Knowledge Base

Your knowledge base is what your agent searches when a customer asks a question. Set it up before you deploy an agent, and update it whenever your policies, products, or answers change.

Content lives in sources (named groups of related knowledge) and entries (files, websites, or text). Everything becomes editable text the agent retrieves with semantic search. Sources belong to a workspace and can be shared across agents in that workspace. See Workspaces.

How it works

1. Add content to a knowledge source as a file, a website crawl, or text.
2. Link AI extracts the text, indexes it with vector embeddings, and makes it searchable.
3. Open any entry to edit it. Saving re-embeds the updated text so search stays current.
4. Train the source so the agent knows which topics it covers.
5. Assign the source to an agent. When a customer asks something, the agent searches that source and answers from the retrieved text.

You don't trigger searches yourself. The agent searches when the conversation needs it.

Sources and entries

Knowledge source: A named group of related knowledge, such as "Product docs" or "Support policies". Switch sources from the dropdown at the top of Knowledge Base. Create, rename, or delete a source from New or the source menu.

Entry: One piece of knowledge inside a source. Every entry is editable text, even if it started as a PDF or a crawled site.

Create a knowledge source

If the workspace has no sources yet, Knowledge Base shows an empty state.

Create separate sources for different topics so you can assign the right source to each agent.

Add knowledge

At the top of a source you see Drag and drop anything, or pick a source. Everything becomes editable text the agent can use.

Files

Supported types: PDF, TXT, DOCX, DOC, CSV, PPTX, XLSX, and XLS. Maximum size is 10 MB per file.

Website

Text

Use the search bar to filter entries by title, file name, URL, or content.

Edit entries

Click an entry to open the editor. On a wide screen it opens beside the list. On smaller screens it opens as a sheet.

• Edit the extracted or written text in place.
• Text entries have an editable title.
• Website entries keep a link back to the original URL.
• Unsaved edits show an Unsaved badge. Save from the editor or the floating bar.
• Delete an entry from the editor or the row menu.

Saving re-embeds the entry. Search uses the last saved version until you save.

Ask Ally to Optimize

Open any entry and click Ask Ally to Optimize. Ally rewrites the text so retrieval works better. It:

• Keeps every fact, number, name, URL, and date
• Adds headings, bullets, and Q&A-style sections
• Replaces vague pronouns with the real noun so each paragraph stands on its own
• Stays in the original language
• Doesn't add marketing copy

The rewrite is a draft. Review it, then save.

Train the source

Training builds a Knowledge Profile: a summary of topics, example questions, and a document count. The profile tells the agent what it has before it searches.

Status lives in the header badge next to the source name:

Click the badge to train or retrain. You can also train from the Knowledge field on the agent's Brain tab.

Train after you add, edit, or delete a meaningful amount of content. Ally reminds you to retrain after it changes knowledge.

Assign a source to an agent

Each agent links to one knowledge source. The same source can be linked from more than one agent.

Starter questions in chat are set on the agent, not generated from knowledge entries. Add up to four on Configure Agent.

For live information that isn't in your knowledge base, turn on Web Search.

Manage knowledge with Ally

Ally can search, create, and update knowledge for you:

• "Search my knowledge base for the return policy."
• "Add this shipping policy to Product docs."
• "List my knowledge sources."

When Ally adds text, it uses a short title so the entry is easy to find. After changes, retrain so agents pick them up.

You need to be a workspace owner or admin. Mention a source with @ to scope Ally to that one source for the turn.

Best practices

• One topic per source. Keep product docs, support policies, and shipping info as separate sources.
• Write for retrieval. Use clear headings, concrete nouns, and short sections. Use Optimize when text is messy.
• Edit the extracted text. PDFs and crawls often need a cleanup pass before the agent can quote them well.
• Retrain after big edits. The profile goes stale when content changes, and the badge turns to Needs Training.
• Keep products in Orders. Catalog, prices, and stock belong in Orders, not in knowledge files.
• Test real questions. Ask the assigned agent the questions customers actually ask.

Troubleshooting

The agent says it doesn't know

1. Confirm the source is assigned under Knowledge on the agent's Brain tab.
2. Open the entry and confirm text was extracted.
3. Retrain if the badge says Needs Training.
4. Try a more descriptive question. Semantic search works best with full sentences.

Answers are outdated

Update the entry, save so it re-embeds, then retrain.

Optimize failed or is disabled

Split long entries. Check the remaining messages on your plan.

The crawl found the wrong pages

Deselect extras, add the missing URL manually, then start the crawl.

A file was rejected

Use PDF, TXT, Word, CSV, PowerPoint, or Excel, under 10 MB.

Changes show as unsaved

The amber Unsaved badge means the last save hasn't run yet. Search still uses the previous version.

Next steps