# Knowledge base grounding prompts

Use these snippets when building or splitting knowledge bases for HighLevel Conversation AI, Voice AI or Super Agents.

---

## KB document header (prepend to each article)

```
TITLE: {short descriptive title}
SCOPE: {when the agent should use this doc}
OUT OF SCOPE: {when to ignore this doc}
LAST REVIEWED: {YYYY-MM-DD}
---

```

## Splitting rule

One document = one decision or one procedure. Bad: "Everything about our services." Good: "How to book a scoping call" and "Support SLA for existing clients" as separate docs.

## Retrieval instruction (add to system prompt)

```
When answering, search the knowledge base for the most specific document first.
Prefer documents whose SCOPE matches the contact's question.
If two documents conflict, say there is conflicting information and offer a human handoff — do not pick one silently.
Cite the document TITLE internally in your reasoning; do not expose filenames to the contact.
```

## FAQ format (works well for RAG)

```
Q: {exact question a customer asks}
A: {answer in 2–4 sentences}
NOTES: {edge cases, links, "if X then Y"}
```

## Anti-hallucination footer (append to sensitive docs)

```
This document is authoritative for {topic} only.
Do not extrapolate to {related topics not covered}.
If the contact's situation is not described here, escalate.
```

## Voice AI variant (shorter chunks)

Keep each chunk under 120 words. Lead with the spoken answer first; put caveats in a second paragraph marked NOTES (Voice AI may skip NOTES unless asked).