--- title: "Why Is My AI Knowledge Base Not Finding Documents?" url: https://www.insulin.dev/blog/ai-knowledge-base-not-finding-documents/ canonical: https://www.insulin.dev/blog/ai-knowledge-base-not-finding-documents/ type: Blog description: "When an agent can't find a document you added, the reason is usually on screen: a Test search, an indexing status, a stale sync, or the bases it searches." --- # Why Is My AI Knowledge Base Not Finding Documents? > Canonical HTML version: https://www.insulin.dev/blog/ai-knowledge-base-not-finding-documents/ 1. [Home](/) 2. / 3. [Blog](/blog/) 4. / 5. Why Is My AI Knowledge Base Not Finding Documents? # Why Is My AI Knowledge Base Not Finding Documents? When an agent can't find a document you added, the reason is usually on screen: a Test search, an indexing status, a stale sync, or the bases it searches. ![Zee Chen](/authors/zee-chen.jpg) Zee Chen Sep 24, 2026 ![Why Is My AI Knowledge Base Not Finding Documents?](/images/blog/ai-knowledge-base-not-finding-documents/hero.png) Explore AI Summary [![](/logos/company/openai.svg)](https://chat.openai.com/?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fai-knowledge-base-not-finding-documents%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Knowledge%20Bases%2C%20Search. "Summarize with ChatGPT")[![](/logos/company/anthropic.svg) ](https://claude.ai/new?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fai-knowledge-base-not-finding-documents%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Knowledge%20Bases%2C%20Search. "Summarize with Claude")[![](/logos/company/gemini.svg)](https://www.google.com/search?udm=50&aep=11&q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fai-knowledge-base-not-finding-documents%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Knowledge%20Bases%2C%20Search. "Summarize with Gemini")[](https://www.perplexity.ai/search/new?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fai-knowledge-base-not-finding-documents%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Knowledge%20Bases%2C%20Search. "Summarize with Perplexity") Table of Contents - [Where should you look first when an agent can’t find a document?](#where-should-you-look-first-when-an-agent-cant-find-a-document) - [Symptom, check, fix: the diagnostic table](#symptom-check-fix-the-diagnostic-table) - [Which statuses mean a document isn’t searchable yet?](#which-statuses-mean-a-document-isnt-searchable-yet) - [Why are documents missing after an embedding-model change?](#why-are-documents-missing-after-an-embedding-model-change) - [Why does the agent quote an old version?](#why-does-the-agent-quote-an-old-version) - [Is the agent searching that knowledge base at all?](#is-the-agent-searching-that-knowledge-base-at-all) - [What if it’s still missing?](#what-if-its-still-missing) - [Frequently asked questions](#frequently-asked-questions) - [Takeaways](#takeaways) _When an AI knowledge base isn’t finding a document you know is there, the cause has usually left a visible trace. The skill is checking for it in the right order._ * * * An agent says it couldn’t find anything on the subject, and the document is sitting in the knowledge base. The instinct is to blame the model or rewrite the agent’s instructions. Hold off: in Insulin, **a missing document has a visible state** — a Test search result, a status, a last-sync time, the list of knowledge bases an agent searches. Read them in order and one of them usually names the cause. ## Where should you look first when an agent can’t find a document? **Run Test search in the knowledge base that holds the document, using the words the user typed.** **Test search is** the knowledge base’s own preview of what an agent would retrieve: enter a query and it returns the matching passages, each with its source document and a relevance score, before any answer is written. The result splits the problem in two. If the passage comes back, the knowledge base is doing its job, and the question is whether the agent searched it. If it doesn’t, the question is whether the document is searchable at all, so read its status next. Use the user’s phrasing rather than the document’s title: a query that repeats the document’s own wording proves very little. ## Symptom, check, fix: the diagnostic table **Work down the steps in order and stop at the first one that matches.** Steps 1–7 cover most missing documents; steps 8–11 are for when everything above them checks out. # What you see What to check What to do 1 The agent says it found nothing **Test search**, in the user’s own words Passage returned: go to step 6. Nothing returned: go to step 2 2 **Pending** or **Indexing** The document’s status; a progress bar and count on a website row or connector card Wait. Only **Indexed** documents are searchable 3 **Failed** Hover the alert icon for the error Fix the cause, then re-index the document 4 **Deprecated** The document’s status **Restore** it; it re-indexes and returns to search 5 Documents missing since an embedding-model change **Settings → Embedding model**: `N of M documents searchable` Wait for the rest. **Index them now** retries failures; **Restart indexing** resumes a stalled rebuild 6 Test search returns the old wording How the document got in, and when it last synced File: re-upload it under the same name. Website or connector: **Re-sync** 7 Test search finds it; the agent doesn’t The agent’s **Knowledge Bases** section; for Insulin, the **Automatic** badge Attach the knowledge base to the agent, or select it ### If it’s still missing # What you see What to check What to do 8 The document isn’t in the list at all The crawl’s **Depth** and 100-page cap; the connector’s scope; the file’s format and size Add the page’s own URL; widen the scope with **Edit scope**; split or convert the file 9 It ranks low in Test search Search mode, vector weight, **Top K** Tune the knowledge base 10 An agent with many knowledge bases misses it How many are attached Attach fewer: an agent searches at most three per turn 11 An Inbox draft ignored it Inbox **Settings → Account → Knowledge bases** Attach fewer, so it is among the first three; auto-sent replies use none ## Which statuses mean a document isn’t searchable yet? **Only an Indexed document can be found; Pending, Indexing, Failed and Deprecated all keep it out of search.** **Pending** means queued and **Indexing** means being chunked and embedded; a website or connector that is indexing shows a progress bar and a count of documents indexed so far. **Failed** carries its reason: hover the alert icon, fix the cause, and re-index the document. When many documents in a personal knowledge base fail together, suspect the embedding provider: a personal knowledge base doesn’t switch models on its own, so choose a different one in **Settings**. **Deprecated** is a decision, not an error: someone retired the file from search without deleting it. **Restore** returns it as **Pending**, searchable once it re-indexes. And if documents in an older knowledge base sit at **Pending** indefinitely, open Settings: one created before models were recorded may have **no model set**, and nothing indexes until you choose one. ## Why are documents missing after an embedding-model change? **Because search hides each document until it has been re-embedded on the new model.** Search keeps working meanwhile, answering only from documents already rebuilt, and an agent searching mid-rebuild is told how much of the knowledge base it covered. Settings counts progress under **Embedding model**, as in `3 of 42 documents searchable`, and reports failures separately, because waiting won’t clear them: **Index them now** retries every outstanding document. If progress stalls, **Restart indexing** picks up exactly what is outstanding without redoing finished work. ## Why does the agent quote an old version? **Because only connectors with Auto-sync refresh on their own; uploaded files and crawled websites never do.** The fix depends on how the document got in: - **Uploaded file.** Re-upload the new version under the same name, and it replaces the existing document in place. A revision uploaded under a different name is a second document, so both versions stay searchable until you [deprecate or delete the old one](/blog/deprecate-replace-or-delete-a-knowledge-base-document/). - **Website.** A site re-crawls only when you trigger **Re-sync**. Its row’s **Last Sync** column says when that last happened. - **Connector.** **Auto-sync** is chosen when the connector is connected. Insulin sweeps due Auto-sync connectors about once an hour: incrementally once the last sync is over an hour old, with a full reconcile once the last full sync is over 24 hours old. Without Auto-sync, use **Re-sync** on the connector’s card. After any refresh, wait for **Indexed** before running Test search again. ## Is the agent searching that knowledge base at all? **A custom agent searches only the knowledge bases attached to it; the Insulin assistant, on Automatic, searches all of your own.** - **A custom agent.** Open its create or edit form and find the **Knowledge Bases** section: a knowledge base that isn’t ticked is never searched. An organization knowledge base attaches to an org-level agent. - **The Insulin assistant.** In **Chat**, select Insulin and click the pencil in the conversation header. An **Automatic** badge means Insulin searches every knowledge base you own. Tick or untick any row and your selection takes over completely: what isn’t ticked is not searched, and nothing ticked means no knowledge base at all. **Reset to automatic** goes back. Automatic doesn’t include knowledge bases shared with you. To include one, tick it in the same list, where it carries a **Shared** badge; that makes the selection explicit, so tick every knowledge base Insulin should search. [How Insulin controls which agents see which documents](/blog/which-agents-see-which-documents/) covers attachment, Read and Edit modes, and sharing. ## What if it’s still missing? **Once status, freshness and scope check out, look at what never made it in, how the passage ranks, and how many knowledge bases are in play.** ### The document never made it in A crawl follows links to its **Depth** (1–10, default 3; depth 1 is the start page alone) and imports up to 100 pages, so a deeper or later page never arrives: add its own URL as a website. Sites added with **Bulk add** aren’t crawled until you start a sync. A connector brings in only its scope — the repositories, spaces or folders picked, up to any maximum set — so widen it with **Edit scope**; GitHub syncs issues and pull requests, never source code. For files, check [which file types and sizes a knowledge base accepts](/blog/what-documents-an-ai-knowledge-base-can-read/). ### It ranks too low A passage far down Test search’s list may never reach an answer. Search mode and vector weight decide which passage wins; [weighting vector against keyword search](/blog/tune-hybrid-search-vector-keyword-weight/) covers that dial. Mind the cut-offs: Test search previews up to 20 results, a knowledge base’s default **Top K** runs from 1 to 50 in Settings, and an agent searching at query time can request up to 100. ### Too many knowledge bases are attached An agent searches at most three knowledge bases per turn, so with more attached, the one holding the document may not be among those searched. Attach fewer, narrower knowledge bases. ### An Inbox draft didn’t use it Inbox keeps its own read-only selection, under **Settings → Account → Knowledge bases**. A draft searches the first three attached knowledge bases you still have access to, in the order you saved them, taking up to four passages from each, and only drafts bound for **Approvals ▸ Pending** consult them: a reply sent automatically consults none. The lookup is best-effort, so a knowledge base you’ve lost access to is skipped silently, and a slow or failed search leaves the draft without it. ## Frequently asked questions ### Why can’t my AI agent find a document that’s in the knowledge base? Usually the document isn’t searchable yet, the search ranks it too low, or the agent isn’t searching that knowledge base. Run Test search with the user’s own words first: if the passage comes back, check the agent’s knowledge bases; if not, check the document’s status. ### When does an uploaded document become searchable? When its status reads Indexed. Pending means it is queued and Indexing means it is being chunked and embedded; neither is searchable yet. A Failed document shows its error when you hover the alert icon, and needs the cause fixed and a re-index. ### Do knowledge bases update automatically when a source changes? Only connectors with Auto-sync do, swept about once an hour. Uploaded files and crawled websites never refresh on their own: re-upload the file under the same name, or re-sync the website. ### Which knowledge bases does the Insulin assistant search? On Automatic, every knowledge base you own. Tick or untick any of them and your selection takes over completely, so an unticked knowledge base is not searched, and nothing ticked means none. Knowledge bases shared with you are not part of Automatic. ### Does Test search show everything an agent could retrieve? No. Test search previews up to 20 results. A knowledge base’s default Top K, set in Settings, can be as high as 50, and an agent searching at query time can request up to 100 results. ## Takeaways - Start with **Test search**, in the user’s own words: it tells you whether the knowledge base or the agent is at fault. - Only **Indexed** documents are searchable. **Failed** explains itself on hover; **Deprecated** needs **Restore**. - Files and websites never refresh on their own. Only connectors with **Auto-sync** are swept, about hourly. - A custom agent searches only what is attached; Insulin on **Automatic** searches your own knowledge bases, and an explicit selection replaces that completely. - Check crawl and upload limits, retrieval tuning and Inbox settings last. For what [Insulin knowledge bases](/knowledge-bases/) are built to do, start with the product page; the [Knowledge Bases documentation](https://doc.fours.com/insulin/knowledge-base/) is the full reference for statuses, syncing and limits. ## Sources Primary sources for the platform rules cited above. Last verified September 24, 2026. Cloud providers change fees, eligibility, and program terms without notice — check the source before relying on a figure. - [Fours docs: Insulin Knowledge Bases](https://doc.fours.com/insulin/knowledge-base/) — Test search and its 20-result preview against the 1–50 default Top K and up to 100 results per agent query; the statuses, and that only indexed documents are searchable; same-name uploads replacing in place; websites re-crawling only on Re-sync; Auto-sync chosen at connect and swept about hourly, with a full reconcile after 24 hours; re-indexing progress, Index them now and Restart indexing; Deprecate and Restore; which knowledge bases custom agents and the Insulin assistant search, and up to three per query; connector scope, crawl depth, the 100-page cap and unsynced bulk imports - [Fours docs: Insulin Inbox](https://doc.fours.com/insulin/inbox/) — Inbox drafts search the first three attached knowledge bases still accessible, in saved order, taking up to four passages from each; the lookup is best-effort and a knowledge base you lose access to is skipped silently; replies sent automatically consult none ## Keep reading - [Knowledge BasesChange a Knowledge Base's Embedding Model SafelySep 30, 2026](/blog/switch-embedding-models-without-breaking-search/) - [Knowledge BasesDeprecate, Replace, or Delete a Knowledge-Base DocumentSep 28, 2026](/blog/deprecate-replace-or-delete-a-knowledge-base-document/) - [Knowledge BasesPermission-Aware Enterprise Search for AI AgentsAug 21, 2026](/blog/permission-aware-enterprise-search/) - [SearchHow to Measure AI Answer Quality and Source CoverageAug 20, 2026](/blog/measure-ai-answer-quality/) [Browse every post on the Insulin Blog](/blog/) ### Stay Updated New posts, product updates and marketplace strategy are shared on LinkedIn as they publish. [Follow Fours on LinkedIn](https://www.linkedin.com/company/suger-inc)