LLM Advisory

This framework allows to easily create multiple advisors using large language models. It uses a simple agent workflow which can be customized.

LLM Advisory uses langchain and langgraph for model and workflow handling, pydantic for data models. It is easily extendable and new advisors can be created by extending the core advisor. Support for OpenAI and Ollama is available.

The advisory object is created by llm_advisory = LLMAdvisory(advisors=[DefaultAdvisor()]) with at least on advisor.

The advisors are invoked with a custom input prompt (or a default prompt to return a advisory) by calling llm_advisory.get_advisory(...) which returns a pydantic model representing a signal containing the advisory.

what can this be used for?

LLM Advisory is a generic toolkit for topic based advisories. Each advisor takes the same raw data but focuses on its specialty.

By dividing the work, you:

1. Avoid confusion — each advisor has a narrow, well-defined role.
2. Get clearer signals — instead of one muddled answer, you get targeted, high-confidence advice from the “right” advisor.

Projects using LLM Advisory:

• Backtrader LLM Advisory - A backtrader trading strategy advisory, which generates trade advises. Multiple advisors are available which generate data from a running strategy.

Quickstart

Set up the LLMAdvisory with ChatGPT:

from llm_advisory import LLMAdvisory
from llm_advisory.advisors import DefaultAdvisor
from llm_advisory.pydantic_models import LLMAdvisorDataArtefact

llm_advisory = LLMAdvisory(
    model_provider="openai",
    model_name="gpt-4o",
    model_config={"OPENAI_API_KEY": "xxx"},
    advisors=[DefaultAdvisor()]
)

Ask the advisory for a advise:

advisory_response = llm_advisory.get_advisory(
    message="Is this a positive or negative message",
    data=[
        LLMAdvisorDataArtefact(
            description="Message",
            artefact="This is a positive message",
        )
    ]
)

Processing the advise:

print(
    advisory_response.advise.signal,
    advisory_response.advise.confidence,
    advisory_response.advise.reasoning,
)

Advisors

• DefaultAdvisor: Default advisor with no speciality
• PersonaAdvisor: Default advisor with additional personality specified

State Advisors

• AdvisoryAdvisor: Advisor that creates the final advisory

Examples

TODO add text how to create a basic advisory object

• Provide a advise from advisors based on provided data (moral)
• Create a advisor based on a famous trading Persona or a fictional Persona with a custom Persona description ()

Frequently Asked Questions

• My advisory is running slow when using ollama Set max_concurrency to a lower value when creating the adivsory.

Future functionality

A list with possible future functions.

• Flexible workflow
• Fetch additional data from provided sources, per advisor
• compile experience into memory for self improvement
• add observer for advisors to overwatch their behaviour
• add feedback for advises from trading so they can learn