agentUniverse Documentation

The PEER Financial Training Dataset is designed for agent training, pre-training, and supervised fine-tuning in financial analysis scenarios. It includes 100 common financial questions and 100 sets of inputs and outputs specifically for the Planning Agent and the Expressing Agent within the PEER framework.

Dataset Components

• FinQA: A collection of 100 professional questions categorized into nine areas: information query, general financial QA, report interpretation, target analysis, strategy advice, major events interpretation, macro analysis, market analysis, and policy interpretation.
• Planning Agent Data: Contains the user's original query (input) and the resulting sub-questions generated by the agent (output), which act as an interpretation framework.
• Expressing Agent Data: Contains the answers provided by the Executing Agent for each sub-question (input) and the final professional synthesized/rewritten answer (output).

Note: Data for the Executing Agent is currently undergoing compliance approval and is not yet available.

The agentUniverse User Guide provides a structured path for developers to build agentic applications. The learning path is divided into three main stages:

1. Getting Started: Covers installation, project structure, running tutorials, building single agents, and multi-agent applications. It also includes advanced techniques like memory management, prompt management, and multimodal agents.
2. Principle Introduction: Explains the core domain components (Agent, LLM, Tool, Knowledge, Memory, Planner) and technical components (RAG, Service registration, Web APIs, Logging, and Data Collection).
3. Component Reference Manual: Provides detailed usage instructions for specific implementations, such as various LLM providers (OpenAI, Claude, Ollama, etc.), tools (LangChain wrappers), and storage backends (Milvus, ChromaDB, SQLDB).

For rapid development, you can also explore the Product Platform for a visual agentic workflow experience.

A Reader is a component responsible for extracting information from various sources (such as local files, web pages, or I/O interfaces) and converting them into the Document format used throughout agentUniverse.

To create a custom Reader, you must subclass the Reader class and override the _load_data method. The _load_data method must return a List[Document].

from abc import abstractmethod
from typing import List, Any, Optional

from agentuniverse.agent.action.knowledge.store.document import Document
from agentuniverse.base.component.component_base import ComponentEnum
from agentuniverse.base.component.component_base import ComponentBase

class Reader(ComponentBase):
    """The basic class for the knowledge reader."""
    component_type: ComponentEnum = ComponentEnum.READER
    name: Optional[str] = None
    description: Optional[str] = None

    def load_data(self, *args: Any, **kwargs: Any) -> List[Document]:
        """Load data from the input params."""
        return self._load_data(*args, **kwargs)

    @abstractmethod
    def _load_data(self, *args: Any, **kwargs: Any) -> List[Document]:
        """Load data from the input params."""
        pass

A QueryParaphraser is a component responsible for refining a Query through processes like rewriting, splitting, or keyword extraction. This refinement helps retrieve more accurate and rich content from a Store.

Key Concepts:

• Input/Output Symmetry: Both the input and the output of a QueryParaphraser must be a Query object. This allows for chaining multiple layers of paraphrasing (e.g., one component extracts keywords, and another rewrites the text).
• The Query Object: The Query object is the data contract used throughout the paraphrasing pipeline. It carries the original string, rewritten text bundles, images, extracted keywords, and embeddings.

from typing import Optional, List, Set
from pydantic import BaseModel, Field
from PIL.Image import Image

class Query(BaseModel):
    query_str: Optional[str] = None
    query_text_bundles: Optional[List[str]] = Field(default_factory=list)
    query_image_bundles: Optional[List[Image]] = Field(default_factory=list)
    keywords: Optional[Set[str]] = Field(default_factory=set)
    embeddings: List[List[float]] = Field(default_factory=list)
    ext_info: dict = {}
    similarity_top_k: Optional[int] = None