agentic-rag-pdf 0.1.3

Agentic RAG over any PDF — grounded, cited answers from research reports, ESG/sustainability disclosures, contracts, manuals, and filings. Multi-agent LangGraph workflow with hybrid retrieval, RAPTOR summaries, NL→SQL over extracted tables, CRAG corrective loops, and a defense-in-depth hallucination stack.

PdfAgent — Agentic RAG over any PDF

Grounded answers from your PDFs, with source citations.

Ask questions about any PDF — research reports, contracts, manuals, ESG / sustainability filings — and get grounded, cited answers. Drop a PDF into the browser UI and start asking.

---

Quick Start

# 1. Install
pip install agentic-rag-pdf

# 2. Install + start Ollama (https://ollama.com/download), pull the embedding model
ollama pull mxbai-embed-large
ollama signin                       # only if using cloud models (e.g. gpt-oss:120b-cloud)

# 3. Launch the UI
pdfagent

Open http://localhost:8000 → click Upload PDF → drop a file → ask questions.

The PyPI distribution name is agentic-rag-pdf (the bare pdfagent name is owned by an unrelated project). The console command is pdfagent.

---

What it does

• Multi-modal ingestion — text, tables (extracted into SQLite), images (OCR + optional vision captions).
• Hybrid retrieval — BM25 + FAISS dense, merged with Reciprocal Rank Fusion, with optional cross-encoder reranking.
• RAPTOR hierarchical summaries for global queries; Map-Reduce for exhaustive aggregation.
• Schema-aware NL → SQL for tabular queries (SELECT-only, sandboxed).
• Hallucination defense (CRAG + 7 layers) — pre-synthesis entity-metric fact extraction, deterministic attribution check, citation verification, claim-level faithfulness, cross-chunk contradiction detection, arithmetic verification of delta tables.
• Three interfaces — browser UI (SSE-streamed agent steps), REST API, CLI.

See docs/ARCHITECTURE.md for the workflow diagram, full feature breakdown, configuration reference, REST API, and project structure.

---

Other ways to install

From the latest GitHub commit (handy between releases):

pip install git+https://github.com/bhagatdas/PdfAgent-Agentic-RAG-over-any-PDF.git

For development (editable, clone the repo):

git clone https://github.com/bhagatdas/PdfAgent-Agentic-RAG-over-any-PDF.git
cd PdfAgent-Agentic-RAG-over-any-PDF
pip install -e .
pdfagent --reload

---

Command-line usage

The wheel installs two console scripts:

pdfagent                              # launch the browser UI (uvicorn on :8000)
pdfagent --host 0.0.0.0 --port 8080   # custom bind

pdfagent-cli ingest                   # CLI ingest (drop files in data/pdfs/ first)
pdfagent-cli query "..."              # one-shot question
pdfagent-cli chat                     # interactive REPL
pdfagent-cli schema                   # print the table schema catalog

Override defaults via a .env file in your working directory. The full settings reference lives in config/settings.py.

---

REST API at a glance

Method	Path	Purpose
POST	/query	Run the pipeline; JSON response
POST	/query/stream	Same, streamed via SSE (one event per agent)
POST	/upload-ingest/stream	Wipe + ingest a single PDF
GET	/health	Vector store doc count + table count
GET	/ollama/health	Probe Ollama readiness + per-OS install help

Full endpoint table and request/response shape: docs/ARCHITECTURE.md#rest-api.

---

Tech stack

LangGraph + LangChain · Ollama (gpt-oss:120b-cloud, mxbai-embed-large) · FAISS + rank-bm25 · cross-encoder reranker · SQLite for tables · FastAPI (SSE) · PyMuPDF · EasyOCR.

---

Publishing

Releases go to PyPI as agentic-rag-pdf on every v* tag push. See PUBLISHING.md for the one-time trusted-publisher setup and the tag → release flow.

---

License

Proprietary — see repository owner.

---

Built by Bhagat Kumar Das.