LitGraph

LitGraph organizes AI literature into 856 research domains, with structured insights, problems, supporting-paper links, and paper cards for research ideation. This release is dated 2026-11-12.

Component	Count
Domains	856
Unique papers	256,227
Domain–paper assignments	256,361
Insights	53,301
Problems	51,858
Papers with structured cards	256,129
Papers with metadata only	98

The insight/problem counts match the frozen export used for the manuscript's library statistics. Domain-level and case-level entries are both included. All available paper cards from that export are included, including 694 cards absent from the older API database's subset of entry-linked papers.

Contents

Two equivalent access formats are supplied:

• snapshot/library.db: complete entries and available structured paper cards, with SQLite FTS5 search data.
• snapshot/api_index.db: field search index, paper metadata and abstracts, domain catalog, and domain–paper mappings for the read-only HTTP API.
• snapshot/domains.json: the complete 856-domain catalog.
• data/*.parquet: typed tables for Hugging Face, Python, SQL engines, and downstream processing. These preserve the original nested records in JSON string columns so heterogeneous extraction fields do not silently disappear.
• api/ and serve.py: read-only API implementation and portable launcher.
• statistics.json, manifest.json, SHA256SUMS, CHANGELOG.md: counts, provenance identifiers, checksums, and differences from the older live API.

The Parquet configs are different tables, not experimental train/test splits; train is a loader label for the full table.

Use the local database and API

Download the complete dataset repository, then run from its root:

pip install huggingface_hub
hf download Einsia/LitGraph --repo-type dataset --local-dir LitGraph
cd LitGraph

python serve.py --verify-only
python serve.py --port 8895

Requires Python 3.10+ with SQLite FTS5. No model key, GPU, vector model, or paid API is needed for retrieval. The launcher verifies the database and index against the release checksums, then creates a local index copy to accommodate changed file timestamps after download. The distributed snapshot files stay intact. Allow approximately 0.8 GB of extra space for that local index copy.

http://127.0.0.1:8895/llms.txt
http://127.0.0.1:8895/openapi.json
http://127.0.0.1:8895/api/v1/stats
http://127.0.0.1:8895/api/v1/domains?limit=100
http://127.0.0.1:8895/api/v1/search?q=distillation&kind=insight&mode=any

The API uses field-specific lexical BM25 with Chinese bigram tokenization. It does not implement vector search. Hugging Face stores the downloadable snapshot; uploading this repository does not itself deploy an HTTP API.

Read Parquet

Load an individual table directly from Hugging Face:

from datasets import load_dataset
insights = load_dataset('Einsia/LitGraph', 'insights', split='train')

Or read the downloaded files locally:

import pyarrow.parquet as pq
insights = pq.read_table('data/insights.parquet').to_pandas()
papers = pq.read_table('data/papers.parquet', columns=[
    'paper_id', 'title', 'abstract', 'source_card_available'
]).to_pandas()

Table	Main columns
domains	id, name, paper_count, domain_json
insights / problems	id, domain_id, kind, tier, statement, supporting_paper_ids, entry_json
papers	paper_id, title, year, venue, arxiv, abstract, source_card_available, domain_ids, metadata_json, card_json
domain_papers	domain_id, paper_id

Parse entry_json, card_json, metadata_json, or domain_json with json.loads to access the complete nested record. Original scientific fields are retained. Pipeline response traces, prompts, usage logs, and internal ingestion-source fields are excluded. Paper years retain their source string representation; no silent correction of known source anomalies is applied. metadata_json also preserves the upstream intuition/problem-definition fields where present, which the current HTTP API does not expose.

Evidence and version scope

This dataset contains metadata, abstracts, and extracted/aggregated research materials. It does not contain a full-text or PDF collection. A card's source=fulltext means its extraction used full text, not that the full text is included. Some cards do not record their extraction depth; see statistics.json. Model-extracted statements can contain errors and should be checked against the linked original publications. Paper IDs are local numeric IDs, not OpenAlex IDs.

This is a current corpus, not a historical-cutoff or benchmark-masked snapshot. The 38-domain IdeaX benchmark data, run trajectories, generated ideas, scores, credentials, and working directories are outside this dataset release.

A dataset-wide reuse license has not yet been specified. This repository does not grant rights to the original publications or assert a blanket license over third-party paper material.