Guardrails

A simple, from-scratch project for building and testing custom LLM guardrails, developed to force model to generate short and concise abstracts with citations.

Features

• Input Guardrails: language detection, injection filtering, and prompt sanitation.
• Output Guardrails: sentence length, citation detection, and text relevence.
• Config-Driven: thresholds (e.g., min sentences, language).
• Extensible Design: add your own checks or plug in domain-specific validators.

Models

• Mistral quantized version: https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.2-GGUF

Data

We used PubMed abstracts for testing the summarization and citation guardrails.

---

Quick Start

# clone the repo
git clone https://github.com/raza4729/guardrails.git
cd guardrails

# create & activate virtual environment
python -m venv venv
source venv/bin/activate   # on Windows: venv\Scripts\activate

# install dependencies
pip install -r requirements.txt

Run the main pipeline:

python -m src.pipeline

---

Structure

src/
 ├── guardrails/
 │    ├── InputGuardrail.py
 │    ├── OutputGuardrail.py
 |    └── guardrails_config.json
 ├── pipeline.py
 └── models.py
tests/
 ├── test_cases.py
 └── test_results.md
requirements.txt
README.md
.gitignore

---

Example Usage

from guardrails.OutputGuardrail import OutputGuardrail
cfg = {"language": "en", "min_sentences": 1, "max_sentences": 3,     "citation_patterns": [
      "\\[\\d+\\]",
      "\\([A-Z][a-z]+ et al\\., \\d{4}\\)",
      "https?://\\S+",
      "\\(Source:.*?\\)"
    ]}
guardrail = OutputGuardrail(cfg)
result = guardrail.check_completeness(input="your prompt here", output="Your model output text")
print(result)

---

Future Work

• Add citation accuracy scoring
• Introduce adaptive confidence thresholds
• Dockerize each service
• Add embedding-based semantic similarity for deeper relevance checks (weighted).
• Introduce rule-based vs. ML-based guardrail hybrid mode.
• Support multilingual guardrail profiles (per language).
• Log violations to a central monitoring service.

---