What happens if a spend request is rejected?

The wallet.pay() call returns status: rejected with a reasons array. The workflow stops before the paid step executes — no money moves and the decision is recorded in the audit trail.

What is the difference between queued and rejected?

Queued means the spend intent exceeded the approval threshold and is waiting for human review in the NORNR control room. Rejected means policy blocked it outright — counterparty not allowed, daily limit exceeded, or mandate not matched.

How is this different from API rate limiting?

Rate limiting caps request volume after the fact. NORNR governs spend intent before money moves — checking amount, counterparty and mandate in real time so blocked or queued requests never reach the paid API.

Do I need to rewrite my existing agent code?

No. Add one wallet.pay() call before the paid step and branch on the returned status. Your existing workflow logic stays unchanged.

How to set a cost limit on a LlamaIndex agent

1. Why LlamaIndex agents need cost limits

LlamaIndex is built for retrieval-augmented generation and agent reasoning over large document sets. When those agents call paid LLMs, embedding APIs or external data services, there is no built-in layer that asks whether the cost is within a mandate before the query executes.

NORNR adds that layer. Place the policy check before any paid tool call and the agent gets a clear approved, queued or rejected outcome. Rejected or queued decisions never reach the paid API.

2. Install the SDK

bash

pip install agentpay llama-index-core llama-index-llms-openai

3. Create the governed wallet

python

from agentpay import Wallet

wallet = Wallet.create(
    owner="llamaindex-rag-agent",
    daily_limit=60,
    require_approval_above=20,
    base_url="https://nornr.com",
)

4. Gate the query engine or tool call

python

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
from llama_index.llms.openai import OpenAI

decision = wallet.pay(
    amount=10.00,
    to="openai",
    purpose="rag query over internal docs",
)

if decision.get("status") == "approved":
    documents = SimpleDirectoryReader("./data").load_data()
    index = VectorStoreIndex.from_documents(documents)
    query_engine = index.as_query_engine(llm=OpenAI(model="gpt-4o-mini"))
    response = query_engine.query("What are the key risk factors?")
    print(response)
elif decision.get("status") == "queued":
    print("Query queued for operator review.", decision)
else:
    print("Query blocked by policy.", decision)

The pattern works the same for agent tool calls — wrap the NORNR check around any step that would result in a paid API interaction.

5. Decision output

output

{"status": "approved", "requiresApproval": false}
{"status": "queued",   "requiresApproval": true}
{"status": "rejected", "reasons": ["counterparty_not_allowed"]}

6. Common mistakes

Calling the query engine before the NORNR check — the cost already incurred at that point.
Setting the approval threshold higher than the daily limit.
Not logging the decision object alongside the query result for audit purposes.