FinLucRAG Corpus (OHLCV + SEC/PR Index)

One-stop corpus metadata for building finance RAG systems.
This repo ships two Parquet tables and a fetcher. It does not re-host filings/press releases.

What’s included

• data/ohlcv.parquet — daily OHLCV for ~30 tickers (2022-01-01 → 2025-09-05).
• data/filings.parquet — document-level index of SEC filings & earnings press releases: doc IDs/keys, statement dates, EDGAR doc URLs (stable, file-specific), checksums, and basic classification.
• scripts/fetch_raw.py — polite downloader that saves originals to raw/{doc_key}.

What’s not included

• No raw HTML/PDF (legal/size reasons). Fetch on demand with the script.
• No paragraph splits (that’s on the user and their workflow).
• No benchmark Q&A (that will live in a separate benchmark repo).

---

File layout

data/ ohlcv.parquet filings.parquet scripts/ fetch_raw.py

Schemas (frozen)

data/ohlcv.parquet

column	type	notes
date	date	trading day (UTC)
ticker	string (UPPER)	e.g., AAPL
open, high, low, close	float64
volume	int64

data/filings.parquet

column	type	notes
doc_id	string	{ticker}_{accession}_{filename}
doc_key	string	{ticker}/{accession}/{source}/{filename}
doc_local_hint	string	raw/{doc_key} target path
ticker	string
cik	string	zero-padded
cik_int	int	numeric CIK
accession	string	with dashes
acc_nodash	string	digits only
form	string	8-K, 10-Q, 10-K, …
source	enum	primary, press_release, exhibits
statement_date	date	doc day (PR/filing)
event_date	date	same as statement_date
next_trading_hint	date	event_date + 1 day (convenience)
event_type	enum	earnings or other (heuristic: 8-K Item 2.02 / EX-99.* PR text)
item_codes	string/NA	e.g., 2.02,9.01 for 8-K
doc_code	string/NA	e.g., EX-99.1
title	string/NA	if known
url	string	doc-level EDGAR URL (ends with filename)
content_type	string	text/html, application/pdf, text/plain
byte_len	int	size when indexed
sha256	string	checksum of local copy at index time
filename	string	file name

---

Quick start

Load with pandas

import pandas as pd
ohlcv = pd.read_parquet("data/ohlcv.parquet")
fil   = pd.read_parquet("data/filings.parquet")
print(len(ohlcv), len(fil))

Fetch originals (on your machine)

python scripts/fetch_raw.py \
  --parquet data/filings.parquet \
  --out raw \
  --only event_type=earnings \
  --sources press_release,primary \
  --rate 2 --verify-sha
# Files land under: raw/{doc_key}

Respectful use: keep the default --rate 2 (2 req/s) and set a real contact email in --user-agent.

Integrity & reproducibility

• Every row has a doc-level URL and filename; many have a sha256 of a known copy.
• You can verify downloads with --verify-sha. If a mismatch occurs, the script will warn.

Licensing & content rights

• This repo: metadata, code, and packaging under the LICENSE below.
• Original documents: owned by their issuers/SEC/third parties. This repo provides links only; you fetch to your machine.

Versioning

• Semantic-ish: bump minor for schema-compatible additions, major for breaking changes.
• filings.parquet and ohlcv.parquet schemas are considered frozen as listed above.

Citation (suggested)

If you use this corpus, cite:

https://huggingface.co/datasets/lakshaychhabra/finlucrag