An AI agent needs reliable business information to answer useful questions. The Knowledge library lets you organize documents, website pages and FAQs into collections that a Thinking Engine agent can search.
Start with a collection such as Delivery and collection, containing current delivery areas, collection hours and escalation instructions. Keep its sources accurate and consistent.
Before you start
You need permission to manage Thinking Engine settings, access to Knowledge, and a business AI agent to attach the collection to. The installation must support knowledge ingestion, and the agent's Thinking Engine must be enabled for agent runs.
Prepare content you want the agent to use in customer answers. Include dates or version notes where policies change. Keep private credentials and unrelated customer records out of general business guidance. Indexing can incur embedding usage charges when embeddings are configured, and the agent's Run test is a billed run. Check the workspace's available credit before testing.
Add and check your sources
- Open Knowledge, choose New collection, enter a descriptive name and select Create. Collections let you attach the same set of sources to the agents that need them.
- In that collection, choose Add knowledge → Write document. Enter a title, such as
Collection hours and delivery areas, then paste verified information. Use short sections with clear headings. State what the agent should do when a destination or exception is not covered. Select Save document. - Add other sources when useful. Upload file supports PDF, Excel
.xlsx, CSV, JSON, text, Markdown and HTML files up to 15 MB; select Upload & index. FAQ set accepts question-and-answer pairs and saves with Save FAQ set. For Website URL, enter the specific page and choose Crawl & index. Treat each URL as a single-page source and add other important pages separately; do not assume the whole website was imported. - Wait for the document status to move from pending through processing to ready, and check that it has indexed chunks. Chunks are the searchable passages extracted from the source. A successful upload alone does not establish that the content is searchable.
- Under Test a search, enter a realistic customer question and select Search. Try “Where can I collect my order?” and a question about an unsupported delivery area. Read the returned passages and confirm they contain the correct facts. This library search can include collections beyond those attached to one agent, so also test the agent itself.
Attach the collection and test the agent
- Open Agent Manager, then the intended agent's Thinking Engine page. Under Knowledge → Collections this agent can use, select the collection. Automatic context adds relevant passages before the model begins; Passages (top-K) controls how many it receives. Select Save changes.
- In Test it, ask a factual question and choose Run test. This uses the last saved configuration and can execute the agent's enabled tools. Keep the request focused on looking up information. Review the answer and available steps, including its supporting sources. Test an unanswered question as well: the agent should acknowledge missing information instead of inventing a policy.
Keep the answers dependable
Success means the sources are ready, the intended collection is saved on the agent, and the agent answers accurately from those sources. An empty collection or a saved upload that is still processing is not ready for customer use.
- Ingestion is unavailable or remains pending: ask your workspace administrator to check installation readiness and processing.
- A document fails: read its error. Check readable text, supported file type, size, URL accessibility and the displayed workspace quota before using Reindex.
- Search finds nothing useful: improve the source wording and headings, then check the agent's selected collections.
- A policy changes: reindex a changed URL source, or replace an outdated upload or written document with the corrected version. Remove superseded material so agents do not retrieve conflicting answers.
Build a short question set before launch
Use questions a customer would actually ask: “Are you open on Sunday?”, “Can someone else collect my order?”, and “Do you deliver to my area?” Add an outdated-policy question and one that has no answer in the collection. Record the expected fact or handover for each.
Test search first, then the saved agent configuration. When search finds the right passage but the agent answers incorrectly, review its instructions, selected collections and available tools. When search itself misses the answer, improve the source or its indexing before changing the agent's tone.
Keep procedures and business facts separate enough to maintain. An opening-hours document can name the current hours; agent instructions can explain how to answer briefly and when to ask a colleague. Avoid putting conflicting versions of the same policy in several collections.
Give somebody responsibility for reviewing changes to prices, locations and service terms. Update the source, check processing, and repeat the affected questions. A document's ready status confirms indexing, not that the policy inside it is still correct.