Metadata-Version: 2.4
Name: validate-llm
Version: 0.1.0
Summary: A composable validation framework for LLM inputs and outputs
License: MIT License
        
        Copyright (c) 2026 Thomas Yeoh
        
        Permission is hereby granted, free of charge, to any person obtaining a copy
        of this software and associated documentation files (the "Software"), to deal
        in the Software without restriction, including without limitation the rights
        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
        copies of the Software, and to permit persons to whom the Software is
        furnished to do so, subject to the following conditions:
        
        The above copyright notice and this permission notice shall be included in all
        copies or substantial portions of the Software.
        
        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
        SOFTWARE.
        
Requires-Python: >=3.11
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: litellm
Requires-Dist: deepeval>=1.0.0
Requires-Dist: anthropic
Requires-Dist: better-profanity
Requires-Dist: detoxify
Requires-Dist: sentence-transformers
Requires-Dist: torch
Requires-Dist: duckduckgo-search>=6.2.0
Requires-Dist: ddgs
Requires-Dist: rank-bm25>=0.2.2
Requires-Dist: pydantic
Provides-Extra: demo
Requires-Dist: fastapi; extra == "demo"
Requires-Dist: uvicorn; extra == "demo"
Provides-Extra: test
Requires-Dist: pytest; extra == "test"
Requires-Dist: datasets>=2.0.0; extra == "test"
Requires-Dist: scikit-learn; extra == "test"
Provides-Extra: dev
Requires-Dist: validate-llm[demo,test]; extra == "dev"
Dynamic: license-file

<picture>
  <source media="(prefers-color-scheme: dark)" srcset="brand/pip-install-dark.svg">
  <img alt="pip install validate-llm" src="brand/pip-install.svg" width="560">
</picture>

<br>

Composable validation guardrails for LLM pipelines — accuracy, relevancy, toxicity, privacy, and bias checks in one pipeline.

**[Documentation →](https://tyeoh9.github.io/llm-validation-framework)**

---

## Install

```bash
pip install validate-llm
```

Requires Python 3.11+. Optional extras:

```bash
pip install "validate-llm[demo]"   # FastAPI demo server + web UI
pip install "validate-llm[test]"   # pytest + datasets
pip install "validate-llm[dev]"    # everything
```

## Quick start

```python
from llm_validation_framework import ValidationFramework, LLMProvider, Pipe
from llm_validation_framework import ToxicityAgent, PrivacyAgent, AccuracyAgent
from llm_validation_framework.config_loader import load_api_key

llm = LLMProvider(provider="anthropic", model="claude-haiku-4-5-20251001", key=load_api_key())

vf = ValidationFramework(
    llm=llm,
    input_guardrail=Pipe(steps=[ToxicityAgent()], verbose=False),
    output_guardrail=Pipe(steps=[ToxicityAgent(), PrivacyAgent(), AccuracyAgent()], verbose=False),
)

result = vf.validate("What is the Pacific Ocean?")
print(result["status"], result["score"])  # PASS 0.87
```

`validate()` returns a structured dict with `status`, `score`, and per-agent `results` for both the input and output guardrails. See the [docs](https://tyeoh9.github.io/llm-validation-framework/core/validation-framework/) for the full schema.

## Agents

| Agent | What it does | Needs API key |
|---|---|---|
| `ToxicityAgent` | Three-layer check: profanity filter → toxicity model → semantic similarity | No |
| `PrivacyAgent` | Regex scan for SSN, credit cards, API keys; optional system prompt leakage detection | No |
| `AccuracyAgent` | LLM-as-a-judge factual accuracy + relevancy, with optional RAG grounding | Yes |
| `RelevancyAgent` | LLM-as-a-judge check that the answer addresses the question | Yes |
| `BiasAgent` | LLM-as-a-judge scan for stereotypes and discriminatory language | Yes |

`ToxicityAgent` and `PrivacyAgent` run fully locally with no external calls.

## Config

```bash
export ANTHROPIC_API_KEY=your-key
```

Or create a `config.ini` at the repo root (gitignored):

```ini
[ANTHROPIC]
API_KEY=your-key
```

Supported providers follow [litellm's naming](https://docs.litellm.ai/docs/providers).

## RAG grounding

Pass a retriever to `AccuracyAgent` to ground factual checks against your own corpus:

```python
from llm_validation_framework import AccuracyAgent, RAGProvider

accuracy = AccuracyAgent(rag=RAGProvider(your_vectorstore.as_retriever()))
```

See the [RAG Integration guide](https://tyeoh9.github.io/llm-validation-framework/guides/rag-integration/) for a full walkthrough.

## Demo

The demo is a FastAPI backend + static web UI.

```bash
# Terminal 1
uvicorn demo.api_server:app --host 127.0.0.1 --port 5050

# Terminal 2
python demo/serve_ui.py
```

Open `http://127.0.0.1:8000`.

## Contributors

- Hitha Shri Nagaruru
- James Wu
- Lewis Lui
- Thomas Yeoh

## License

MIT — see [LICENSE](LICENSE)
