Knowledge Base
Grounds the agent in your real content. The setup form’s “Where’s your info?” step and the bot’s Knowledge panel both feed this node — on the canvas it’s the dashed wire into the agent’s Knowledge & tools port.
- Out (top, dashed) → an AI Agent’s resource port.
Sources
Section titled “Sources”| Kind | How | Notes |
|---|---|---|
| Web page | Paste a URL, Fetch page | Fetched server-side, stripped to text. Re-fetch any time with the refresh button. Private/internal addresses are refused. |
| Pasted text | Paste text with a name | Policies, FAQ answers, product sheets — anything textual. |
| Upload PDF (max 5 MB) | Text is extracted server-side; only the text is stored, never the file. Scanned/image-only PDFs have no extractable text and are rejected honestly. |
Caps (enforced server-side, shown in the panel): 50k characters per source,
100k characters per bot. The node card shows the truth at a glance —
2 sources · 18k chars.
Runtime behavior
Section titled “Runtime behavior”- Up to 24k characters of knowledge are injected into the agent’s instructions per reply, with the source list and an instruction to ground answers in it — and to escalate to a human rather than invent when the answer isn’t there.
- When a bot’s stored content exceeds the per-reply budget, excerpts are picked by keyword overlap with the customer’s message (no embeddings — the selection is inspectable in the test-chat trace, which reports exactly what was injected).
- Replies grounded this way are tagged in conversation history
(e.g.
KB · 2 sources) and show up in usage breakdowns.
Scope hints
Section titled “Scope hints”The old topic chips still exist as optional scope hints — topic names added to the instructions as guardrails. They carry no content; real grounding comes from sources.