How I Built Bob: My n8n AI Chatbot for Docs (RAG), and What It Really Costs in AUD

By | November 20, 2026
Estimated reading time: 8 minutes · Costs measured from my own OpenRouter logs, July to October 2026
Affiliate disclosure: this post contains an affiliate link to RackNerd, marked “affiliate link”. If you buy through it, I earn a commission at no extra cost to you. The costs in this post are what I actually paid either way.
Bob is my personal chatbot. Ask him a question and he searches ServiceNow’s documentation, answers with direct quotes, and tells you which page each quote came from. Here’s how he works, what he’s made of, what went wrong along the way, and what he really costs to run, worked out from 81 real questions.
Short answer: Bob has no fixed monthly cost of his own. His database (Weaviate) and web search (Tavily) are on free plans, and n8n runs on the small server this blog already uses. What he does cost is the AI model: about A$0.17 per question (the median of 81 questions I asked him), paid to OpenRouter for Claude. Loading the documentation in the first place cost A$2.11, once.

Why I built Bob

ChatGPT and Claude are good at explaining ServiceNow in general. Where they slip is the detail: they’re often a release or two behind, and they can get a setting name or a step quietly wrong. When I need the answer for the release I’m actually on, I want it straight from the documentation, with a quote I can check.
So I built a chatbot that only answers from the docs. This kind of setup is usually called RAG (retrieval-augmented generation): before the AI answers, it looks up the relevant pages, and it answers from those first.

What Bob is made of

Part What I used Its job What it costs me
Workflow n8n, self-hosted Runs everything below Free software
Server A small RackNerd VPS Runs n8n (and this blog) Already paid for: about A$6.66 a month, shared
Chat window n8n’s Chat Trigger Where I type questions Free
Brain Claude Sonnet, through OpenRouter Decides what to search for and writes the answer Pay per use: about A$0.17 a question – Can be changed to less expensive or free model
Memory n8n’s window memory, last 50 messages Remembers the conversation Free
Documentation store Weaviate Cloud, Free plan Holds the docs, split into chunks, ready to search A$0
Embeddings OpenAI text-embedding-3-small, through OpenRouter Turns text into numbers so similar text can be found A$2.11 once to load the docs; a tiny fraction of a cent per search after that
Web search Tavily, free plan Backup when the docs don’t have the answer A$0
OpenRouter is a single account that gives you access to models from Anthropic, OpenAI, Google and others, billed per request. I use it so I only have one prepaid balance to keep an eye on.

How a question moves through Bob

  1. I ask a question in n8n’s chat window.
  2. The AI Agent node sends Claude my question, Bob’s rules and the recent conversation.
  3. Claude decides what to search for. n8n turns that search into an embedding and asks Weaviate for the 20 closest chunks of documentation.
  4. The chunks go back to Claude, which can search again with different words if they don’t answer the question. Half of my questions needed three calls to Claude or fewer.
  5. Claude writes the answer in a fixed shape: Answer, Details, Supporting quotes, Sources.
  6. If the docs don’t have it, Bob can search the web with Tavily, and his rules make him say so.
A full answer usually takes 45 to 60 seconds. That’s slow for a chatbot, but it’s the price of searching first instead of guessing.

The rules I gave Bob

Most of what makes Bob useful is in his instructions (the system prompt), about 5,500 characters of them. In short:
  • Search the documentation first, every time, even when you think you know the answer.
  • Don’t answer from general knowledge. If it isn’t in the docs, say so.
  • Back every answer with short, direct quotes and the page each one came from.
  • Don’t invent steps and don’t mix up different releases.
  • Be upfront when an answer comes from the web instead of the docs.
  • For broad questions, give an overview and point to the pages that go deeper.
The fixed answer format matters more than I expected. Because every answer ends with its quotes and sources, I can check Bob in a few seconds instead of taking his word for it.

Loading the documentation

Bob only knows what’s in Weaviate, so the first job was loading the documentation. I had Claude Code write a script that reads the docs as Markdown files, splits them into chunks, gets an embedding for each chunk and stores them in a Weaviate collection.
The OpenRouter key I made for that job shows 73.2 million tokens and US$1.45 (A$2.11) in total. That’s a one-off. Bob only needs new embeddings when the docs change.

What Bob actually costs

These numbers come from my own OpenRouter activity logs: 81 questions Bob answered between 6 July and 16 September 2026, all on Claude Sonnet 5.
What US$ A$
A typical question (median) 0.12 0.17
The middle half of questions 0.06 to 0.23 0.09 to 0.33
Average (pulled up by a few long ones) 0.22 0.31
The dearest answered question 1.28 1.87
My busiest month, August (48 questions) 11.84 17.21
Loading the docs (once) 1.45 2.11
Weaviate, Tavily and n8n 0 0
Claude Sonnet 5 and Sonnet 5.5 both cost US$2 per million input tokens and US$10 per million output tokens on OpenRouter, checked 12 November 2026. Bob’s model is set to “latest Sonnet”, which moved to Sonnet 5.5 at the end of September at the same price. Converted at the Reserve Bank rate for 11 November 2026 (US$1 = A$1.4539). These are OpenRouter’s charges per request. OpenRouter is prepaid and charges a fee when you buy credits, and your bank may add an international fee: see how to pay for AI without bank fees.

Why one question costs that much

Every time Claude searches, the 20 chunks it gets back are added to the conversation, and the whole conversation is sent again on the next call. Here’s one real question from 16 September:
Call to Claude Tokens sent Tokens back Cost (US$)
1. Read the question, decide what to search 2,509 181 0.007
2. Read 20 chunks, search again 18,659 195 0.039
3. Read 20 more chunks, search again 39,294 301 0.082
4. Write the answer 61,428 1,819 0.141
Total 121,890 2,496 0.269 (A$0.39)
About 98% of the tokens Bob uses are input, mostly documentation he’s re-reading. The answer itself is the cheap part. If you want to understand tokens better, AI token pricing explained walks through it.

What went wrong (and the fixes)

The OpenRouter connection that didn’t work

My first attempt used n8n’s OpenRouter credential, and the chat node failed with “Incorrect API key provided”, pointing at OpenAI. The request was going to OpenAI’s address with my OpenRouter key. The fix: use n8n’s OpenAI Chat Model node with an OpenAI-type credential, set its Base URL to https://openrouter.ai/api/v1, and paste the OpenRouter key in as the API key.

The question that gave up after ten rounds

On 18 July, Bob searched ten times in a row without finding what he wanted, hit n8n’s limit of 10 rounds and stopped without an answer. Each round sent more text than the last, so that non-answer cost US$1.78 (A$2.59), more than 15 typical questions. n8n’s AI Agent has a Max Iterations setting (10 by default). A lower limit makes a stuck question give up sooner and cheaper.

Running out of credit

OpenRouter is prepaid. When the balance runs out, Bob doesn’t warn you, he just stops answering. OpenRouter has an auto top-up option on its Credits page if you’d rather not keep checking.

Lock the chat down

If you turn on Make Chat Publicly Available in the Chat Trigger, anyone with the link can ask questions, and every question is paid for from your OpenRouter balance. Set the trigger’s authentication to Basic Auth (a shared username and password) or n8n User Auth (only people logged in to your n8n).

How to make a bot like Bob cheaper

  • Return fewer chunks. With 20 chunks per search, each round added roughly 16,000 to 22,000 tokens to every call after it. Fewer, better chunks shrink every later call.
  • Lower Max Iterations so a question that isn’t going anywhere stops early.
  • Look at prompt caching. OpenRouter lists cached input for Sonnet 5.5 at US$0.10 per million tokens, against US$2 normally. My logs show no cached tokens at all, so every round paid full price to re-read the same rules and chunks. That’s the next thing I want to test.
  • Keep conversations short. Bob remembers the last 50 messages, and all of them are sent with every call. A new chat for a new topic is cheaper.

Should you build your own n8n AI chatbot?

If you keep looking things up in the same set of documents, yes, it’s worth a weekend. Nearly all of Bob is drag-and-drop nodes in n8n, and the only code (the loading script) was written by Claude Code. The running cost is small if you use it for real questions, and you can see every cent of it in OpenRouter’s logs.
The cheapest way to run n8n is on your own server. I use a small RackNerd VPS (affiliate link) that renews at US$54.99 a year and also runs this blog. Self-hosted n8n vs n8n Cloud compares the options, and the cheapest VPS for Australians covers servers.

FAQ

How much does it cost to run an AI chatbot in n8n?

For Bob, about A$0.17 per question on Claude Sonnet, plus A$2.11 once to load the documentation. The database and web search are on free plans, and n8n runs on a server I already pay for. A bot that searches less, or uses a cheaper model, would cost less per question.

Do I need to know how to code?

Not much. The chatbot itself is built from n8n’s nodes. Loading documents into the database is the one part that needed a script, and I had Claude Code write it.

Is Weaviate Cloud free?

Its Free plan costs US$0 and includes 100,000 objects, 10 GB of disk and one collection, according to Weaviate’s pricing page. The cheapest paid plan, Flex, starts at US$45 a month (about A$65).

Can I build this on n8n Cloud instead?

Yes. The same nodes work on n8n Cloud, but you pay for the n8n plan on top of the AI costs. Self-hosted n8n vs n8n Cloud has the prices in AUD.
Costs are what I paid on my own accounts and will be different for your documents, questions and model. Prices can change, so check them before you build.

Get the monthly AI price email

Once a month: what changed in AI prices for Australians, plus links to any new FNG guides. Free, and you can unsubscribe any time.

Written by Scott (Freaky Newguy), a Sydney-based IT worker. How I check prices.

Sources



Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.