DSPy Integration

Integrate Sentinel's THSP safety validation into DSPy pipelines.

Installation

pip install dspy sentinelseed

Overview

DSPy is Stanford's framework for programming language models. This integration adds safety validation to DSPy modules, ensuring outputs pass through THSP gates.

Quick Start

import dspy
from sentinelseed.integrations.dspy import SentinelGuard

dspy.configure(lm=dspy.LM("openai/gpt-4o-mini"))

base_module = dspy.ChainOfThought("question -> answer")
safe_module = SentinelGuard(
    base_module,
    api_key="sk-...",
    mode="block"
)

result = safe_module(question="What is machine learning?")
print(result.answer)
print(result.safety_passed)  # True

Components

SentinelGuard

Wraps any DSPy module and validates output.

from sentinelseed.integrations.dspy import SentinelGuard

guard = SentinelGuard(
    module,
    api_key="...",
    provider="openai",  # or "anthropic"
    mode="block",       # block, flag, heuristic
)

SentinelChainOfThought

Validates both reasoning AND output.

from sentinelseed.integrations.dspy import SentinelChainOfThought

cot = SentinelChainOfThought(
    "problem -> solution",
    api_key="...",
    validate_reasoning=True,
    validate_output=True,
)

result = cot(problem="...")
print(result.safety_fields_validated)  # ["reasoning", "solution"]

Tools for ReAct

from sentinelseed.integrations.dspy import create_sentinel_tool

safety_tool = create_sentinel_tool(api_key="...")

agent = dspy.ReAct("task -> result", tools=[safety_tool])

Output Metadata

result.safety_passed    # bool: Did content pass?
result.safety_gates     # dict: Individual gate results
result.safety_reasoning # str: Explanation
result.safety_method    # str: "semantic" or "heuristic"

Validation Modes

Agent Modules

SentinelToolValidator

Validates tool calls before execution.

from sentinelseed.integrations.dspy import SentinelToolValidator

validator = SentinelToolValidator(api_key="...")

@validator.wrap
def search_web(query: str) -> str:
    return web_search(query)

SentinelAgentGuard

Validates each step of agent execution.

from sentinelseed.integrations.dspy import SentinelAgentGuard

safe_agent = SentinelAgentGuard(
    agent,
    api_key="...",
    validate_input=True,
    validate_steps=True,
    validate_output=True,
)

Links

• DSPy Documentation