Native PDF RAG Pipeline

A corrective RAG pipeline that embeds raw PDF bytes using Gemini Embedding 2 — no text extraction needed for embedding. Uses Gemini 2.5 Pro for vision-model OCR (text for search and LLM context), Claude for grading and generation, and LangGraph for the corrective retrieval loop. All steps are traced with LangSmith.

Prerequisites

• Python 3.11+
• Google API key (Gemini Embedding 2 + Gemini 2.5 Pro)
• Anthropic API key
• LangSmith API key

Setup

python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
# Edit .env with your API keys

Environment Variables

Variable	Required	Description
GOOGLE_API_KEY	Yes	Gemini Embedding 2 + Gemini 2.5 Pro OCR
ANTHROPIC_API_KEY	Yes	Claude for grading + generation
LANGSMITH_API_KEY	Yes	LangSmith tracing and evals
LANGSMITH_TRACING	Yes	Set to true to enable

Running

1. Add PDF files to the docs/ directory
2. Run the pipeline:

python pipeline.py

To run evaluations:

python evals.py

Article

Stop Extracting Text from PDFs — Embed the Document Directly