...

An AI analytics question looks like a single request. However, inside the agent, it becomes a sequence of tasks. Ask an analytics agent why gross margin fell in the Northeast last quarter, and it plans the analysis, generates SQL, interprets the results and summarizes them in plain language. Across a longer conversation, it also condenses earlier turns to keep the context window manageable. These tasks do not have identical requirements. Only some of them call for a frontier model’s reasoning. Others are more focused and can be handled by models optimized for that specific job.

That is the premise behind model routing in Kyvos: using the best-fit, most cost-efficient model for each task.

Why do different AI tasks need different models?

AI tasks differ in how much reasoning they require, so a single model priced for the hardest task overpays for the rest.

Consider two ends of the spectrum. Planning a multi-step analysis and generating SQL against enterprise data require reasoning and precision. An error at either task can produce a wrong answer that looks right.

Other tasks are routine. For example, compaction runs throughout a conversation, shrinking earlier turns so the context window stays within limits. Knowledge base summarization distills documents and reference material into what the agent needs for the current question. Both are high-frequency, high-volume work that smaller, faster models handle well.

Since these tasks run so often, they account for a meaningful share of total token volume. Tokens are the units of text that LLM providers bill for. Lightweight models process tasks at a fraction of the per-token price of frontier models, and at enterprise scale, that gap compounds across thousands of users and millions of questions.

Enterprise AI cost comes down to token volume multiplied by token price. Kyvos addresses both factors: model routing manages price, and governed business context reduces volume.

How Kyvos enables model routing

Kyvos lets users independently configure the LLM and embedding connections for each AI-driven capability, optimizing performance, cost and response quality. A capability is a distinct type of work an AI agent performs, such as planning or query generation. A typical model routing configuration in Kyvos looks like this:

Capability Typical model tier Role
Planning Reasoning-tier Sequences tool calls and interprets results
Query generation Frontier Converts natural language questions into structured queries
Code generation Frontier Writes and runs Python for forecasting and statistical analysis
Summary generation Summary generation General-purpose, or lightweight in Quick mode Explains query results in plain language Explains query results in plain language
Web and URL search Search-enabled Retrieves and synthesizes external information
Metadata generation General-purpose Drafts descriptions, tags and business context for semantic models
Knowledge base summarization Lightweight Distills retrieved policies, playbooks and documentation
Compaction Lightweight Compaction Lightweight Condenses conversation history to keep long sessions efficient

Enterprises can also set two guardrails on spend. In Deep Dive mode, three effort levels (Low, Medium and High) cap how many tool calls the planner can make, so simple questions get short plans and only complex analysis takes extra steps. A ceiling on summary generation limits how much result data reaches the model and keeps large result sets from inflating the token bill.

With these settings, enterprises can reserve frontier models for capabilities where a mistake is expensive and run every other capability on the most cost-efficient reliable model.

What does model routing look like in Kyvos?

What does AI context look like in practice

Which LLM providers and models does Kyvos support?

Kyvos supports certified LLM and embedding models from Azure OpenAI, Google Vertex AI and Amazon Bedrock, spanning the OpenAI GPT, Google Gemini and Anthropic Claude model families. Certified models provide reliable, secure and optimized performance for conversational analytics and semantic search.

Kyvos certifies each model for compatibility with specific AI capabilities, so administrators know which models suit which work. Enterprises keep the cloud provider and LLM agreements already in place. Each capability takes its own connection, so a single deployment can mix models across tiers and providers.

How does Kyvos reduce the number of tokens AI needs?

Model routing addresses one part of AI economics: the cost of the model performing each task. However, the number of tokens an AI system needs to process matters just as much. Kyvos semantic layer addresses that side.

Kyvos unifies semantics, entities, knowledge and relationships into complete business context for AI, with consistent metric definitions and governed business logic enforced at query time. The agent resolves what a question means directly from that context, which cuts exploratory queries and extra reasoning steps. The result is fewer tokens per question and less compute on the underlying data platform.

Governed context and model routing work together in Kyvos. With model routing configured, each question uses fewer tokens, and each token is priced at the tier its capability requires.

Frequently Asked Questions

Does model routing apply to every AI agent on Kyvos?
Yes. Any AI agent using Kyvos follows the same per-capability model routing, based on the connections users configure.
Can administrators control whether data is sent to the LLM in Kyvos?
Yes. Kyvos includes an Allow Sending Data to LLM setting that administrators use to permit or prevent data from being sent to the configured LLM during query processing.
Which providers support web and URL search in Kyvos?
Web and URL search in Kyvos runs on Azure OpenAI and Google Vertex AI. Amazon Bedrock does not support the capability.