knowledge_base_report.json

Overview

knowledge_base_report.json defines a sophisticated Knowledge Base Retrieval Q&A Agent designed for generating research reports by leveraging a local knowledge base. It encapsulates the configuration of an intelligent agent system that performs advanced task planning, decomposition of user queries, multi-perspective research, iterative retrieval, integration of evidence, and structured report generation.

The system is particularly recommended for academic research paper question-answering scenarios, providing in-depth, verified, and structured responses based strictly on retrieved knowledge base content without fabrication.

Detailed Components Explanation

This JSON file is structured primarily as a DSL (Domain Specific Language) representation describing an agent-based workflow with components, their parameters, connections, and execution logic.

Top-Level Fields

• id: Unique numeric identifier (20).

• title: Object with multilingual titles:

  • en: "Report Agent Using Knowledge Base"

  • zh: "知识库检索智能体"

• description: Multilingual description of the agent's purpose.

• canvas_type: "Agent" indicating this is an agent workflow design.

• dsl: The core definition describing components, globals, graph topology, and operational parameters.

• avatar: Base64-encoded PNG image representing the agent's icon.

DSL Breakdown

1. Components

The key elements of this DSL are its components, which represent the nodes in the agent's workflow graph.

• Begin

  • Component Type: Begin

  • Purpose: Entry point for the conversational interaction.

  • Parameters:

    • enablePrologue: true — enables a greeting message.

    • prologue: "你好！ 我是你的助理，有什么可以帮助到你的吗？" (Chinese for "Hello! I am your assistant, how can I help you?")

    • mode: "conversational" — the interaction mode.

  • Connections:

    • Downstream: Agent:NewPumasLick

    • Upstream: None (start node)

• Agent:NewPumasLick

  • Component Type: Agent

  • Purpose: The core Knowledge Base Retrieval Q&A Agent.

  • Parameters: Extensive configuration for how the agent handles queries.

    • LLM Settings:

      • llm_id: "qwen3-235b-a22b-instruct-2507@Tongyi-Qianwen"

      • temperature: 0.1 (low randomness for consistency)

      • max_tokens: 128000 (very large token limit)

      • max_retries: 3

      • max_rounds: 3

    • Penalties:

      • frequency_penalty: 0.5 (disabled by default)

      • presence_penalty: 0.5 (disabled by default)

    • Message History and Prompting:

      • message_history_window_size: 12

      • prompts: User prompt template uses {sys.query} placeholder.

      • sys_prompt: Very detailed system prompt describing the agent's role, execution framework, quality checks, core principles, LaTeX usage, failure handling, and language handling.

    • Tools:

      • A retrieval tool configured to query the knowledge base with parameters like similarity threshold (0.2), top-k results (1024), and top-n results (8).

    • Outputs:

      • Produces a string output in the content field.

  • Connections:

    • Upstream: begin

    • Downstream: Message:OrangeYearsShine

    • Tool: Links to Tool:AllBirdsNail (purpose not elaborated)

• Message:OrangeYearsShine

  • Component Type: Message

  • Purpose: Delivers the agent's generated content as a message output.

  • Parameters:

    • content: Dynamic field referencing the agent's output ({Agent:NewPumasLick@content})

  • Connections:

    • Upstream: Agent:NewPumasLick

    • Downstream: None (end node)

• Tool:AllBirdsNail (Referenced but not connected visibly in main flow)

  • Component Type: Tool

  • Description: "This is an agent for a specific task."

  • Purpose: Likely auxiliary or specialized processing tool for the agent.

  • No explicit connections in main flow except linked as a tool to the agent.

2. Globals

Global system variables used in the workflow:

• sys.conversation_turns: Initialized to 0; tracks number of dialogue turns.

• sys.files: Empty list; possible placeholder for user-uploaded files or references.

• sys.query: Empty string; placeholder where user's question input is stored.

• sys.user_id: Empty string; user identification placeholder.

3. Graph Structure

The workflow graph defines nodes and edges representing the flow of data and control.

• Nodes: begin, Agent:NewPumasLick, Message:OrangeYearsShine, Tool:AllBirdsNail

• Edges:

  • begin → Agent:NewPumasLick

  • Agent:NewPumasLick → Message:OrangeYearsShine

  • Agent:NewPumasLick tool edge → Tool:AllBirdsNail (auxiliary tool invocation)

Important Implementation Details and Algorithms

Core Algorithmic Flow Described in System Prompt

The agent's behavior is meticulously defined in the sys_prompt string with explicit stages:

1. Assessment & Decomposition: Automatically extracting key elements (topics, entities, scope) and decomposing the query into 5-20 factual data points.

2. Query Type Determination: Rule-based decision between:

   • Depth-first approach for method comparisons or multiple explanations.

   • Breadth-first for 3+ independent sub-questions.

   • Simple query for straightforward fact/specification questions.

3. Research Plan Formulation: Tailors search strategies based on query type, defining perspectives, keywords, document types, and output format.

4. Retrieval Execution: Uses the linked retrieval tool to fetch relevant knowledge base content, with automatic query rewriting and retry loops if quality or coverage standards are unmet.

5. Integration & Reasoning: Constructs responses as fact-evidence-reasoning chains, attaching 1-2 strongest supporting evidence items per conclusion.

Quality Gate Checklist

The agent performs verification at three stages:

• Decomposition: checks key concepts and data points.

• Retrieval: ensures quality, diversity, and timeliness of sources.

• Generation: verifies evidential support, assumptions, and user expectations.

Core Principles

• Strict reliance on explicitly retrieved knowledge base content.

• No generation or fabrication of information.

• Preference for accuracy, even if incomplete.

• Structured, logical, professional output formatting.

• Support for LaTeX in responses.

Failure and Interaction Strategy

• Explicit user notification if critical facts are missing.

• Time filtering and indication of retrieval date for time-sensitive queries.

• Responds in user's preferred language.

Usage Example

While this file itself is configuration data, here is a conceptual usage example:

{
  "sys.query": "Compare the effectiveness of different AI models for academic research.",
  "sys.user_id": "user_123"
}

• The begin node initiates the interaction with a greeting.

• The query passes to Agent:NewPumasLick which decomposes the question, determines depth-first strategy, formulates research plans, retrieves relevant documents, and synthesizes a structured report.

• The final report content is passed to Message:OrangeYearsShine which outputs the answer to the user interface.

Interaction with Other System Parts

• Knowledge Base: The agent relies on a local knowledge base accessed via the Retrieval tool component.

• Language Model: Uses a specific LLM instance (qwen3-235b-a22b-instruct-2507@Tongyi-Qianwen) for reasoning and text generation.

• User Interface: The Begin and Message components interface with the user, receiving queries and delivering answers.

• Auxiliary Tools: The Tool:AllBirdsNail component could represent additional processing or external calls, integrated with the agent as needed.

This file serves as a self-contained agent workflow module within a larger AI assistant platform supporting multi-turn conversational Q&A with knowledge base grounding.

Mermaid Flowchart Diagram

The following Mermaid flowchart illustrates the main components and their relationships in this file:

flowchart LR
    Begin[Begin: Conversation Start]
    Agent[Agent: Knowledge Base Retrieval Q&A Agent]
    Message[Message: Output Response]
    Tool[Tool: Auxiliary Processing]

    Begin --> Agent
    Agent --> Message
    Agent -- uses --> Tool

• Begin initiates the conversation.

• Agent processes user queries with complex reasoning and retrieval.

• Message delivers final content.

• Tool is an auxiliary component employed by the agent.

Summary

knowledge_base_report.json is a comprehensive agent configuration file defining a powerful, multi-stage knowledge base retrieval and report generation assistant. It combines advanced query decomposition, rule-based planning, iterative retrieval, and evidence-based synthesis to serve academic research needs, ensuring high accuracy and professional output constrained strictly by available knowledge base data. This file is a critical module for enabling structured, grounded Q&A interactions within a broader AI assistant ecosystem.