EVA: RAG + LangGraph agent + cross-encoder reranker

question
user query
route
LangGraph agent
bi-encoder
top-N=10 recall
cross-encoder
top-K=3 precision
Claude Haiku 4.5
generation with sources
grounded answer
FinOps guardrails
  • relevance gate
  • input length cap
  • API GW rate limit
  • budget alarms
memory + coref
rewrites pronouns using conversation history
canned
off-topic - short refusal
reformulate guard
retry if weak score. revert if new score is worse