cmpunkmannu/financebench-voyage-finance-2-embeddings

FinanceBench (voyage-finance-2 embeddings)

Pre-computed embeddings for the FinanceBench corpus. Skips ~$5-15 of Voyage API cost and ~30 minutes of ingest time vs re-embedding from raw PDFs. Intended consumer: the RAG agent at Rishabhmannu/financebench-rag-agent (install: pip install financebench-rag-agent).

What's in the box (frozen)

Field	Value
Source corpus	FinanceBench (SEC filings: 10-K, 10-Q, 8-K, earnings releases)
Parser	pypdf (canonical)
Documents	84 distinct SEC filings
Chunks	68,059
Dense vectors	voyage/voyage-finance-2 (1024-dim, cosine)
Sparse vectors	BM25 tokens (Qdrant fastembed)
File format	Parquet (snappy) + manifest.json
License	CC BY-NC 4.0 (inherited from FinanceBench)

Sector coverage

Sector	Chunks
Consumer Staples	15,714
Consumer Discretionary	10,563
Information Technology	10,221
Industrials	7,095
Financials	7,030
Health Care	6,465
Utilities	5,534
Communication Services	3,568
Materials	1,869

Use it

Prerequisites: install the CLI and bring up the local stack — see the project's docs/setup.md. The CLI command below assumes the api container is already running.

CLI consumer (recommended)

financebench seed --from-hf cmpunkmannu/financebench-voyage-finance-2-embeddings

Direct parquet (any RAG stack — no project setup needed)

from huggingface_hub import hf_hub_download
import pandas as pd

path = hf_hub_download(
    repo_id="cmpunkmannu/financebench-voyage-finance-2-embeddings",
    filename="chunks.parquet",
    repo_type="dataset",
)
df = pd.read_parquet(path)

Schema (chunks.parquet)

Column	Type	Description
chunk_id	string (UUID)	Stable point ID
content	string	Header-annotated chunk text
dense_vector	fixed_size_list[1024]	voyage-finance-2 embedding
sparse_indices	list	BM25 sparse vector indices
sparse_values	list	BM25 sparse vector values (parallel to indices)
doc_type	string	e.g. 10k, 10q, earnings
fb_company	string	FinanceBench canonical company
fb_doc_period	string	Fiscal period (year)
fb_gics_sector	string	GICS sector
financebench_doc_name	string	FB document ID
page_number	int32	PDF page number
chunk_index	int32	Position within document
source_file	string	PDF filename
company, company_name, fiscal_year, confidentiality, num_pages	mixed	Additional payload fields

manifest.json contains pipeline config, the exact dense distance metric, and the snapshot timestamp.

License and attribution

The chunk text is derived from FinanceBench (Islam et al., 2023), released under CC BY-NC 4.0. This dataset inherits that license — non-commercial use only.

Voyage AI's voyage-finance-2 is a commercial API. Embeddings are redistributed under their terms of service. Cite Voyage AI when using.

@article{islam2023financebench,
  title={FinanceBench: A New Benchmark for Financial Question Answering},
  author={Islam, Pranab and Kannappan, Anand and Kiela, Douwe and others},
  journal={arXiv preprint arXiv:2311.11944},
  year={2023}
}

Snapshot provenance (frozen)

• Snapshot generated: 2026-06-01T17:46:12.918581+00:00
• Generated by: financebench-rag-agent v0.3.0
• Source project: https://github.com/Rishabhmannu/financebench-rag-agent
• For current eval results, methodology, and pipeline updates: see the project repo's docs/. This README is frozen at snapshot time; the project may have moved on.