promptsage 0.0.1

Promptsage helps you combine all the components of your prompt into a single LLM-ready payload

Promptsage

Promptsage is an LLM prompt builder with AI safety guardrails. Apply fine-grained access control over the context, and prevent unwanted behavior with filters. Compatible with all major datastores, LLMs and gateways.

Quick Install

With pip:

pip install promptsage

Examples

from promptsage import text_prompt, load_examples
from promptsage.sources import PDF, LangchainDocuments
from promptsage.filters import PromptInject, Anonymize, TrimTokens

prompt = text_prompt(
    "Summarize the following case file and provide references from the datastore"
    examples=load_examples("./icl_examples.json")
    sources=[
        PDF('case_file.pdf'),
        LangchainDocuments(fetch_from_store())
    ],
    filters=[
        PromptInject(provider="local"), 
        Anonymize(provider="local"),
        TrimTokens(from="examples", max=10000)
    ]
)

# Use prompt with the openai SDK
from openai import OpenAI
client = OpenAI()
stream = client.chat.completions.create(
    model="gpt-4",
    messages=prompt.to_openai_messages(),
)

# Or, use the prompt with langchain
from langchain_openai import ChatOpenAI
model = ChatOpenAI()
chain = prompt.to_langchain_messages() | model
chain.invoke()

Access Control

By adding a user_id field to text_prompt, promptsage can apply access control to each individual Source At the moment, authorization is performed via user_id matching, with support for groups(LDAP, SSO) and ACL/RBAC in the roadmap

From our unit tests:

    sources = [
    EchoSource("User 1 knows that the password is XXX", "user1"),
    EchoSource("User 2 knows that the password is YYY", "user2"),
    EchoSource("User 3 knows that the password is ZZZ", "user3"),
    ]
    with pytest.raises(UnauthorizedError):
        text_prompt(
            "What do I know?"
            sources=sources,
            user_id="user2",
        )

Instead of raising an exception, you may also exclude unauthorized sources:

    text_prompt(
        "What do I know?"
        sources=sources,
        user_id="user2",
        access_control_policy=AccessControlPolicy.skip_unauthorized,
    )

Datastores

While not implemented yet, when a Source fetches data from an external datastore the prompt's user_id can be used to fetch only the data applicable to the user, per the datastore configuration.

Basic Concepts

A typical LLM invocation is comprised of three parts:

• The user prompt, instructions for the LLM
• The context, all accompanying data necessary for the LLM to complete generation
• The examples, for in-context learning

promptsage introduces the following contexts

• A Source is a chunk of text within the context
• A Filter is applied over the entire prompt to catch undesired inputs
• A Prompt is a sanitized prompt ready to be used with the target LLM

Current Features

• Compatible with openai SDK and langchain formats
• Access control only applies with user_id matching
• Filters: Prompt Injection

Roadmap

• llama_index support
• Acess Control: LDAP & SSO integration
• Vectorstores: Langchain, Pinecone, Milvus
• Filters:
  • PII anonymization
• Templates:
  • Prompt compression