Prompt Guard

Composable prompt-safety guards as a lightweight Python library. It provides:

• TF‑IDF guard for fast lexical similarity against known phrases.
• RAG guard for embedding-based retrieval against a sentence corpus.
• LLM judge for a final model-based classification.

Quick start

from prompt_guard import GuardPipeline, PromptGuardConfig

cfg = PromptGuardConfig(
    embed_model_name="mxbai-embed-large",
    judge_model_name="qwen2.5:3b-instruct",
)
guard = GuardPipeline(cfg)
result = await guard.check("Some prompt")

Notes

• TF‑IDF and RAG build in‑memory indexes at first use.
• RAG and LLM judge use local Ollama models by default; ensure Ollama is running and the models are pulled.

Text file formats

• phrases.txt: one phrase per line (blank lines are ignored).
• sentences.txt: one sentence per line (blank lines are ignored).