GraphScholar is a research assistant for graph learning and GraphRAG papers. It turns local PDF papers into structured metadata, chunk-level evidence, and a graph-based literature store for retrieval and query answering.

• parses PDFs from data/
• extracts title, abstract, tasks, applications, datasets, method summary, and contribution summary
• builds paper summaries and chunk-level evidence
• constructs a paper-topic-method-task-application-dataset graph
• answers research questions with retrieval and graph querying
• saves each run into outputs/answer/

• outputs/paper_summaries.json
• outputs/paper_chunks.json
• outputs/paper_graph.json
• outputs/paper_metadata_cache.json
• outputs/answer/*.md

1. paper_organization.py reads PDFs and extracts structured metadata.
2. src/build_graph.py builds the paper graph from the summaries.
3. src/tools.py handles paper search, chunk search, and graph queries.
4. src/agent.py routes the question, gathers evidence, and produces the answer.
5. run_agent.py runs demo or single-question mode and writes a report.

Each paper summary includes:

• title
• abstract
• tags
• category
• paper_type
• tasks
• applications
• datasets
• method_summary
• contribution_summary
• confidence

Rebuild the paper store:

python paper_organization.py

Run in local deterministic mode:

python run_agent.py --no-llm

Run one question:

python run_agent.py --question "If I am working on GraphRAG, help me organize the most representative papers in recent years by method, evaluation, and survey."

Run the preset demo set:

python run_agent.py --demo

src/llm_client.py uses explicit in-code settings:

• DEFAULT_BASE_URL
• DEFAULT_API_KEY
• DEFAULT_MODEL_ID

Fill those values before using LLM mode.

This is not a plain keyword search tool. It combines:

• structured paper extraction
• graph-aware literature organization
• task/application/dataset-aware retrieval
• chunk evidence with page references
• saved answer reports for review and reuse

The project is intentionally lightweight, but it already behaves like a compact research workflow for graph literature analysis.