LlamaIndex¶

LlamaIndex LLM classes accept a base URL. Point it at Privyx and the prompts LlamaIndex builds, including retrieved context, are masked before they reach the provider.

OpenAI models¶

privyx proxy --upstream https://api.openai.com

from llama_index.llms.openai import OpenAI

llm = OpenAI(model="gpt-4o-mini", api_base="http://localhost:8000/v1")

print(llm.complete("Write a short greeting to alice@example.com"))

OPENAI_BASE_URL is not enough

LlamaIndex's OpenAI class reads OPENAI_API_BASE. With only OPENAI_BASE_URL set, it calls the provider directly and nothing is masked. Pass api_base, or set OPENAI_API_BASE.

Anthropic models¶

privyx proxy --upstream https://api.anthropic.com

from llama_index.llms.anthropic import Anthropic

llm = Anthropic(model="claude-opus-5-5", base_url="http://localhost:8000")

print(llm.complete("Write a short greeting to alice@example.com"))

Any OpenAI-compatible model¶

For a model name the OpenAI class does not know, use OpenAILike:

from llama_index.llms.openai_like import OpenAILike

llm = OpenAILike(
    model="my-model",
    api_base="http://localhost:8000/v1",
    api_key="unused",
    is_chat_model=True,
)

print(llm.complete("Write a short greeting to alice@example.com"))

is_chat_model=True matters: without it, OpenAILike uses the legacy completions API, a path Privyx does not mask.

Good to know¶

Embeddings are not masked. Building an index sends your documents to /v1/embeddings, which Privyx forwards as sent. Mask the documents first with privyx mask, embed with a local model, or refuse the path with proxy.passthrough_unknown: false.
Retrieved context is masked. What a query engine puts into the prompt goes through the chat endpoint, like any other message.