High-Level Architecture

Relevant source files

The following files were used as context for generating this wiki page:

• agents.py
• main.py
• LLM.py
• searcher/sementic_search.py
• prompts/deep_research_agent_prompts.py
• prompts/review_agent_prompts.py

The Chain-of-Ideas (CoI) Agent is an autonomous research framework designed to generate novel scientific ideas and experimental designs through iterative literature analysis and multi-agent collaboration. The system leverages Large Language Models (LLMs) to perform deep research, simulate peer review, and refine research directions based on a "chain of ideas" extracted from existing scientific publications.

Sources: README.md:10-15, agents.py:108-112

System Components

The architecture is composed of several specialized modules and agents that interact to move from a research topic to a finalized experimental plan.

Core Agents

Component	Description	Key Responsibilities
DeepResearchAgent	The primary orchestration agent for idea synthesis.	Generates search queries, builds research chains, synthesizes novel ideas, and designs experiments.
ReviewAgent	A simulation of a peer reviewer.	Provides critical feedback on generated ideas and experiments to ensure feasibility and novelty.
SementicSearcher	The data retrieval interface.	Interfaces with Semantic Scholar API to find, download, and parse research papers.
openai_llm	The LLM interface wrapper.	Manages synchronous and asynchronous calls to OpenAI or Azure OpenAI services for text generation and embeddings.

Sources: agents.py:65-75, agents.py:108-135, searcher/sementic_search.py:65-80, LLM.py:64-75

Architectural Overview Diagram

The following diagram illustrates the high-level interaction between the user, the core agents, and external research databases.

Functional Logic and Data Flow

The Research Chain Process

The system operates by creating a "chain" of related papers. Starting from an anchor paper or a topic-based search, the DeepResearchAgent performs both forward search (citations) and backward search (references) to understand the evolution of a research trend.

1. Search & Discovery: Uses SementicSearcher.search_async to find initial papers.
2. Chain Building: Iteratively explores references and citations using deep_research_paper_with_chain.
3. Extraction: Extracts entities, summary ideas, and experimental designs from paper content using specialized prompts.
4. Trend Analysis: Summarizes the historical progression from early to late papers in the chain.

Sources: agents.py:240-270, agents.py:300-340, searcher/sementic_search.py:180-210

Sequence of Idea Generation

The following sequence diagram detail the execution flow from initialization to the final result output.

Implementation Details

LLM Integration

The LLM.py file provides a unified interface for model interactions. It supports both Azure and standard OpenAI endpoints and handles exponential backoff through the tenacity library.

• response_async: Used for non-blocking text generation.
• get_embbeding: Used by the SementicSearcher to re-rank papers based on cosine similarity to the research topic.

Sources: LLM.py:140-160, searcher/sementic_search.py:145-160

Prompt Management

Prompts are modularized into specific files based on the agent's role.

• deep_research_agent_prompts.py: Contains prompts for query rewriting, reference extraction, trend analysis, and novelty checking.
• review_agent_prompts.py: Contains prompts for assessing feasibility, novelty, and experimental design.

Sources: prompts/deep_research_agent_prompts.py:10-25, prompts/review_agent_prompts.py:45-55

Data Structures

The primary data structure used for paper metadata is the Result class in sementic_search.py.

class Result:
    def __init__(self,title="",abstract="",article=None,citations_conut = 0,year = None) -> None:
        self.title = title
        self.abstract = abstract
        self.article = article
        self.citations_conut = citations_conut
        self.year = year

Sources: searcher/sementic_search.py:65-72

Summary

The CoI Agent architecture prioritizes the iterative refinement of research ideas by mimicking the scientific process. By combining deep literature traversal (DeepResearchAgent) with adversarial feedback (ReviewAgent), the system ensures that generated outputs are not only relevant but also grounded in existing academic trends and verified for novelty.

Sources: README.md:5-10, agents.py:450-460