LlamaIndex Integration

Safety callbacks and LLM wrappers for LlamaIndex applications.

Installation

pip install sentinelseed[llamaindex]

Components

Quick Start

Option 1: Global Callback Handler

from llama_index.core import Settings
from llama_index.core.callbacks import CallbackManager
from sentinelseed.integrations.llamaindex import SentinelCallbackHandler

handler = SentinelCallbackHandler(
    seed_level="standard",
    on_violation="log",  # log, raise, flag
)

Settings.callback_manager = CallbackManager([handler])

index = VectorStoreIndex.from_documents(documents)
response = index.as_query_engine().query("Your question")

print(handler.get_stats())

Option 2: Wrap LLM

from llama_index.llms.openai import OpenAI
from llama_index.core import Settings
from sentinelseed.integrations.llamaindex import wrap_llm

Settings.llm = wrap_llm(
    OpenAI(model="gpt-4o"),
    seed_level="standard",
    inject_seed=True,
)

Option 3: SentinelLLM Directly

from llama_index.llms.openai import OpenAI
from sentinelseed.integrations.llamaindex import SentinelLLM

sentinel_llm = SentinelLLM(
    llm=OpenAI(model="gpt-4o"),
    seed_level="standard",
    inject_seed=True,
    validate_input=True,
    validate_output=True,
)

response = sentinel_llm.chat(messages)
response = sentinel_llm.complete(prompt)

Option 4: Quick Setup

from sentinelseed.integrations.llamaindex import setup_sentinel_monitoring

handler = setup_sentinel_monitoring(
    seed_level="standard",
    on_violation="log",
)

Callback Events

Configuration

SentinelCallbackHandler(
    seed_level="standard",
    on_violation="log",
    event_starts_to_ignore=[],
    event_ends_to_ignore=[],
)

SentinelLLM(
    llm=base_llm,
    seed_level="standard",
    inject_seed=True,
    validate_input=True,
    validate_output=True,
)

Links

• LlamaIndex Docs