LiteratureGraphContext

Structure paper PDFs into a knowledge graph (Paper, Method, Dataset, Claim, Limitation, Evidence) and expose it via MCP for literature review workflows.

Licensed under the MIT License.

Design philosophy

LiteratureGraphContext is a structured evidence layer for literature review—not an AI that writes your related work or discovers research gaps for you.

What LGC owns

• Graph connections — Papers link through shared tasks, methods, datasets, citations (CITES), contrasts (CONTRASTS_WITH), and extensions (EXTENDS). Multiple studies stay connected across topics instead of living in isolated PDFs.
• Evidence-backed facts — Claims, limitations, and contributions are stored with provenance: paper_id, page, section, and evidence_text.
• MCP as the data plane — Tools return structured lists, comparisons, and matrices grounded in the graph.

What the connected agent owns

• Interpreting limitations as research gaps
• Drafting related work prose
• Open-ended synthesis and recommendations

A typical agent flow:

User: "What are the gaps in prediction research?"

  Agent → search_papers("prediction research")
        → find_limitations("prediction")   # or paper_id= from search
        → explore_paper_graph(paper_id, hops=1|2)
        → compare_papers([...])

  Agent synthesizes the answer, citing paper_id / page / evidence_text from MCP

LGC does not pre-compute narrative conclusions. It makes the graph and evidence reliable enough for agents to reason over. See ROADMAP.md for how known limitations are prioritized under this model.

Documentation

• Tutorial — install → setup → index → MCP → viz → Zotero (full walkthrough)
• Roadmap — planned features and known gaps

Quick start

pip install literature-graph
litgraph setup --papers-dir ./my-papers
# or non-interactive:
# litgraph init --papers-dir ./my-papers && litgraph index -y
litgraph serve-mcp

For a clone / development install: pip install -e ".[dev]".

litgraph setup walks through project init, LLM / API keys (~/.litgraph/.env), optional Zotero, MCP client config, and a first index.
litgraph index alone runs scan → parse → extract → build.

For the full flow (including Kuzu single-writer rules and Windows daemon HTTP), see the Tutorial.

Where to put papers

Place PDF, BibTeX, and Markdown note files under your papers folder:

my-papers/
  mobility_gnn_2024.pdf
  mobility_gnn_2024.bib
  mobility_gnn_2024_notes.md

Configure the folder path

Method	Command
At init	litgraph init --papers-dir ./my-papers
On scan (auto-save)	litgraph scan ./my-papers
Explicit set	litgraph config set papers_dir ./my-papers
Manual edit	.litgraph/config.yaml → papers_dir

Relative paths are resolved from the project root. Absolute paths are also supported.

After setup, litgraph index (or the individual scan / parse / extract / build commands) uses the configured papers_dir when no path argument is given.

Verify MCP tools: litgraph test-mcp

Supported file types

Type	Role
.pdf	Full parse → LLM extract → graph
.md	Paper notes (# Abstract, ## Introduction, etc.) → extract → graph
.bib	Metadata only (title, authors, venue, doi) merged into graph at build

Configuration

Each repository has its own project at <repo>/.litgraph/ (config, cache, graph DB). Run litgraph init once per repository before other commands.

Location	Purpose
<repo>/.litgraph/config.yaml	Project settings (papers_dir, LLM, database)
<repo>/.litgraph/cache/, db/	Per-project scan cache and graph
~/.litgraph/.env	Preferred place for API keys (OPENAI_API_KEY, ZOTERO_*, …)
~/.litgraph/logs/	Global log files
<repo>/.env	Optional local override (do not commit)

Important: ~/.litgraph is for API keys and logs only—it is not a project. Commands like index / scan require an initialized project under your repository. If you see unexpected paths or split graph data, run litgraph doctor.

Project resolution order:

1. LITGRAPH_PROJECT_ROOT (MCP / IDE)
2. Walk up from cwd to find <repo>/.litgraph/config.yaml (skips ~/.litgraph)

Setup & MCP (Cursor)

litgraph setup          # recommended onboarding
# litgraph mcp setup    # alias of the same wizard

Background daemon (optional)

For Zotero auto-sync and to avoid stdio lock conflicts on Windows, run litgraph daemon and choose daemon-http in setup. Details: Tutorial §4 and §7.