-
OpenAI GPT-4o
-
Azure AI GPT-4o
-
Anthropic Claude Sonnet 3.7 via Amazon Bedrock Converse
Validated AI Providers and LLM Models
Last updated 12 minutes ago
Learn which AI providers and LLM models are integrated and tested for use with KIO Co-Pilot.
Table of Contents
CoreMedia implements and tests KIO Co-Pilot with several external AI providers and their Large Language Models (LLMs) following our customers' demands. This guide provides an overview of the AI providers and LLM models with which CoreMedia integrates out-of-the-box and which are used for internal testing and validation with KIO Co-Pilot. It also describes how to configure different providers in KIO Co-Pilot.
In addition, the guide outlines limitations, validation status, and reasons why not every available model is immediately suitable for integration.
Integrated AI Providers
KIO Co-Pilot integrates with the following external AI/LLM providers, selected due to customers demand:
-
AWS Bedrock
-
Microsoft Azure AI
-
OpenAI
These providers offer a broad set of foundation models. CoreMedia uses a subset of these models to ensure stable, predictable behavior in production environments.
|
Because prompt engineering, response quality, and model behavior differ across vendors—and sometimes even between models from the same vendor, CoreMedia cannot ensure uniform performance across all LLM providers. We have not conducted long-term testing with every LLM vendor in the market. |
Validated LLM Models
While the selected providers offer many different models, CoreMedia performs continuous evaluation and validation on a selected subset to ensure they behave reliably with KIO Co-Pilot.
A validated model is a model that CoreMedia continuously tests with KIO Co-Pilot. KIO Co-Pilot will run the most stable when using one of these models.
|
CoreMedia integrates with external third-party AI providers but does not create, control, or fine-tune any of the models they offer. Model behavior may vary due to the probabilistic nature of LLMs, and identical prompts may not always produce identical responses. Accuracy, consistency, or response quality cannot be guaranteed for any model by CoreMedia. |
The following LLM models are currently used and tested for production use:
| KIO Backend Version | Validated LLM Models |
|---|---|
< 1.4.0 |
|
>= 1.4.0 |
|
>= 2.0.0 |
|
|
For KIO Backend Version >= 2.0.0, OpenAI/Azure AI GPT-5.6-Terra is highly recommended over Anthropic Claude Sonnet 4.6, as it delivers significantly faster response times. Claude Sonnet 4.6 remains a validated option but is comparatively slower. |
|
The list of validated models may expand over time as further models are evaluated. While other models might also work, we recommend using the validated models listed above. |
Other Providers and Models
It may be possible to connect additional AI providers or use other models offered by the selected providers. However, configurations outside the validated list should be considered as experimental.
Possible limitations include, for example:
-
reduced or inconsistent tool/function-calling support
-
differences in streaming or response formatting
-
unpredictable conversational behavior
-
stricter rate or throughput limits
Models outside the validated list may work, but are not recommended for production use without customer-side testing.
You might contact CoreMedia for assistance in evaluating additional models.
LLM Provider Configuration
Each LLM provider requires specific configuration properties for the KIO Co-Pilot backend. The Blueprint workspace provides a configuration example for OpenAI as the default configuration. You can override this configuration by activating a specific Spring profile for your LLM provider and adapting the necessary properties from the examples below.
OpenAI Configuration
# OpenAI Example Configuration
# Configure your OpenAI API Key
# You can get an API Key from https://platform.openai.com/account/api-keys
# Permissions: Restricted and List models: Read, Model capabilities: Request
spring.ai.openai.api-key=CONFIGURE_ME
spring.ai.model.chat=openai
spring.ai.model.image=openai
spring.ai.model.embedding=openai
spring.ai.openai.chat.model=gpt-5.6-terra
spring.ai.openai.chat.reasoning-effort=none
# tool search embeds tool descriptions/queries; text-embedding-3-large gives the best retrieval quality
# spring.ai.openai.embedding.model=text-embedding-3-large
# similarity threshold tuned for text-embedding-3-large (0.0-1.0); raise for stricter matching, lower for more recall
# kio.backend.agents.tool-search.similarity-threshold=0.32
# to extract image data we can use a faster model with less complexity
kio.backend.image-extraction-model=gpt-5.4-nanoThe example configuration is also available as the Spring profile openai and can be activated by setting the environment variable
KIO_SPRING_PROFILES_ACTIVE=dev,openai in the deployment configuration.
Azure OpenAI Configuration
# Azure OpenAI Example Configuration
# For more details see https://docs.spring.io/spring-ai/reference/2.0/api/chat/azure-openai-chat.html
spring.ai.openai.api-key=CONFIGURE_ME
spring.ai.model.chat=openai
spring.ai.model.image=openai
spring.ai.openai.base-url=https://my-endpoint.cognitiveservices.azure.com
spring.ai.openai.model=gpt-5.6-terra
spring.ai.openai.microsoft-deployment-name=my-gpt-5.6-terra-deployment-name
kio.backend.image-extraction-model=${spring.ai.openai.model}
spring.ai.model.embedding=openai
# tool search embedding: on Azure the model is chosen by deployment — point this at a text-embedding-3-large deployment
# spring.ai.openai.embedding.microsoft-deployment-name=my-text-embedding-3-large-deployment-name
# similarity threshold tuned for text-embedding-3-large (0.0-1.0); raise for stricter matching, lower for more recall
# kio.backend.agents.tool-search.similarity-threshold=0.32The example configuration is also available as the Spring profile azure and can be activated by setting the environment variable
KIO_SPRING_PROFILES_ACTIVE=dev,azure in the deployment configuration.
Anthropic Claude via Amazon Bedrock Configuration
# Bedrock Converse Example Configuration
spring.ai.model.chat=bedrock-converse
spring.ai.model.image=bedrock-converse
spring.ai.model.embedding=bedrock-converse
spring.ai.bedrock.converse.chat.model=us.anthropic.claude-sonnet-4-6
kio.backend.image-extraction-model=${spring.ai.bedrock.converse.chat.model}
spring.ai.bedrock.aws.region=us-east-1
# the default is 500 tokens, which is not enough and caused JSON parse errors when updating larger rich text properties
spring.ai.bedrock.converse.chat.max-tokens=4000
# the default is 30s, which turns out to be too short for long-running operations
spring.ai.bedrock.aws.asyncReadTimeout=120s
# Enable prompt caching for system prompt and tool definitions.
# Anthropic Claude requires explicit cache breakpoints (unlike OpenAI's implicit prefix caching).
# SYSTEM_AND_TOOLS places breakpoints on the system prompt and tools description,
# making the cached prefix identical across all users and conversations. This maximizes
# cache hits in multi-user scenarios where many concurrent sessions share the same
# system prompt and tools. Cache TTL is 5 minutes, reset on each hit. Requires 1024+
# tokens to activate.
# See https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html
# Alternative: CONVERSATION_HISTORY caches tools + system prompt + all previous turns,
# which yields deeper per-session caching but does not share across users.
spring.ai.bedrock.converse.chat.cache-options.strategy=SYSTEM_AND_TOOLSThe example configuration is also available as the Spring profile converse and can be activated by setting the environment variable
KIO_SPRING_PROFILES_ACTIVE=dev,converse in the deployment configuration.
How to Continue
To learn more about KIO Co-Pilot, explore the other guides: KIO Guides Overview
Need support? Feel free to contact us at support@coremedia.com.
If you encounter issues or unexpected behavior, our Support Team will be glad to assist.
Is this page useful?