visual-parser 2.0.4

Standalone Visual-RAG PDF Parser - text extraction and Vision-LLM figure descriptions to JSONL

visual-parser (Standalone Visual-RAG PDF Ingestion)

visual-parser is a standalone document-ingestion tool that converts PDFs into a multi-modal JSONL knowledge base (text chunks + figure descriptions + metadata). The intended workflow is:

1. Run visual-parser on curated PDFs to generate JSONL KB files.
2. Run RADIANT-LLM Visual-RAG for QA over the generated KB.

Outputs (JSONL KB)

By default, the pipeline writes:

• 01_chunks_kb.jsonl: chunked text extracted from PDFs (Nougat by default).
• 02_visuals_kb.jsonl: figure/page visual descriptions (Vision LLM).
• 03_metadata_kb.jsonl: document metadata rows (title/author/etc.).
• 04_processed_pdfs.txt: a tracker so re-runs only process new PDFs (unless --rebuild).

API keys (.env)

Provide at least one provider:

• OPENAI_API_KEY (OpenAI)
• GEMINI_API_KEY (Gemini)

Optional:

• HF_TOKEN (if you use gated Hugging Face models)

Run with Docker (Docker Hub)

Prebuilt images are on zev94/radiant-llm under the visual-parser tags:

Tag	Description
visual-parser-latest	Always latest build (rolling)
visual-parser-2.0.0	Pinned release (Apache 2.0 release)
visual-parser-1.0	Legacy — v1.0.0, stale

1) Install Docker

• Docker Desktop (Windows/macOS) or Docker Engine (Linux)

2) Pull the image

docker pull zev94/radiant-llm:visual-parser-latest

3) Run (input + output on the same mounted folder)

Windows PowerShell:

docker run --rm --env-file .env `
  -v "C:\path\to\pdfs:/data" `
  zev94/radiant-llm:visual-parser-latest `
  --input-dir /data --output-dir /data

Linux / WSL:

docker run --rm --env-file .env \
  -v "/path/to/pdfs:/data" \
  zev94/radiant-llm:visual-parser-latest \
  --input-dir /data --output-dir /data

4) Run (separate output directory)

Windows PowerShell:

docker run --rm --env-file .env `
  -v "C:\path\to\pdfs:/data" `
  -v "C:\path\to\out:/out" `
  zev94/radiant-llm:visual-parser-latest `
  --input-dir /data --output-dir /out

Offline install (legacy .tar)

docker load -i .\visual-parser_0.1.0.tar
docker images   # use the tag printed by Docker

Model overrides (optional)

Default vision model is GPT-5.4 when using --vision-provider gpt. Override on the command line:

docker run --rm --env-file .env -v "C:\path\to\pdfs:/data" `
  zev94/radiant-llm:visual-parser-latest `
  --input-dir /data --output-dir /data --vision-model gpt-5.4

Common configuration flags

After pulling the image, run:

docker run --rm zev94/radiant-llm:visual-parser-latest --help

For copy-paste Docker examples (vision presets, text modes, workers, rebuild), see docker-usage-examples.md.

Paths:

• --input-dir / -i (required)
• --output-dir / -o (default: same as input)

Text extraction:

• --text-mode nougat|lightweight (default: nougat)
• --nougat-model facebook/nougat-small
• --chunk-size 500
• --chunk-overlap 100

Vision LLM:

• --vision-provider gpt|gemini (default: gpt)
• --vision-model gpt-5.2 (or gpt-4o, gemini-2.5-flash, etc.)
• --vision-detail low|high|auto
• --reasoning-effort none|low|medium|high|xhigh
• --metadata-pages 2

Performance / misc:

• --max-workers 4
• --rebuild (reprocess everything; ignore 04_processed_pdfs.txt)
• --log-level DEBUG|INFO|WARNING|ERROR

---

Citation

If you use RADIANT-LLM or the accompanying evaluation materials, please cite the journal article:

@article{ndum2026retrieval,
  title={A retrieval-augmented, domain-intelligent agentic framework for reliable decision support in safety-critical nuclear engineering},
  author={Ndum, Zavier Ndum and Tao, Jian and Ford, John and Yim, Mansung and Liu, Yang},
  journal={Reliability Engineering \& System Safety},
  pages={113057},
  year={2026},
  publisher={Elsevier}
}

Journal: Reliability Engineering & System Safety (2026), article 113057
Preprint: https://arxiv.org/abs/2604.22755

---

License

Copyright 2026 Zavier N. Ndum

This project is licensed under the Apache License 2.0. See the LICENSE file in the RADIANT_LLM repository for the full license text.