PromptGuard

Never parse LLM output again.

LLMs are probabilistic. Software is deterministic.

PromptGuard bridges the gap.

PromptGuard is a production-grade reliability layer that turns Large Language Models into safe, structured, testable software components.

If your application depends on LLM outputs — agents, workflows, background jobs, voice systems, document parsing — then PromptGuard prevents the failures that eventually break every LLM app in production.

---

Why PromptGuard Exists

Every LLM developer eventually runs into this:

json.loads(llm_output)  # crashes in prod

Because LLMs:

• return invalid JSON
• hallucinate fields
• change output formats
• break silently after model updates
• fail one out of every N requests

These failures cause:

• background job crashes
• broken agents
• corrupted pipelines
• silent data loss
• 2am production incidents

PromptGuard eliminates this entire class of problems.

---

What PromptGuard Guarantees

• Schema-valid outputs — always
• Automatic repair when models misbehave
• Deterministic structured data
• Prompt regression testing
• Provider-agnostic execution

No regex. No fragile parsing. No silent failures.

---

Why PromptGuard Over Alternatives?

Feature	PromptGuard	Instructor	Outlines
Auto repair loop	Yes (N retries)	1 retry	No
Multi-provider	4 built-in	OpenAI-centric	Multiple
Prompt versioning	Built-in	No	No
Regression testing	Built-in	No	No
Schema types	4 (Pydantic, TypedDict, dataclass, JSON)	Pydantic	Limited
CLI tooling	Yes	No	No

PromptGuard is a reliability layer, not just a parser. Versioning + testing + repair in one package.

---

Installation

pip install llm-promptguard

# With provider-specific dependencies
pip install llm-promptguard[openai]
pip install llm-promptguard[anthropic]
pip install llm-promptguard[google]
pip install llm-promptguard[all]

---

Quick Start

from promptguard import llm_call
from pydantic import BaseModel

class Person(BaseModel):
    name: str
    age: int

result = llm_call(
    model="gpt-4o",
    prompt="John is 30 years old",
    schema=Person
)

print(result.data)
# Person(name='John', age=30)

If the model returns invalid output, PromptGuard automatically:

1. detects the schema violation
2. explains the error
3. re-prompts the model
4. repairs the output
5. returns guaranteed valid data

---

Real-World Examples

Resume & Document Parsing

class Resume(BaseModel):
    name: str
    skills: list[str]
    years_experience: int

resume = llm_call(
    prompt=resume_text,
    model="gpt-4o",
    schema=Resume
)

• No missing fields
• No malformed JSON
• No broken pipelines

---

Email Triage Automation

class EmailIntent(BaseModel):
    intent: str
    urgency: int
    requires_reply: bool

intent = llm_call(
    prompt=email_body,
    model="gpt-4o",
    schema=EmailIntent
)

Safe to run in background workers. Safe to store in databases. Safe to trigger workflows.

---

AI Agents (Tool Calling)

class ToolArgs(BaseModel):
    action: str
    resource_id: str

args = llm_call(
    prompt=agent_prompt,
    model="gpt-4o",
    schema=ToolArgs
)

run_tool(**args.data.model_dump())

PromptGuard prevents agents from:

• hallucinating arguments
• calling tools incorrectly
• freezing execution chains

---

Voice Agents & Call Automation

Speech → LLM → Action

If structure breaks, the call fails.

PromptGuard ensures voice systems always receive valid commands.

---

Background Jobs & Queues

@worker.task
async def process_document(text):
    result = await allm_call(
        prompt=text,
        model="gpt-4o",
        schema=Extraction
    )
    save(result.data)

No retries. No poison messages. No corrupted jobs.

---

LangChain Integration

PromptGuard works seamlessly with LangChain.

from langchain.tools import Tool
from promptguard import llm_call
from pydantic import BaseModel

class SearchArgs(BaseModel):
    query: str

search_tool = Tool(
    name="search",
    func=lambda q: search_api(q),
    description="Search the web"
)

args = llm_call(
    prompt="Search for Tesla earnings",
    model="gpt-4o",
    schema=SearchArgs
)

search_tool.run(args.data.query)

PromptGuard becomes the type-safe boundary between agents and tools.

---

Prompt Regression Testing

promptguard test

• detects prompt drift
• catches model behavior changes
• prevents silent regressions

LLM prompts finally become testable.

---

The Demo

class Order(BaseModel):
    product: str
    quantity: int
    price: float

order = llm_call(
    prompt="Buy two iPhones for $999 each",
    model="gpt-4o",
    schema=Order
)

print(order.data)
# Order(product='iPhone', quantity=2, price=999.0)

No parsing. No retries. No crashes.

Just software-safe AI.

---

Mental Model

Without PromptGuard:

LLM → text → hope → bugs

With PromptGuard:

LLM → contract → software

---

When You Should Use PromptGuard

If your application:

• runs LLMs in production
• executes workflows or agents
• parses model output
• stores results in databases
• depends on structure

Then PromptGuard is not optional.

---

CLI Commands

promptguard init           # Initialize in a project
promptguard run prompt.yaml  # Run a prompt from YAML
promptguard test           # Run regression tests
promptguard history        # Show version history
promptguard diff <hash>    # Compare versions
promptguard stats          # Show statistics

---

Supported Providers

Provider	Model Examples
OpenAI	gpt-4o, gpt-4o-mini, gpt-4-turbo
Anthropic	claude-3-opus, claude-3-sonnet, claude-3-haiku
Google	gemini-1.5-pro, gemini-1.5-flash
Local	Ollama, LM Studio, vLLM (OpenAI-compatible)

---

Testing with real LLM Calls

• export OPENAI_API_KEY=sk-...
• ./scripts/test_all.sh

---

Documentation

Full docs at metanoia-oss.github.io/promptguard.

---

License

Apache 2.0

---

PromptGuard — because production AI must be dependable.