How it works
RAG vs tools for customer support
By the RoutMe team. Last updated .
Retrieval-augmented generation (RAG) answers a question by retrieving relevant passages from documents and giving them to the model. Tool use answers it by calling a live system, such as an order API. RAG suits policies and how-to content that changes slowly, and tools suit records that change constantly.
What each approach does
RAG was described in a 2020 paper as combining a language model with a retriever over an external document collection. In practice, documents are split into passages, indexed so they can be searched by meaning, and the best matches are placed in front of the model when a question arrives.
A tool call is different. The model chooses a named function, such as get_order, supplies arguments, and receives the live result. Nothing is indexed in advance.
Comparison
| RAG | Tools | |
|---|---|---|
| Source | Documents you add | Live systems via API or MCP |
| Best for | Policies, FAQs, sizing, how-to | Orders, stock, accounts, availability |
| Freshness | As of last indexing | As of the moment of the question |
| Per-customer data | Not suited | Suited, with verification |
| Setup | Paste, link or upload content | Build or connect a tool server |
| Main risk | Stale or missing passages | Oversharing or unverified access |
Examples from a store
- "What is your return policy?" is a RAG question. The answer lives in a page you wrote.
- "Where is order 1043?" is a tool question. The answer is in the order system right now.
- "Can I return the jacket I bought last week?" needs both: the order date from a tool and the return window from a document.
How RoutMe does each
Knowledge sources in RoutMe can be pasted text, a URL or an uploaded spreadsheet up to 20 MB. Documents are split into passages of about 1,200 characters on paragraph boundaries with a short overlap, and the assistant searches by meaning and receives the six most similar passages, each labeled with its source. A URL source is a snapshot taken when you add it and is not re-fetched, so re-add a page after it changes. Details at docs.routme.io/knowledge.
Tools come in through MCP. A single assistant turn allows up to six rounds of tool calls before it must answer, and a failed call returns to the assistant as an error so it can say it could not look something up.
Choosing the mix
- List your top ten questions.
- Mark each as a document question, a record question, or both.
- Add the documents first, since that takes minutes.
- Build tools for the record questions, starting with order status.
- Test with questions your documents do not answer, and check that the assistant says it cannot confirm.
Common mistakes
- Putting live data in documents. An exported price list goes stale. Use a tool.
- Giving a tool to answer a policy question. It is slower and adds risk. Use a document.
- Skipping the "I do not know" test. The assistant should say so when nothing relevant is found.
Keeping both sources honest
Both approaches fail quietly if nobody owns them. For documents, assign an owner to each policy page and review it on a schedule, because the assistant will confidently quote whatever is indexed. For tools, write down what each tool returns and what it deliberately leaves out, and review the list whenever the underlying API changes. A useful habit is a weekly sample: read ten real conversations, mark where the answer came from, and note any that came from neither a document nor a tool. Those are the answers the model supplied from general knowledge, and they are the ones to fix by adding a source or by teaching the assistant to say it cannot confirm.
Sources
Explore the product: RoutMe for Shopify, solutions and the documentation.