Deep Research Agent

Relevant source files

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

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

The Deep Research Agent is a core component of the Chain-of-Ideas (CoI) framework designed to automate scientific discovery. Its primary purpose is to perform exhaustive literature reviews, identify research trends, and generate novel, feasible research ideas and experimental designs. By leveraging Large Language Models (LLMs) and semantic search engines, it simulates the reasoning patterns of human researchers to transition from existing methodologies to innovative future directions.

Sources: agents.py:101-131, prompts/deep_research_agent_prompts.py:177-219, README.md:5-10

Architecture and Core Logic

The Deep Research Agent operates through a multi-stage pipeline that includes paper discovery, idea chain construction, novelty checking, and experimental design. It utilizes a SementicSearcher to interact with external academic graphs (like Semantic Scholar) and multiple LLM instances (Main and Cheap) for complex reasoning and data extraction tasks respectively.

Deep Research Chain Workflow

The agent constructs a "research chain" by traversing paper citations (forward search) and references (backward search) to establish a historical context for a specific topic.

The diagram above illustrates the iterative process of building a research chain by traversing academic relationships.
Sources: agents.py:221-322, prompts/deep_research_agent_prompts.py:116-175

Key Components

1. Literature Search and Reranking

The agent uses semantic search to find relevant literature. If a query fails to yield results, the agent employs a re-writing logic to generate more concise or general queries. Results are often reranked using cosine similarity between query embeddings and paper content.

Component	Function	File Reference
get_search_query	Generates search strings based on topic and idea.	agents.py:133-145
SementicSearcher.search_async	Executes API calls to find papers and downloads PDFs.	searcher/sementic_search.py:141-180
rerank_papers	Reorders candidates based on LLM embeddings.	searcher/sementic_search.py:117-135

2. Idea Generation and Novelty Validation

The agent generates ideas by analyzing the "Idea Chain" and "Trend Summary." To ensure scientific integrity, it performs a strict novelty check against existing papers to avoid plagiarism.

Functional Implementation

Chain Construction Logic

The deep_research_paper_with_chain function is the primary driver for historical analysis. It maintains a list of idea_chain, experiments, and total_entities extracted from the literature.

• Forward Search: Uses search_related_paper_async with need_reference=False to find papers citing the current anchor.
• Backward Search: Iterates through extracted references and searches for them specifically.
• Relevance Judgment: Every discovered paper is passed through a judge prompt to ensure it contributes to the specified research topic.

Sources: agents.py:221-295, prompts/deep_research_agent_prompts.py:155-175

Experimental Design and Refinement

After an idea is finalized, the agent generates a detailed experimental plan. This plan includes:

1. Initial Design: Created via generate_experiment using past related research experiments as context.
2. Review Feedback: A ReviewAgent assesses the plan for feasibility and clarity.
3. Refinement: The agent uses refine_experiment to incorporate suggestions, potentially searching for new technical information to solve specific implementation hurdles identified by the reviewer.

Sources: agents.py:377-407, prompts/deep_research_agent_prompts.py:327-362, prompts/review_agent_prompts.py:46-72

Data Structures and Configuration

The agent relies on a Result data structure to track paper metadata and a central configuration file for API management.

# From searcher/sementic_search.py:61-67
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

Configuration Parameter	Description	Default
MAIN_LLM_MODEL	Powerful model (e.g., GPT-4o) for reasoning.	gpt4o-0513
CHEAP_LLM_MODEL	Efficient model for summarization/extraction.	gpt-4o-mini
max_chain_length	Maximum papers in a single research chain.	5
improve_cnt	Number of refinement iterations for experiments.	1

Sources: agents.py:12-25, main.py:31-41, searcher/sementic_search.py:61-67

The Deep Research Agent serves as the intellectual engine of the CoI-Agent repository, transforming raw academic data into structured research trajectories. By automating the transition from "what exists" to "what could be," it significantly reduces the manual labor required for the initial stages of the scientific process.