Structured LLM router analyzes the user prompt to decide whether to query local knowledge base or initiate web search.

Queries high-performance Arrow vector store, filtering by authenticated user_id with sub-millisecond retrieval.

Evaluates each retrieved chunk against the query. Discards irrelevant noise and flags web search fallback if context is insufficient.

If document context is inadequate or external real-time data is required, queries Tavily Search API to supplement context.

Synthesizes final answer strictly grounded in the filtered context, copying entities, dates, and numbers accurately.

Evaluates whether the generation is fully supported by the facts. If hallucinated claims are detected, retries generation.