vitaai 1.0.1

Python SDK for VitaAI self-hosted LLM endpoints — drop-in Gemini API replacement

vitaai · Python SDK

Python SDK for VitaAI self-hosted LLM endpoints.
Works just like the Google Gemini SDK — client.models.method() pattern with a stream parameter on every endpoint.

---

Install

pip install vitaai
pip install requests  # required for image endpoints

---

Quick start

from vitaai import VitaAI

client = VitaAI(
    base_url="http://your-server-ip:8000",
    api_key="your-api-key"   # optional — not required if server has no auth
)

response = client.models.generate_content(
    model="gemma-4-E2B-it",
    contents="What is AI?",
    max_output_tokens=100
)
print(response.text)
print("Tokens:", response.usage_metadata.total_token_count)

---

Client parameters

Parameter	Type	Default	Description
base_url	str	https://api.vitaai-api.com	Your self-hosted server URL
api_key	str	None	Optional API key. Or set VITAAI_API_KEY env var
timeout	int	120	Request timeout in seconds

---

Streaming

Every endpoint supports a stream parameter:

• stream=False (default) — waits and returns the full response object
• stream=True — yields tokens live as the model generates them

# Normal — full response at once
r = client.models.generate_content(model="gemma-4-E2B-it", contents="What is AI?")
print(r.text)

# Streaming — tokens print live
for chunk in client.models.generate_content(model="gemma-4-E2B-it", contents="What is AI?", stream=True):
    print(chunk.text, end="", flush=True)

---

Endpoints

1. Generate content

r = client.models.generate_content(
    model="gemma-4-E2B-it",
    contents="What is machine learning?",
    max_output_tokens=150,
    stream=False,  # default
)
print(r.text)
print("Tokens:", r.usage_metadata.total_token_count)

With config (system instruction, temperature, etc.)

r = client.models.generate_content(
    model="gemma-4-E2B-it",
    contents="Explain deep learning",
    config={
        "system_instruction": "You are a teacher. Explain simply.",
        "temperature": 0.7,
        "top_p": 0.9,
        "max_output_tokens": 200,
    }
)
print(r.text)

Multi-turn conversation

r = client.models.generate_content(
    model="gemma-4-E2B-it",
    contents=[
        {"role": "user",  "parts": [{"text": "My name is Alex."}]},
        {"role": "model", "parts": [{"text": "Nice to meet you, Alex!"}]},
        {"role": "user",  "parts": [{"text": "What is my name?"}]},
    ]
)
print(r.text)

Streaming

for chunk in client.models.generate_content(
    model="gemma-4-E2B-it",
    contents="Write a short poem.",
    max_output_tokens=100,
    stream=True,
):
    print(chunk.text, end="", flush=True)

Response field	Type	Description
r.text	str	Full generated text
r.usage_metadata.prompt_token_count	int	Input tokens
r.usage_metadata.candidates_token_count	int	Output tokens
r.usage_metadata.total_token_count	int	Total tokens

---

2. Summarize

# Normal
r = client.models.summarize(
    contents="Long article text here...",
    format="bullets",       # "paragraph" or "bullets"
    max_output_tokens=150,
)
print(r.summary)

# Streaming
for chunk in client.models.summarize(contents="Long article...", format="paragraph", stream=True):
    print(chunk.text, end="", flush=True)

Parameter	Type	Default	Description
contents	str	—	Text to summarize
format	str	"paragraph"	"paragraph" or "bullets"
max_output_tokens	int	200	Max summary length
stream	bool	False	Stream tokens live

Response field	Type	Description
r.summary	str	Summarized text
r.format	str	Format used
r.usage_metadata.total_token_count	int	Total tokens

---

3. Embeddings

r = client.models.embed_content(
    contents="Artificial intelligence is transforming the world."
)
print("Dimensions:", len(r.embedding.values))
print("Vector:", r.embedding.values[:5])

Response field	Type	Description
r.embedding.values	List[float]	The embedding vector

---

4. Analyze image

# Normal
r = client.models.analyze_image(
    file_path="photo.jpg",
    prompt="What objects are visible in this image?",
    max_output_tokens=512,
)
print(r.analysis)

# Streaming
for chunk in client.models.analyze_image(file_path="photo.jpg", prompt="What do you see?", stream=True):
    print(chunk.text, end="", flush=True)

Parameter	Type	Default	Description
file_path	str	—	Path to image (JPG, PNG, WEBP, GIF)
prompt	str	—	Question or instruction
system_instruction	str	None	Optional system prompt
temperature	float	1.0	Sampling temperature
max_output_tokens	int	512	Max response length
image_token_budget	int	280	Detail level: 70/140/280/560/1120
stream	bool	False	Stream tokens live

Response field	Type	Description
r.analysis	str	Model's analysis
r.image_info.size	tuple	Image dimensions
r.usage_metadata.total_token_count	int	Total tokens

---

5. Describe image

# Normal
r = client.models.describe_image(file_path="photo.jpg", max_output_tokens=512)
print(r.description)

# Streaming
for chunk in client.models.describe_image(file_path="photo.jpg", stream=True):
    print(chunk.text, end="", flush=True)

Response field	Type	Description
r.description	str	Full image description
r.image_info.size	tuple	Image dimensions
r.usage_metadata.total_token_count	int	Total tokens

---

6. OCR — extract text from image

# Normal
r = client.models.ocr_image(file_path="document.png", language="English")
print(r.extracted_text)

# Streaming
for chunk in client.models.ocr_image(file_path="document.png", stream=True):
    print(chunk.text, end="", flush=True)

Parameter	Type	Default	Description
file_path	str	—	Path to image
language	str	"English"	Language of text in image
max_output_tokens	int	512	Max response length
image_token_budget	int	560	Higher = better OCR accuracy
stream	bool	False	Stream tokens live

Response field	Type	Description
r.extracted_text	str	All text found in image
r.image_info.language	str	Language used
r.usage_metadata.total_token_count	int	Total tokens

---

7. Ping

r = client.models.ping()
print(r.status)   # "ok"
print(r.model)    # "gemma-4-E2B-it"
print(r.device)   # "CUDA"
print(r.version)  # "2.3.0"
print(r.limits.max_input_tokens)
print(r.limits.max_output_tokens)

---

8. Health

r = client.models.health()
print(r.status)          # "healthy"
print(r.cuda_available)  # True / False
print(r.limits.max_input_tokens)
print(r.limits.max_output_tokens)

---

9. Stats

s = client.models.stats()
print(s.device)                    # "CUDA"
print(s.gpu_name)                  # "Tesla T4"
print(f"{s.gpu_memory_total_gb:.1f} GB")
print(f"{s.gpu_memory_allocated_gb:.1f} GB")

---

10. List models

for m in client.models.list():
    print(m["name"])
    print(m["capabilities"])
    print(m["max_output_tokens"])

---

Error handling

from vitaai.transport import AuthenticationError, ServerError, VitaAIError

try:
    r = client.models.generate_content(model="gemma-4-E2B-it", contents="Hello")
    print(r.text)
except AuthenticationError:
    print("Invalid API key")
except ServerError as e:
    print(f"Server error ({e.status_code}): {e}")
except VitaAIError as e:
    print(f"Error: {e}")

---

Server endpoints

Method	Path	stream support
GET	/	—
GET	/health	—
GET	/v1/stats	—
GET	/v1/models/list	—
POST	/v1/models/generate	✅ ?stream=true
POST	/v1/models/summarize	✅ ?stream=true
POST	/v1/models/embed	—
POST	/v1/models/analyze-image	✅ ?stream=true
POST	/v1/models/describe-image	✅ ?stream=true
POST	/v1/models/ocr-image	✅ ?stream=true

---

Environment variables

Variable	Description
VITAAI_API_KEY	API key (alternative to api_key= parameter)
VITAAI_BASE_URL	Server URL (alternative to base_url= parameter)