An agent inventing prices or delays is worse than no agent. With Deep Agent Studio, your documents (pricing, procedures, FAQ, terms) feed a RAG knowledge base: the agent searches it before speaking, instead of improvising. Here is how to keep it reliable.
Keep the right documents, not all documents
RAG quality starts with triage. Keep what advisors really consult: current price grids, refund procedures, delivery times, account-access terms. Drop obsolete archives, drafts and duplicates — every dubious document is a potential dubious answer. Name each file clearly: the agent finds its way, and so will you at re-index time.
- Current, dated pricing and plans
- Procedures advisors actually apply
- Delays, terms, edge cases
- Zero obsolete archives, zero duplicates
Chunk to be found
An 80-page PDF can't be retrieved by chunks: the agent pulls a fuzzy passage and answers beside the point. Split content per question (one topic = one section), with explicit headings and short sentences. A section “Double-charge refund: 5-day delay, no fees, systematic ticket” will be found and quoted correctly. It is the highest-ROI work in any RAG project.
If a human takes over a minute to find info in your doc, the agent will miss it too. Split, head, simplify.
Re-index on every change
A price changing with no re-index means guaranteed wrong answers for weeks. In Deep Agent Studio, each bot has its Files & RAG tab: import, check indexing, and hit Re-index after every significant edit. Then test in the simulator with the trick question (“how much is plan X?”) before publishing. Make re-indexing a team reflex, like updating the website.
Frame the agent: sourced or silent
Give the bot one simple rule in its system prompt: answer from the base when info is there, say it doesn't know otherwise — and offer advisor handoff. “I don't have that information, I'm passing your request on” beats a polite invention every time. Verify with off-base questions: the agent must decline, not embroider.
FAQ
How many documents to start? Five to ten well-chunked sections suffice for a first intent. Little and exact beats lots and fuzzy.
What if two documents contradict? The agent will hesitate or err: delete or fix the obsolete version, then re-index. Contradiction is a content problem, not an AI problem.
Does RAG replace my APIs? No: documents for knowledge (prices, rules), APIs for facts (invoice, order, account). Both combine in one flow.
