Combined Router

The Combined router is an orchestrator that routes queries to VecSearch, NL2SQL, or both. Unlike the other agents and flows, it does not have an AgentSpec definition — it coordinates existing sub-sessions at runtime.

• The classifier prompts the LLM to respond with exactly one word — nl2sql, vecsearch, or both — based on the user's question. Unrecognized responses default to both.
• When routed to a single tool, the query is dispatched directly to the corresponding sub-session.
• When routed to both, the sub-sessions run in parallel. If VecSearch finds relevant documents, both results are fed into a synthesis LLM call to produce a unified response; otherwise, the router returns the NL2SQL answer directly.
• The system prompt is fetched from the MCP server (optimizer_tools-default). If unavailable, a default instruction is used.
• Token usage from the classifier, sub-sessions, and synthesis calls is aggregated.
• Requires both a configured VecSearch flow and an NL2SQL agent to be available.

Model Use and Fallbacks

The router uses the selected primary language model for both classification and synthesis, as well as for the NL2SQL and VecSearch sub-sessions. There is no separate classifier-model setting in the client. Applications that need a smaller, dedicated model for the lightweight classification and synthesis calls must provide it through custom runtime wiring.

Classification adds one language-model call to every combined request; synthesis adds another when both routes return relevant results. Consider this additional latency and token use when selecting a model for high-throughput or cost-sensitive workloads.

If classification fails or produces an unrecognized result, the router runs both routes. If synthesis fails, it returns the NL2SQL and VecSearch answers under labelled headings instead.