Pawa AI Python SDK

Official Python library for the Pawa AI API.

Pawa AI provides African-built small language models for chat, voice, embeddings, document parsing, agents, and knowledge bases.

Installation

pip install pawa-ai

Quickstart

Set your API key:

export PAWA_AI_API_KEY="your_api_key_here"

Get your key from the Builders Dashboard.

Chat

from pawa_ai import PawaAI

client = PawaAI()

response = client.chat.create(
    model="pawa-v1-ember-20240924",
    messages=[
        {
            "role": "user",
            "content": [{"type": "text", "text": "Hello! How can I use AI in my app?"}],
        }
    ],
    stream=False,
)

print(response.text)  # typed ChatCompletion response
print(response.usage)  # token usage when available

Streaming

with client.chat.create(
    model="pawa-v1-ember-20240924",
    messages=[
        {"role": "user", "content": [{"type": "text", "text": "Explain RAG in simple terms"}]}
    ],
    stream=True,
) as stream:
    for delta in stream.text_deltas():
        print(delta, end="", flush=True)

    # Or collect the full response after streaming
    completion = stream.collect()
    print(completion.text)

Async streaming:

stream = await client.chat.create(..., stream=True)
text = await stream.collect_text()

Text-to-Speech

audio = client.voice.text_to_speech.create(
    model="pawa-tts-v1-20250704",
    text="Hello, this is Pawa AI speaking!",
    voice="liora",
)

with open("output.mp3", "wb") as f:
    f.write(audio)

Embeddings

response = client.vectors.create(
    model="pawa-embeddings-v1-20241001",
    sentences=["Embed this sentence.", "And this one too."],
    lang="multi",
)

embeddings = response.embeddings

Pass raw=True on any resource method to get the original JSON dict instead of typed models.

Retries with exponential backoff

from pawa_ai import PawaAI, RetryConfig

client = PawaAI(
    retry_config=RetryConfig(
        max_retries=3,
        initial_delay=0.5,
        max_delay=8.0,
        exponential_base=2.0,
        jitter=0.1,
    )
)

Retries automatically apply to rate limits (429), server errors (500/502/503/504), and connection failures. The SDK respects Retry-After response headers when present.

Async

import asyncio
from pawa_ai import AsyncPawaAI

async def main():
    async with AsyncPawaAI() as client:
        response = await client.chat.create(
            model="pawa-v1-ember-20240924",
            messages=[
                {"role": "user", "content": [{"type": "text", "text": "Habari yako?"}]}
            ],
        )
        print(response["data"])

asyncio.run(main())

API coverage

Error handling

from pawa_ai import PawaAI, AuthenticationError, RateLimitError, ChatCompletion

client = PawaAI()

try:
    completion: ChatCompletion = client.chat.create(model="pawa-v1-ember-20240924", messages=[...])
except AuthenticationError as e:
    print(f"Auth failed: {e.message}")
except RateLimitError as e:
    print(f"Rate limited: {e.status_code}")

Documentation

• Pawa AI Docs
• API Reference
• Quickstart Guide

License

MIT