The Problem

Workflow Steps

1. 1
   Chunk your KB documents at 512 tokens with 50-token overlap. Claude does not provide a native embedding model — use a third-party embedding model such as OpenAI text-embedding-3-small, Cohere embed-v3, or any sentence-transformers model. Store embeddings in Pinecone or FAISS with metadata: {doc_id, section, last_updated}
2. 2
   Query Router Agent: classify each incoming query as factual_lookup / how_to / billing / complaint / out_of_scope — route out_of_scope directly to human escalation; route complaint queries to the Answer Agent with an additional instruction to lead with empathy before policy
3. 3
   Retriever: run semantic search, return top-3 chunks with cosine similarity scores; flag as low-confidence if max score < 0.72
4. 4
   Confidence Gate (LangGraph conditional edge): similarity < 0.72 → Clarification Agent; similarity >= 0.72 → Answer Agent
5. 5
   Clarification Agent: generate one specific clarifying question (no compound questions), re-run retrieval with original query + user clarification concatenated; if still low-confidence, escalate to human queue
6. 6
   Answer Agent: generate answer grounded only in retrieved chunks, cite source chunk IDs, flag answer_confidence as high/medium/low
7. 7

Prompt Example