Knowledge bases

Create a knowledge base from files, URLs, a crawled website or text, test what it answers, measure it with question sets, and keep it current.

A knowledge base is the material an agent can search during a conversation: your policies, product pages, price lists, help articles. You build and test it under Knowledge Bases, then give it to an agent with the Knowledge Base tool.

Keep the two jobs apart. Facts about your company belong here. What the agent learns about one person belongs in Memory.

What goes into one

A knowledge base is made of sources: files, URLs, or text.

Files. Documents, JSON, and spreadsheets in Excel or CSV.
URLs. Add pages one by one, or switch on Crawl website, give a Website URL, and choose which of the discovered pages to ingest. A Google Doc can be added by its Google Doc URL. A page has to be publicly reachable to be fetched.
Text. Paste content directly. Markdown is supported.

Sources lists what the knowledge base is currently made of.

Creating one

Under Knowledge Bases, choose Add Resource.
Choose your source: Text, URL, File or Google Doc. Clicking a card moves straight on to the next step.
Give the knowledge base a Name and configure the source. For URL, pick a Mode: URLs takes pages one by one (Add URL queues another), with an optional Auto-update that refreshes them every 6, 12 or 24 hours; Crawler takes a Website URL and discovers the pages reachable from it. A Google Doc must be shared so that anyone with the link can view it.
Choose Create. With Crawler, choose Discover instead, pick the pages to include, then Crawl them.
The wizard shows the content being processed, then opens Test your knowledge base so you can ask a first question.

Adding more data

Add data extends a knowledge base that already exists. Each addition either appends or replaces:

Add to existing data keeps what is stored and adds the new content on top.
Replace removes all existing content and its search index, then stores the new content in its place. The app asks for confirmation, because it cannot be undone.

URL sources can also re-fetch their content on a schedule, so a knowledge base built from your website follows it as it changes.

Checking that ingestion worked

Adding a source starts an ingestion job. Live ingestion shows the jobs in progress, and Ingestion history keeps every one, including crawled pages that failed, counted as pages ok and pages failed. When a job fails, an Ingestion failed banner says so and Retry ingestion runs it again.

Testing what it answers

Test has two ways in.

Test via chat. Ask questions as a caller would. Answers stream in with the sources they came from, so you can see where each claim was found, and suggested questions show what the knowledge base can answer.
Semantic search. Search for passages directly. Under Retrieval options, Fast search skips the answer model and returns the matching text only, with no generated answer and no sources: the quickest way to check that a passage is in there at all.

Test here before connecting the knowledge base to an agent. An agent can only answer as well as its knowledge base can.

Measuring it

For more than spot checks, Evaluation runs a fixed set of questions against the knowledge base and keeps each run.

Create a dataset of questions. Add them one at a time, import a CSV with question and expected_answer columns (5 MB at most), or generate them from the knowledge base.
Create an evaluation and start a run.
Check Run history after you change the sources.

Rerun the same set after every significant change: a run that gets worse tells you which change to undo.

Clearing and deleting

Clear empties a knowledge base of its content and keeps the knowledge base itself, ready for new data. Delete removes it entirely. Knowledge bases can also be deleted in bulk from the list.