LlamaIndex

Option 1: install() (recommended)

LlamaIndex’s OpenAI and Anthropic LLMs call the official SDKs, so install() covers them, caching included:

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

parry.install()
Settings.llm = OpenAI(model="gpt-4o-mini")

index = VectorStoreIndex.from_documents(SimpleDirectoryReader("docs/").load_data())
engine = index.as_query_engine()
engine.query("What does max_cost_usd do?")
engine.query("What does max_cost_usd do?")    # the LLM call is a cache hit

Option 2: ParryHandler

The handler applies rate limiting, the budget and cost tracking at the LlamaIndex event level:

from llama_index.core import Settings
from llama_index.core.callbacks import CallbackManager
from parry import Parry
from parry.integrations.llamaindex import ParryHandler

guard = Parry(max_cost_usd=1)
Settings.callback_manager = CallbackManager([ParryHandler(guard)])

engine = index.as_query_engine()
engine.query("...")
print(guard.stats.summary())

The handler reads token usage from the raw provider response, then from additional_kwargs, and estimates with tiktoken as a last resort. BudgetExceeded raised in on_event_start stops the query before the LLM is called.

ServiceContext is gone

Older LlamaIndex tutorials use ServiceContext.from_defaults(callback_manager=...). LlamaIndex has removed it. Use Settings.callback_manager, or pass callback_manager= to the LLM directly.

Using both

As with LangChain, you can use both with the same guard without double counting:

guard = parry.install()
Settings.callback_manager = CallbackManager([ParryHandler(guard)])

Embeddings calls (OpenAIEmbedding) are not intercepted in v0.1. Only LLM calls are.