DocRAG-NVIDIA (MVP)

Multimodal PDF RAG with table-aware retrieval using DuckDB + Chroma.

Why this repo

Most RAG demos ignore tables. This project:

• extracts PDF tables and stores full rows in DuckDB
• stores text chunks + table summaries in Chroma
• performs hybrid retrieval and returns answers with citations (doc/page/table)

Quickstart

cp .env.example .env
docker compose up --build

LLM Runtime Selection (CUDA vs Mac)

The /ask endpoint uses an OpenAI-compatible chat completion API. Switch the backend by setting LLM_MODE.

CUDA (NIM/Triton):

• LLM_MODE=cuda (default)
• NIM_URL (e.g., http://<host>:8000/v1)
• NIM_MODEL
• NIM_KEY (optional)

Mac (local OpenAI-compatible server):

• LLM_MODE=mac
• MAC_LLM_URL (e.g., http://127.0.0.1:8000/v1)
• MAC_LLM_MODEL
• MAC_LLM_KEY (optional)

Ingest Notes

• Text blocks and table summaries are indexed in Chroma.
• Full table rows are stored in DuckDB for exact lookup.
• Image metadata (page, size, bbox) is stored in DuckDB for now.