vexrag 0.2.1

Most RAG security tools focus on jailbreaking or prompt injection. VexRAG is different: it injects poisoned passages directly into the retrieval index and measures the system’s answer functional correctness under adversarial data manipulation.

Threat model — when to use VexRAG

VexRAG is for security testing your own RAG in a controlled, isolated environment (not production).

Use VexRAG if:

• Retrieval data may be untrusted or partially adversarial (uploads, crawls, third-party corpora).
• You need to measure behavior when an attacker can poison or skew the index.
• You are building trust-aware RAG and want evidence of resilience, not only prompt-level guards.

You probably do not need VexRAG if:

• Every indexed document is fully trusted, ingestion is strictly controlled, and you accept that risk without red-team validation.
• You only care about query-time prompt injection or jailbreaks (VexRAG targets retrieval and corpus attacks).

Science-first approach

VexRAG implements paper-backed attacks, not ad-hoc heuristics:

See vexrag/attack_algorithms/ for implementation details and fidelity notes.

Warning — real corpus mutation
Scans write poisoned passages into the target retrieval index. Configs default to cleanup: true, but interrupted runs may still leave poison behind.
Never target production or shared indexes. Back up the retrieval database before testing on your own data.

Versioning Policy

VexRAG is an early-stage library. Until 1.0.0, treat any release as potentially breaking — configs, CLI flags, and APIs may change without a major version bump.

When we deprecate public functionality, it stays available for two minor releases before removal (e.g. deprecated in 0.3.0, removed in 0.5.0).

From 1.0.0 onward we plan to follow SemVer (breaking changes in major releases only).

Prerequisites

Use a sample RAG target from rag_examples: start the example app (each folder’s README has the command), then scan it with vx.

For the default Ollama-based configs you need Python 3.11+ and a running Ollama daemon. CI and releases are tested on 3.11 only; newer versions may work but are not officially supported yet.

python --version  # 3.11+ required; 3.11 tested in CI
ollama list

Install/pull Ollama models for scan configs:

ollama pull llama3:8b

For full benchmarks (ollama-default.yaml and some advanced configs), also pull nomic-embed-text:latest.

All rag_examples targets default to http://localhost:8080; run one example at a time.

Installation

pip install vexrag

Optional extras (install what your stack needs):

pip install "vexrag[qdrant]"
pip install "vexrag[chroma]"
pip install "vexrag[faiss]"
pip install "vexrag[sentence-transformers]"

The sentence-transformers extra enables the in-process embedding provider in scan configs; model weights download from Hugging Face on first use.

Verify the CLI:

vx --help

Run a scan

vx scan --config path/to/scan.yaml

Use sample configs from rag_examples/ as a starting point.

First successful scan

From rag_examples/small/rag_01_in_memory_en:

python3 -m venv .venv
source .venv/bin/activate
pip install vexrag
pip install -r requirements.txt
python3 small_rag.py
vx scan --config scan_configs_examples/ollama-smoke.yaml

Expected outcome:

• python3 small_rag.py serves the target API on http://localhost:8080 (embeddings via Hugging Face; first run may download model weights).
• vx scan finishes with a scan report. Smoke config needs only ollama pull llama3:8b.

Project roadmap

Done

• Implementation of PoisonedRAG (arXiv: 2402.07867)
• Implementation of HijackRAG (arXiv: 2410.22832)
• Automatic generation of attack cases for both methods
• Support for vLLM and Ollama
• Simple RAG examples for quick onboarding to VexRAG
• Support for Qdrant, FAISS, Chroma, and file-based retrieval backends
• Codebase hardening: refactors, typing, tooling, removing AI slop

In Progress

• Stable Python API to run scans and generate cases from code, not only via vx

Ideas / Backlog

• Expand red-team methods in VexRAG
• Expand supported retrieval backends
• Implement a web version of VexRAG

Feedback

Feel free to open a GitHub issue for bugs, questions, or attack methods you would like to see in VexRAG. Pull requests and local development notes are in CONTRIBUTING.md.