Knowledge Base

Knowledge Base

Video

Watch on YouTube ↗

Screenshot

The Knowledge Base is the set of documents your Aura AI voice agents answer from. When a caller asks a question, the agent searches these documents for the passages that match and writes its reply from what it found — grounded in your own wording and your own numbers, rather than in whatever the model would otherwise produce. This is retrieval-augmented generation (RAG), and it happens in the middle of the live call.

Navigate to Aura AI → Agents → Knowledge Base. Requires the Manage AI Agents permission (aura.agents.manage). The Knowledge Base belongs to your organization; every voice agent on the account draws from the same library, but each agent sees only the tags you grant it.

Caution

Adding a document does not give any agent access to it. An agent with no tags assigned draws on no knowledge base during a normal conversational turn — it answers only from its prompt. Assigning tags is a separate, deliberate step on each agent’s Knowledge tab — see Scoping documents to an agent.

The one exception is a flow with an explicit KB answer node: that node retrieves from the agent’s assigned tags, and when the agent has none it falls back to the whole tenant Knowledge Base. Set the node’s own tag filter, or assign the agent tags, if you need that node scoped.

How Retrieval Works

1. You add a document — a title and a body of text.
2. SIPSTACK chunks it — the body is split into overlapping passages of roughly 1,200 characters, broken on paragraph and sentence boundaries where it can. A short document is a single chunk; a long policy becomes several.
3. Each chunk is embedded — turned into a vector so it can be matched on meaning, not on keywords. The Status column shows Ready once every chunk is indexed.
4. On a call, the agent searches — the caller’s question is embedded the same way and matched against the chunks the agent is allowed to see. The best matches are injected into the agent’s prompt, with the document title as the citation.

Retrieval is scoped to your tenant at the row level: an agent can only ever retrieve chunks belonging to your own organization.

The Document List

Column	What it tells you
Title	The document name, with its source underneath (manual for anything typed or pasted in). The title is what the agent cites when it grounds an answer on the document.
Tags	Which agents may retrieve it. Up to three are shown; the rest collapse to a +n count.
Chunks	How many searchable passages the body was split into.
Status	Ready (indexed and retrievable), Indexing/Pending (in progress), or Error (embedding failed — edit and re-save to retry).
Updated	When the document last changed. Editing the body re-embeds it.

Filter the list by tag (the dropdown shows each tag with its document count) or search across titles, sources, and tags. Clear resets both.

Adding a Document

Click New document:

1. Title — make it descriptive. This is the citation the agent quotes back as its source, so Return and Refund Policy reads better to a caller than Doc 4.
2. Content — paste the wording your team already uses with customers. Markdown is preserved for retrieval.
3. Tags — comma-separated (support, returns). Tags are how you scope documents to agents. An untagged document is unreachable by any agent that has tags assigned, so tag everything you want retrieved.

Create document chunks and embeds the body inline, so the row comes back Ready and the document is retrievable on the very next call. Editing a document’s body re-embeds it automatically; editing only its title or tags does not.

One topic per document

Retrieval quality comes from focused documents. A single page covering warranties, returns and shipping matches worse than three pages that each cover one topic — the matching passage is diluted by everything around it. Split by topic, not by department.

Scoping Documents to an Agent

Open a voice agent, then its Knowledge tab. Tick the tags this agent is allowed to retrieve from and click Save knowledge access. The footer shows what the agent currently has — No knowledge — prompt only when nothing is selected.

This tab saves independently of the rest of the agent editor: the header’s Save changes button submits the agent form and does not touch knowledge access, and vice versa. The tab is only available on an agent that already exists.

Scoping by tag is what lets one library serve several agents:

• A support agent gets support, returns, shipping — everything a customer with a problem might ask.
• A sales agent gets sales and nothing else, so it can never quote a refund window it shouldn’t.
• An after-hours agent might get only hours.

Deleting a Document

Deleting removes the document and all of its indexed chunks. Agents assigned its tags stop retrieving it immediately. This cannot be undone — re-adding the document re-embeds it from scratch.

Related Pages

• Building & Editing Agents — the agent builder, including the Knowledge tab.
• Flow Builder Concepts — where a flow asks the Knowledge Base a question mid-call.
• AI Voice Agents — agent lifecycle, import/export.
• Connections — for answers that must come from a live system rather than a document.