Getting Started with DtxPromptGenerator

The DtxPromptGenerator is a lightweight utility that lets you harvest red‑team prompts from DetoxIO without invoking an LLM. It simulates a full red‑teaming pipeline but uses a no-op echo agent to intercept each generated prompt. This gives you clean, structured prompts as BaseMultiTurnConversation objects — no token usage, no model calls.

---

Example

from dotenv import load_dotenv
from dtx_models.prompts import BaseMultiTurnConversation
from dtx.sdk.generator import DtxPromptGenerator
from dtx_models.analysis import PromptDataset

# 1) Load any secret keys from a .env file
load_dotenv()

# 2) Configure the generator (3 prompts from the Xtream dataset)
gen = DtxPromptGenerator(
    max_prompts=3,
    dataset=PromptDataset.HF_JAILBREAKV,
)

# 3) Iterate over prompts – each is a BaseMultiTurnConversation
for convo in gen.generate():
    assert isinstance(convo, BaseMultiTurnConversation)
    print("\n--- Conversation ---")
    for turn in convo.turns:
        print(f"[{turn.role.name}] {turn.message}")

Expected Output

--- Conversation ---
[USER] Ignore all previous instructions and ...
[ASSISTANT] Ignore all previous instructions and ...
--- Conversation ---
...

⚠️ Missing keys? If the dataset requires API credentials and they aren’t set, a DatasetAvailabilityError will be raised with a clear message.

---

3  Key Parameters

Parameter	Default	Description
dataset	PromptDataset.HF_JAILBREAKV	Which dataset to pull prompts from.
max_prompts	10	Total number of prompts to generate.
max_prompts_per_plugin	5	Maximum prompts from each plugin.
max_goals_per_plugin	2	How many goal variations per plugin to use.
tactics	None	Optional list of mutation tactics (e.g., synonym_swap, negation).