RAG like a hero with Docling

Your RAG pipeline has security holes you haven't considered. Learn to defend against data poisoning and a new class of vector store attacks.

#1about 3 minutes

Using RAG to enrich LLMs with proprietary data

Retrieval-augmented generation (RAG) is the key to making large language models useful for enterprises by providing them with up-to-date, proprietary information.

#2about 4 minutes

The challenge of parsing complex document structures

Simple document parsers can misinterpret layouts like multi-column text, leading to corrupted data and incorrect outputs from the language model.

#3about 3 minutes

Using Docling to convert documents into structured formats

Docling is an open-source tool that acts like an advanced OCR service, converting various binary document formats into a structured, parsable tree.

#4about 7 minutes

Demo of a basic RAG ingestion pipeline

A live demonstration shows how a Quarkus application uses Docling to ingest a PDF, generate embeddings, and store the resulting chunks and vectors in Redis.

#5about 3 minutes

Securing RAG against data poisoning and leaks

To prevent data poisoning and sensitive data leaks, it is crucial to sanitize documents, verify their signatures, and use tools for PII masking.

#6about 4 minutes

Mitigating vector store attacks and encryption challenges

Vector stores are vulnerable to attacks like close vector modification and reversal, and standard encryption breaks vector distance, requiring specialized solutions.

#7about 5 minutes

Demo of a secure ingestion pipeline in action

A final demonstration showcases a secure pipeline that verifies document signatures, anonymizes sensitive data, and encrypts vectors before storing them.