Patterns

​

A support assistant that knows your orders

1. Models. Add a chat model named chat.
2. Knowledge. Create a knowledge base help and add your shipping and returns articles.
3. A flow. Build a request-refund flow that creates a refund request and notifies your team.
4. The agent. Create support:
   • Instructions: You help customers with their orders. Answer from the help articles. Be brief. If asked for a refund, use the refund flow and say a person will confirm it.
   • Tools: the orders table with Find and Open; the request-refund flow ticked This flow changes things; the help knowledge base.
   • Who may run it: signed-in users.
   • Limits: most cost $0.25.
5. Your app.

const reply = await pawabase.ai.agents.run("support", question, { thread });

reply = pb.ai.agents.run("support", question, thread=thread)

​

Answers from your own documents

1. Add an embedding model embed and a chat model chat.
2. Create a knowledge base docs and add your pages (web page or Notion).
3. Search and answer:

const hits = await pawabase.ai.knowledge.search("docs", question, { limit: 4 });
const context = hits.map((h) => h.text).join("\n\n");
const answer = await pawabase.ai.chat("chat", [
  { role: "system", content: "Answer using only the notes below. If they don't say, say you don't know.\n\n" + context },
  { role: "user", content: question },
]);
// show answer, and hits[].source as the references

hits = pb.ai.knowledge.search("docs", question, limit=4)
context = "\n\n".join(h["text"] for h in hits)
answer = pb.ai.chat("chat", [
    {"role": "system", "content": "Answer using only the notes below. If they don't say, say you don't know.\n\n" + context},
    {"role": "user", "content": question},
])

​

Search your products by meaning

​

Classify or summarise every new record

​

A research helper that writes to Notion

1. Connect Notion and Tavily under AI → Connections.
2. Create an agent with Search the web (Tavily) and Notion (create pages).
3. Keep Wait for a person after reading anything from outside on. The agent reads the web, then asks before it writes anything. You approve, and the page appears.

​

An assistant you review before it acts