HAPI SDK

A plug-and-play SDK for detecting and reducing hallucinations in Large Language Model outputs using semantic entropy analysis and trained classifiers.

Features

Hallucination Detection: Uses trained classifiers to detect hallucinations in LLM outputs
Semantic Entropy Analysis: Advanced semantic analysis to identify uncertain or inconsistent outputs
Easy Integration: Simple API that works with any Hugging Face model
Multiple Detection Methods: Combines classifier-based and semantic entropy-based approaches
Real-time Analysis: Generate and analyze outputs in real-time

Installation

pip install HAPI-SDK

Quick Start

from dehallucinate_sdk import DeHallucinationClient

# Initialize the client
client = DeHallucinationClient(
    model_id="meta-llama/Llama-2-7b-chat-hf",
    license_key="your-license-key"
)

# Generate and analyze output
prompt = "Explain quantum entanglement in simple terms."
output, flagged_sentences = client.generate_output(prompt)

print("Generated Output:", output)
print("Flagged Sentences:", flagged_sentences)

Available Methods

The DeHallucinationClient exposes the following main methods:

1. `generate_output(prompt, max_tokens=512)`

Generates a response and performs hallucination analysis in one step.

output, flagged_sentences = client.generate_output(
    "What is the capital of France?", 
    max_tokens=100
)

2. `semantic_entropy_check(prompt, num_generations=5)`

Analyzes semantic entropy to detect potential hallucinations.

entropy_score, is_hallucinated = client.semantic_entropy_check(
    "Tell me about the history of artificial intelligence"
)

3. `sentence_contains_hallucination(sentence, context="")`

Checks if a specific sentence contains hallucinations.

is_hallucinated = client.sentence_contains_hallucination(
    "The capital of France is Berlin.",
    context="Geography facts"
)

4. `generate(prompt, max_tokens=512)`

Simple text generation without hallucination analysis.

output = client.generate("Write a story about space exploration", max_tokens=200)

5. `generate_sentence(prompt)`

Generates a single sentence response.

sentence = client.generate_sentence("Complete this: The best way to learn programming is")

Supported Models

Llama-2-7b-chat (meta-llama/Llama-2-7b-chat-hf)
Falcon-40b (tiiuae/falcon-40b)
Llama-2-7b (meta-llama/Llama-2-7b-hf)
MPT-7b (mosaicml/mpt-7b)
More models coming soon!

Configuration

Environment Variables

export OPENAI_API_KEY="your_openai_api_key"
export HUGGINGFACE_API_KEY="your_hf_token"

Advanced Usage

# Configure with custom parameters
client = DeHallucinationClient(
    model_id="meta-llama/Llama-2-7b-chat-hf",
    license_key="your-license-key",
    device="cuda",  # or "cpu"
    temperature=0.7,
    use_semantic_entropy=True,
    confidence_threshold=0.8
)

# Batch processing
prompts = [
    "What causes climate change?",
    "How do vaccines work?", 
    "Explain machine learning basics"
]

results = []
for prompt in prompts:
    output, flagged = client.generate_output(prompt)
    results.append({"prompt": prompt, "output": output, "flagged": flagged})

Error Handling

try:
    output, flagged = client.generate_output("Your prompt here")
except Exception as e:
    print(f"Error during generation: {e}")

Contributing

We welcome contributions! Please see our GitHub repository for more information.

License

MIT License - see LICENSE file for details.

Support

For questions, suggestions, or issues, please contact us or open an issue on GitHub.

Metadata

Release files for HAPI-SDK 0.2.23

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for HAPI-SDK 0.2.23
File	Size	Uploaded
hapi_sdk-0.2.23.tar.gz	15.9 MB	Jul 12, 2025	Details

Built distribution (wheel)

Table of built distributions (wheels) for HAPI-SDK 0.2.23
File	Interpreter	ABI	Platform	Reset
hapi_sdk-0.2.23-py3-none-any.whl 16.0 MB Jul 12, 2025	Python 3	none	any	Details

Total release size: 31.9 MB

Release files / hapi_sdk-0.2.23.tar.gz

Download URL	hapi_sdk-0.2.23.tar.gz
Size	15.9 MB
Tags	Source
SHA-256 checksum How to use checksums	`e3e640d7a65b222978152b6bf5ab235e22282b62a64982ce92ab2e34d333f8fb`
BLAKE2b-256 checksum How to use checksums	`21ba70324a18d7c961a152445f62eb1dc8595c62c8fb98a09feb54db5b0a296c`
Upload date	Jul 12, 2025
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/6.1.0 CPython/3.13.5`

Release files / hapi_sdk-0.2.23-py3-none-any.whl

Download URL	hapi_sdk-0.2.23-py3-none-any.whl
Size	16.0 MB
Tags	Python 3
SHA-256 checksum How to use checksums	`c80ffa95f511396ed15e4285c53f437d0e817ff1521567fbc211e8187420d271`
BLAKE2b-256 checksum How to use checksums	`905ed96b33c71fcf92767c549594a3ae4916c08ec08164c711df0a7b4abb7167`
Upload date	Jul 12, 2025
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/6.1.0 CPython/3.13.5`