> ## Documentation Index > Fetch the complete documentation index at: https://docs.cxp.crescendo.ai/llms.txt > Use this file to discover all available pages before exploring further. # Knowledge in Applied Insights > Explore Knowledge Base evidence and build widgets that summarize conversations alongside selected policies. # Knowledge in Applied Insights **Coming soon.** This guide describes the first KB-enabled Applied Insights release. Production availability has not yet been confirmed. Use Analyst to inspect knowledge while investigating conversations. Use Widget Builder to save a repeatable comparison with a selected policy, a fixed knowledge search, or a bounded document sample. Each widget refresh fetches its configured sources before optional AI summarization. ## Before you start You need access to Applied Insights and an existing Knowledge Base in your organization. Building widgets requires your existing widget-building access. Your administrator must enable the Knowledge Base data source in Applied Insights settings for widgets that use it. Knowledge content and search also require an available Knowledge service. Existing Insights history and widget sharing rules continue to apply. A widget containing knowledge evidence has the same audience as your other widgets. Knowledge management permissions remain unchanged. ## Explore a policy In Analyst, try: ```text theme={null} Find our returns Knowledge Base and list documents about return eligibility. Read the Current version of the returns policy. Give me the document ID, observation time, and any excerpt limitations. ``` The assistant discovers the available IDs before reading a document. Metadata reads do not automatically fetch content. A content request uses **Current** by default; ask explicitly for **Draft** if that is what you need. Missing Current content is reported as missing rather than replaced with Draft. You can also ask for ranked evidence: ```text theme={null} Search our support Knowledge Base for the conditions for returning a damaged item. Quote a short relevant excerpt and identify the source document when known. ``` Search results show evidence for that query. They do not prove that the whole Knowledge Base has no answer when nothing relevant is returned. The same exploration is available through existing Analyst sessions in the [CXP Plugin](../developers/connect-chatgpt-codex); this feature does not add a separate set of plugin KB tools. ## Build a returns-policy widget In Widget Builder, try: ```text theme={null} Build a widget comparing a small sample of recent returns conversations with our Current returns policy. First find the correct policy document. Save that selected document as a Knowledge Base data source so every refresh reads it again. Summarize recurring questions and possible discrepancies. Show the policy observation time, whether the policy excerpt was truncated, and the conversation sample limits. Do not present this as a complete compliance audit. ``` Review the policy selection and preview before using the widget. Expect a summary based on the fetched conversations and selected policy, with explicit limits. Builder can combine multiple configured knowledge and conversation sources. The execution model summarizes those inputs; it does not make follow-up searches. When the policy changes, refresh the widget to fetch its currently available content. Opening a stored historical result keeps the evidence from that run. Choosing a historical conversation date range does not reconstruct the policy as it existed at that time. KB-backed widget cards show a **Knowledge sources** note automatically in Builder and dashboards. If you have Knowledge access, click a selected article's title to open its Knowledge page; otherwise the saved title appears as plain text. The green checked-document icon identifies the published version; the pencil-document icon identifies an unpublished draft. Hover over an icon for its label. These icons describe the content read for that run; **Read at** is the fetch time, not the article’s last edit time. The link opens the live article page, not a historical snapshot. The note also reports whether the result is partial or shortened, even when the widget's main content only shows a summary. **More articles remain to check** means the listing stopped before inspecting every article; later pages may or may not contain matches for the same filters. Ranked searches always represent partial KB coverage. The note belongs to the displayed execution, so historical results keep their original source limitations. Older saved results without source metadata do not have this note; refresh the widget to obtain a new observation. ## Limits and failures * Document listings return up to 50 matches after inspecting at most 100 descriptors per read. Analyst and Builder can continue through pages while exploring; a saved listing source fetches one page per run. An empty partial page does not establish that the Knowledge Base has no matching documents. * Search returns up to 10 ranked results. Exact content reads return an excerpt of up to 16,000 characters; output-size limits can shorten it further. Search shares the available excerpt space across matches and labels shortened excerpts while retaining their ranking and available source IDs. * Document update dates describe metadata updates, which may differ from the last substantive change to the article. Unknown dates stay unknown. * A missing or failed source does not become zero matches. If combined inputs are too large for the summarizer, ask Builder to narrow the sources. * This first release supports selected evidence and labelled samples. It does not calculate exhaustive stale-article counts, prove that an article was never used, or automatically search knowledge based on findings generated mid-run. For editing and publishing source content, see [Knowledge Base management](../admin/knowledge-base-management).