simplex-ai-inc/LiteResearcher-Browse-Corpus

LiteResearcher Browse Corpus

Full web-page text for reproducing the visit / /web_parser tool of LiteResearcher.

---

What this is

The LiteResearcher local environment exposes two tools, backed by two separate datasets:

Tool	Endpoint	Dataset	Content
search	/search	LiteResearcher-Corpus	url / title / doc — doc is snippet-level (~200 chars), used to build the BGE-M3 vector index
visit	/web_parser	this dataset	url / title / text — text is the full page body as markdown

Until now only the search side was released, so an agent that found a URL could never read past the snippet. This dataset completes the loop, making the full think → search → visit → answer trajectory reproducible locally.

Format

Each shard is zstd-compressed JSON Lines, one record per line:

{
  "url":   "https://arxiv.org/html/2405.05538v1",
  "title": "A Survey on Personalized Content Synthesis with Diffusion Models",
  "text":  "# A Survey on Personalized Content Synthesis...\n\n(full page body, markdown)"
}

Bodies were produced from the originally crawled HTML with BeautifulSoup: script, style, iframe, svg and noscript are stripped; headings, lists, tables and links are preserved.

Statistics

Pages	14,390,949
Compressed size	936.9 GB (zstd level 10)
Shards	104 × data/part_*.jsonl.zst
Body length	mean ≈ 16 KB, median ≈ 3 KB, long tail beyond 500 KB
title coverage	~100% (back-filled by URL from the search corpus)
URLs shared with the search corpus	13,441,577 of 25,986,968 (51.7%)

Body length varies a lot across crawl batches — some batches are dominated by short/listing pages (median ~2 KB), others by long-form articles (50 KB+ per page). Statistics are worth recomputing yourself if your experiment is length-sensitive.

/web_parser serves any URL present in this corpus and reports not found otherwise; the rest of the retrieval service is unaffected either way. Coverage will grow in future revisions.

Composition

About half of the records are Wikipedia articles, the rest are crawled web pages:

Source	Share
*.wikipedia.org	7,173,828 (49.8%)
Other crawled web pages	7,217,121 (50.2%)

Wikipedia records can be separated with a simple URL prefix filter if you only want the open web portion (or if you would rather use an official Wikipedia dump).

---

Usage

1. Download

huggingface-cli download simplex-ai-inc/LiteResearcher-Browse-Corpus \
    --repo-type dataset --local-dir ./browse_corpus

The full dataset is ~937 GB. For a smaller trial, fetch a few shards only:

huggingface-cli download simplex-ai-inc/LiteResearcher-Browse-Corpus \
    --repo-type dataset --local-dir ./browse_corpus \
    --include "data/part_000*.jsonl.zst"

2. Start your own PostgreSQL

Any PostgreSQL 14 or newer works — this dataset is plain JSONL, not a version-locked database dump. Docker is the quickest route:

mkdir -p /your/big/disk/pgdata

docker run -d --name lr-browse-db \
    -p 5432:5432 \
    -v /your/big/disk/pgdata:/var/lib/postgresql/data \
    -e POSTGRES_PASSWORD=yourpassword \
    --shm-size=1g \
    postgres:17 \
    -c shared_buffers=4GB \
    -c work_mem=64MB \
    -c maintenance_work_mem=2GB \
    -c max_wal_size=8GB \
    -c checkpoint_timeout=30min

Requirements before you start:

Disk	1.2–1.5 TB for the full dataset (PostgreSQL TOAST compresses the bodies)
RAM	8 GB is enough; more only speeds up the load
Load time	~10 hours for the full dataset, dominated by disk write throughput

The tuning flags above matter: max_wal_size and checkpoint_timeout keep PostgreSQL from checkpointing every few seconds during a bulk load, and maintenance_work_mem speeds up the index build at the end. None of them are required for correctness.

Verify it is up:

docker exec lr-browse-db pg_isready -U postgres

3. Load the shards

pip install psycopg2-binary zstandard

python load_browse_db.py ./browse_corpus \
    --host localhost --port 5432 \
    --user postgres --password yourpassword

load_browse_db.py is in this repository. It:

• streams each .zst shard straight into COPY — never decompresses to disk, never buffers a whole shard in memory
• creates the unique index on url only after all data is loaded (building it incrementally is several times slower)
• records a checkpoint per shard, so interrupting and re-running is safe — completed shards are skipped

Do not load this with row-by-row INSERT/UPSERT; that is tens of times slower and will take days rather than hours.

The table is created as litesearch_sql.documents, which is what the retrieval service reads by default.

4. Enable the browse endpoint

In environment/server/diskann_config.py of the LiteResearcher repository:

ENABLE_SQL_FULLTEXT = True
SQL_HOST     = "localhost"
SQL_PORT     = 5432
SQL_DATABASE = "postgres"
SQL_USER     = "postgres"
SQL_PASSWORD = "yourpassword"
SQL_SCHEMA   = "litesearch_sql"
SQL_TABLE    = "documents"

Restart the retrieval service and check:

curl -X POST http://localhost:8018/web_parser \
     -H "Content-Type: application/json" \
     -d '{"url": "https://arxiv.org/html/2405.05538v1"}'
# → {"found": true, "url": "...", "title": "...", "text": "full page body…"}

"found": true means browse is live. If SQL is unreachable, /web_parser returns 503 and every other endpoint keeps working, so a misconfiguration will not take the service down.

Not using PostgreSQL?

Lookups are exact-match on url and use no PostgreSQL-specific features. Loading the shards into SQLite, DuckDB or RocksDB works equally well — point sql_fulltext.py at your store of choice.

---

Provenance and licensing

Bodies come from publicly reachable web pages, crawled in accordance with each site's robots.txt. The dataset itself is released under Apache-2.0, but the underlying page content remains the property of its respective owners and is provided for research purposes only.

Approximately 50% of the records are Wikipedia articles, whose original content is licensed CC BY-SA 4.0; attribute accordingly if you use that portion separately.

If you own the rights to a page included here and want it removed, please open an issue on the GitHub repository and we will drop it in the next revision.

Citation

@article{literesearcher2026,
  title={LiteResearcher: A Scalable Agentic RL Training Framework for Deep Research Agent},
  author={Li, Wanli and others},
  journal={arXiv preprint arXiv:2604.17931},
  year={2026}
}