Elsai Model

Package: elsai-model v1.4.1

A unified interface to connect with multiple LLM providers. Every connector exposes the same .invoke() and .stream() API so you can swap providers without changing application code.

Installation

bash

pip install --extra-index-url https://core-packages.elsai.ai/root/elsai-model/ elsai-model==1.4.1

Requirements: Python >= 3.9, .env with API keys

Connectors

1. OpenAIConnector

python

from elsai_model.openai import OpenAIConnector

llm = OpenAIConnector(
    openai_api_key="your_openai_api_key",   # or OPENAI_API_KEY env var
    model_name="gpt-4o-mini",
    temperature=0.1,
    implementation="native"   # or "langchain" (v1.3.0+)
)

messages = [{"role": "user", "content": "Tell me a short joke."}]
response = llm.invoke(messages=messages)

# Streaming (v1.0.1+)
for chunk in llm.stream(messages=messages):
    print(chunk, end='', flush=True)

Environment variables: OPENAI_API_KEY, OPENAI_MODEL_NAME, OPENAI_TEMPERATURE

:::note v1.3.0 added implementation parameter — choose 'native' (default) or 'langchain'. :::

2. AzureOpenAIConnector

python

from elsai_model.azure_openai import AzureOpenAIConnector

llm = AzureOpenAIConnector(
    azure_endpoint="https://your-resource.openai.azure.com/",
    openai_api_key="your-key",
    openai_api_version="2023-05-15",
    deployment_name="gpt-4o-mini",
    temperature=0.1,
    implementation="native"   # or "langchain" (v1.3.0+)
)

response = llm.invoke(messages=[{"role": "user", "content": "Hello!"}])

Environment variables: AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT, OPENAI_API_VERSION, AZURE_OPENAI_DEPLOYMENT_NAME

3. BedrockConnector

Native boto3-based implementation — lower latency, fewer dependencies than the previous LangChain version.

python

from elsai_model.bedrock import BedrockConnector

llm = BedrockConnector(
    aws_access_key="your_key",
    aws_secret_key="your_secret",
    aws_region="us-east-1",
    model_id="us.anthropic.claude-3-7-sonnet-20250219-v1:0",
    max_tokens=500,
    temperature=0.1,
)

result = llm.invoke([{"role": "user", "content": "What is the capital of France?"}])

# Streaming (v1.1.0+)
for chunk in llm.invoke_model_with_response_stream(messages):
    print(chunk, end='', flush=True)

# Image OCR (v1.2.0+)
ocr_output = llm.invoke_with_image(image_path="scan.png", prompt="Transcribe all text.")

Environment variables: AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_REGION, BEDROCK_MODEL_ID

:::note v1.1.1 — Updated to use the Messages API, fixing ValidationException with Claude Sonnet 4 models. v1.2.2 — .invoke() now returns the full response object instead of plain text. :::

4. AnthropicBedrockConnector

Specialized connector for Anthropic Claude models through AWS Bedrock.

python

from elsai_model.anthropic_bedrock import AnthropicBedrockConnector

connector = AnthropicBedrockConnector(
    model_id="your_anthropic_bedrock_model_id",
    max_tokens=500,
    temperature=0.7,
    aws_access_key="your_key",
    aws_secret_key="your_secret",
    aws_region="us-east-1",
)

response = connector.invoke(messages)

for chunk in connector.invoke_with_stream(messages):
    print(chunk, end='', flush=True)

:::note Available from elsai-model v1.1.0. :::

5. GeminiService

python

from elsai_model.gemini import GeminiService

service = GeminiService(api_key="your_gemini_api_key", model="gemini-2.5-flash")

# Single-shot
result = service.generate_text("What is AI?")

# Streaming
for chunk in service.stream_text("Tell me a short story"):
    print(chunk, end="", flush=True)

# Multi-turn chat
chat = service.create_chat()
chat.send_message("I have 2 dogs.")
chat.send_message("How many paws in my house?")

# Streaming chat
for chunk in chat.send_message_stream("Explain in detail"):
    print(chunk, end="", flush=True)

# View chat history
history = chat.get_history()

# Multimodal
response = service.generate_with_image(image_path="photo.png", prompt="Describe this image.")

Environment variables: GEMINI_API_KEY

Logprobs support v1.2.1+

Logprobs expose token-level log probabilities and the top alternative tokens the model considered at each position. Use them for confidence scoring, uncertainty detection, calibration analysis, and debugging model behaviour.

Two parameters control logprobs:

Parameter	Type	Description
`response_logprobs`	`bool`	Enable logprobs in the response
`logprobs`	`int`	Number of top alternative tokens to return per position

Logprobs are supported on generate_text(), stream_text(), and generate_with_image().

Basic example:

python

from elsai_model.gemini import GeminiService

service = GeminiService(api_key="your_gemini_api_key", model="gemini-2.5-flash")

result = service.generate_text(
    "What is Artificial Intelligence?",
    response_logprobs=True,
    logprobs=5,   # return top 5 alternative tokens at each position
)

print(result)

Streaming with logprobs:

python

for chunk in service.stream_text(
    "Explain quantum computing",
    response_logprobs=True,
    logprobs=3,
):
    print(chunk, end="", flush=True)

Multimodal with logprobs:

python

result = service.generate_with_image(
    image_path="chart.png",
    prompt="What trend does this chart show?",
    response_logprobs=True,
    logprobs=5,
)

Confidence scoring example:

python

import math

result = service.generate_text(
    "Is the capital of France Paris?",
    response_logprobs=True,
    logprobs=1,
)

# Convert log probability to linear probability for the first token
# result.candidates[0].logprobs_result contains the token data
candidates = result.candidates[0].logprobs_result.chosen_candidates
first_token_logprob = candidates[0].log_probability
confidence = math.exp(first_token_logprob)
print(f"Model confidence on first token: {confidence:.2%}")

6. LiteLLMConnector

Single consistent interface across all providers.

python

from elsai_model.litellm import LiteLLMConnector

connector = LiteLLMConnector(
    model_name="openai/gpt-4o-mini",   # format: provider/model
    temperature=0.1
)

response = connector.invoke(messages=[{"content": "Hello!", "role": "user"}])

Model name formats: openai/gpt-4o-mini · azure/deployment-name · bedrock/model-id

7. Async connectors

Both OpenAI and Azure OpenAI have async variants for high-throughput workloads.

python

from elsai_model.openai import AsyncOpenAIConnector
from elsai_model.azure_openai import AsyncAzureOpenAIConnector

async def run():
    connector = AsyncOpenAIConnector(model_name="gpt-4o-mini", temperature=0.1)
    response = await connector.invoke([{"role": "user", "content": "Hello!"}])

    async for chunk in connector.stream(messages=[{"role": "user", "content": "Hello!"}]):
        print(chunk, end='', flush=True)

Version history

Version	Changes
1.4.1	Current stable release
1.3.0	`implementation` param for native vs LangChain in OpenAI/Azure connectors
1.2.2	All connectors return full response object instead of plain text
1.2.1	Gemini logprobs support
1.2.0	Bedrock `invoke_with_image` for OCR
1.1.2	Gemini response format change
1.1.1	Bedrock updated to Messages API (Claude Sonnet 4 support)
1.1.0	`AnthropicBedrockConnector` added; Bedrock streaming
1.0.1	Streaming for OpenAI and Azure OpenAI
1.0.0	Gemini and LiteLLM added; native Bedrock boto3 implementation

Elsai Model ​

Installation ​

Connectors ​

1. OpenAIConnector ​

2. AzureOpenAIConnector ​

3. BedrockConnector ​

4. AnthropicBedrockConnector ​

5. GeminiService ​

Logprobs support v1.2.1+ ​

6. LiteLLMConnector ​

7. Async connectors ​

Version history ​

Elsai Model

Installation

Connectors

1. OpenAIConnector

2. AzureOpenAIConnector

3. BedrockConnector

4. AnthropicBedrockConnector

5. GeminiService

Logprobs support v1.2.1+

6. LiteLLMConnector

7. Async connectors

Version history