bigos

High-level data ingestion library for RAG pipelines.

Status

Early PoC — APIs unstable.

Quickstart

uv sync && uv run pytest

Quick example

uv pip install -e ".[docling]"
bigos parse contract.pdf --format=md

Supported formats

Currently PDF via Docling. More backends planned.

Sanity-checking on your own documents

Drop PDFs in data/, then:

uv run python scripts/sanity_check.py data/ --output-md-dir=data/_parsed/

This parses every PDF, prints per-file stats, and writes markdowns to data/_parsed/. Open them in VS Code to visually verify quality.

Evaluation

Install optional dependencies including eval extras (see pyproject.toml), then run OmniDocBench with optional per-sample dumps:

uv sync --all-extras
uv run bigos eval --benchmark=omnidocbench --subset=tables \
  --max-samples=20 --dump-dir=eval/results/dumps/tables/
# Heavy mathematics PDFs / OmniDocBench academic_paper with Granite-Docling VLM:
# uv run bigos eval --benchmark=omnidocbench --subset=academic_paper --vlm ...
# Compare against older baselines that used text-only GT:
# uv run bigos eval ... --gt-strategy=legacy

Inspect individual samples:

cat eval/results/dumps/tables/SAMPLE_ID.json | jq .

OmniDocBench academic_paper (subset, 20 próbek, json2md GT):

	v3 (standard Docling)	v4 (--vlm)
Mean NED	0.783	0.524
Mean len(pred)/len(gt)	0.26	1.02
Σ pred / Σ GT formuł	0 / 248	98 / 248 (~40%)

Szczegóły: eval/results/comparison-v3-vs-v4-academic.md, baseline baseline-omnidocbench-academic_paper-v4.{md,json}.

Baseline results (current GT = json2md-aligned): eval/results/baseline-*-v3.md (tables + starsze subsety); academic z VLM: v4 powyżej. Starsze baseline (legacy GT): baseline-*-v2.md. See docs/eval/metrics-explained.md, docs/eval/omnidocbench.md, docs/eval/docling-vlm-api.md, and docs/eval/json2md-reference.md.