Introduction to CoI Agent

Relevant source files

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

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

The Chain-of-Ideas (CoI) Agent is a multi-agent system designed to revolutionize scientific research by automating the development of novel research ideas using Large Language Models (LLMs). The system systematically explores existing literature, builds "idea chains" representing historical research progression, and leverages evolutionary-style reasoning to propose innovative, feasible, and novel research directions.

Sources: README.md:1-5, agents.py:99-105

System Architecture

The CoI Agent architecture is built upon a modular framework where specialized agents interact with external search APIs and LLM providers to synthesize new knowledge. The primary workflow involves a "Deep Research" phase that traces literature backwards and forwards to establish context before generating a final research proposal.

Core Components

Component	Responsibility	Relevant Files
DeepResearchAgent	Manages the primary workflow of searching, chaining ideas, and generating novel proposals.	agents.py:99-130
ReviewAgent	Evaluates generated ideas and experiments, providing critical feedback for refinement.	agents.py:53-75
SementicSearcher	Interfaces with the Semantic Scholar API to retrieve and parse academic papers.	searcher/sementic_search.py:59-100
openai_llm	Handles communication with OpenAI or Azure OpenAI for all text generation and embedding tasks.	LLM.py:59-110

High-Level Workflow

The following diagram illustrates the high-level process from a research topic input to the final refined experiment.

Deep Research and Idea Chaining

The DeepResearchAgent is responsible for building a chronological sequence of research advancements. It performs both "Forward" searches (finding citations) and "Backward" searches (finding references) to understand how a specific field has evolved.

Idea Chain Logic

1. Anchor Paper Identification: The system starts with an anchor paper related to the topic.
2. Expansion: It searches for future papers that cited the anchor and past papers that the anchor referenced.
3. Extraction: For each paper in the chain, it extracts entities, the core idea (background, novelty, contribution), and the experimental design.
4. Trend Analysis: It summarizes the progression (e.g., "how Paper 0 leads to Paper 1").

Sources: agents.py:228-280, prompts/deep_research_agent_prompts.py:84-120

Evolutionary Reasoning Pattern

The agent uses four modes of thinking to transition from the established trend to a new "Future Direction":

• Reflection: Identifying where current methods fail.
• Analogy: Applying solutions from other fields to the current problem.
• Deep Dive: Modifying specific aspects of existing approaches.
• Imitate: Following the successful progressive strategies of previous researchers.

Sources: prompts/deep_research_agent_prompts.py:175-195

Review and Refinement Mechanism

The system employs a ReviewAgent to ensure the quality and feasibility of the generated output. This agent simulates the peer-review process by analyzing the proposed experiment against established literature.

Technical Implementation Details

Configuration and Environment

The system supports both OpenAI and Azure OpenAI. Configuration is managed via a config.yaml file, which is loaded into environment variables at runtime.

# config.yaml example
SEMENTIC_SEARCH_API_KEY: ""
is_azure: True
AZURE_OPENAI_ENDPOINT: ""
MAIN_LLM_MODEL: "gpt-4o"

Sources: README.md:40-65, main.py:10-20

PDF Parsing

The project utilizes scipdf_parser and grobid to convert raw PDF research papers into structured JSON dictionaries containing titles, abstracts, sections, and references.

# searcher/sementic_search.py:246-255
def read_arxiv_from_path(self, pdf_path):
    if not os.path.exists(pdf_path):
        return None
    try:
        article_dict = scipdf.parse_pdf_to_dict(pdf_path)
    except Exception as e:
        return None
    return article_dict

Sources: searcher/sementic_search.py:246-255

Data Models

The Result class in searcher/sementic_search.py serves as the primary data structure for academic papers, storing metadata required for ranking and chaining.

Field	Type	Description
title	String	The paper title.
abstract	String	A summary of the paper.
article	Dict	The full parsed content of the PDF.
citations_count	Int	Number of citations for ranking relevance.
year	Int	Publication year for chronological ordering.

Sources: searcher/sementic_search.py:49-55

Conclusion

The CoI Agent represents a sophisticated application of multi-agent orchestration for scientific discovery. By grounding its generation process in actual literature chains and utilizing a rigorous review-refine loop, it minimizes hallucination and maximizes the novelty of its research proposals. This system provides a robust framework for researchers to explore new domains and generate high-quality hypotheses autonomously.