RAG: advanced
TemplatesArchitecture
Production RAG: input guard, semantic cache, hybrid (vector + keyword) retrieval, cross-encoder reranking, a system prompt, PII redaction on the way out, and traces feeding groundedness evals.
What is inside
- Hybrid retrieval + reranker
- Guardrails in and out
- Traces and evals
Use this template
Opens in the playground, no account needed. Sign up to save it.
Show the Inkline text
diagram architecture "RAG: advanced"
direction down
person user "User"
client reply "Answer in the chat"
zone offline "Offline: ingestion" {
corpus docs "Knowledge base"
chunker chunk "Chunk + metadata"
embedder embed "Embed chunks"
}
vectordb vectors "Vector index"
store keyword "Keyword index"
zone online "Online: answering" {
guardrail inguard "Input guard"
cache semcache "Semantic cache"
retriever hybrid "Hybrid retriever"
reranker rerank "Cross-encoder rerank"
prompt system "System prompt"
llm model "Chat model"
guardrail outguard "PII redaction"
}
zone quality "Quality" {
tracer traces "LLM traces"
eval evals "Groundedness eval"
}
docs -> chunk -> embed
embed -> vectors
chunk -> keyword
user -> inguard : question
inguard -> semcache
semcache -> hybrid : miss
vectors -> hybrid : dense hits
keyword -> hybrid : keyword hits
hybrid -> rerank
rerank -> model : top 5 chunks
system -- model
model -> outguard
outguard -> reply : answer
model ~> traces
traces ~> evals : nightly