AI Cockpit sums up all Cloud data sources of the workspace for the selected period — the last 30 days by default, at most 90 days are evaluated: | Section | What it shows | | ---------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------- | | **Searches**, **No results**, **Not searched** | How often people searched, the share of searches that found not a single passage, and how many sources nobody searched | | **Searches per day** | Searches over time, with searches without results highlighted | | **Where searches come back empty** | Per source, the share of searches without results | | **Not searched in this period** | Sources without a search, with the last search date and a "Sleep mode" badge if the source already sleeps | | **Most cited** | Files that answers name as a source in their text | In the list under **Settings → Data sources**, the **Usage (30 d)** column shows the same numbers per source for the last 30 days, for example "241 searches" with "36 % without results". "never searched" means the source was never searched at all; "no search in 30 d" means it was, just not in the last 30 days. You can sort by it. How to read the numbers: - **A search** is the search of one source. A question that searches three sources counts three times. File-name lookups in search (Cmd+K) are not counted. - **No results** means the source returned not a single passage. Whether the passages it found answered the question is not measured. - **Citations** are counted from this update on; older answers are missing from the list. - Sources with [permissions from the source](/en/docs/integrations/data-pool-search-settings#permissions-from-the-source) switched on appear without file names, because there every person sees different files. Outpost sources are missing: their searches stay on the device. **No search terms.** MeinGPT doesn't store what people searched for — only that they searched and whether anything was found. That's why there is no list of frequent or unanswered search terms. Every view is in the audit log. **What to do with it:** If a source often comes back empty, documents are missing there, or they use other words than the questions — that's what the glossary is for. Sources nobody searches go to [sleep mode](/en/docs/integrations/data-pool-search-settings#sleep-mode) on their own after 60 days; if you don't need them at all, you can remove them. ## Glossary Open the glossary under **Settings → Data sources** with the **Glossary · N terms** button at the top right; it opens as a side panel. There you enter words that mean the same in your documents — for example **KST** and **Kostenstelle**, or **company car** and **pool car**. When someone searches for one of them, MeinGPT also searches for the others, and the assistant learns that they mean the same. This applies to every Cloud data source of the workspace from the next search on. - **Add:** the word or abbreviation on the left, the words with the same meaning on the right, several separated by commas. - **Change:** edit an entry's words right in the list; they are saved when you leave the field. - **Remove:** with the bin icon, after a confirmation. - The glossary holds up to 500 entries with up to eight words each. **Suggestions from the assistant.** When the assistant notices in a chat that your documents use another word — for example because a search for "KST" found nothing and one for "Kostenstelle" found the answer — it suggests an entry. A hint on the button counts waiting suggestions (for example "3 suggestions"); in the side panel they are at the top; they change search only once you accept them. A suggestion doesn't record who asked or in which chat. **Applies to Cloud data sources.** Outpost folders are searched by the device itself; the glossary has no effect there. ## Marking files as authoritative or outdated A Cloud data source's **Knowledge map** shows two lists below the map: - **Most cited:** the files that answers from the last 30 days cite. Next to each you choose **Authoritative**, **Outdated** or **No mark**. - **Large and long unchanged:** the largest folders in which no file changed for over a year — candidates for clean-up or for marking as outdated. You find any other file in the **Files** tab through its file search and mark it the same way; marked files are listed below in a list of their own. Sources with permissions from the source switched on show no file names on the knowledge map. - **Authoritative:** search slightly prefers the file, and the assistant follows it when sources contradict each other. - **Outdated:** search moves the file back a little. The assistant doesn't answer from it when another source answers — and says so when it is the only one. Its citation reads "Outdated". - A mark moves a file only a few places; an outdated file stays findable. When someone asks for a specific edition ("policy 2022"), it isn't moved back. - The mark belongs to the path: if the file is renamed or moved, the mark is gone. Each source holds up to 1,000 marks. Older editions of the same file in the same folder ("Policy_2022", "Policy_2024") are recognized without a mark. Marking pays off for everything the name doesn't tell. Every change to the glossary and to marks is in the audit log. ## Related pages - [Data pool search settings](/en/docs/integrations/data-pool-search-settings): permissions from the source, sleep mode, Try it - [Data sources & RAG](/en/docs/integrations/data-pools-rag): connecting and scoping data sources - [Data quality](/en/docs/integrations/guide-data-quality): preparing documents for search