How much does a RAG knowledge base cost? (2026 Australian guide)

How much does a RAG knowledge base cost to build?

Most RAG knowledge bases cost $25,000 to $150,000 (AUD, ex GST) to build, with enterprise systems above that. A RAG system (sometimes sold as an AI knowledge base, AI search or “chat with your documents”) answers questions by first retrieving relevant passages from your documents, then giving them to a language model to write a cited answer. If the term is new, start with what RAG is. If the audience is customers rather than staff, the AI chatbot cost guide covers the extra work.

Very few Australian firms publish RAG-specific pricing. The bands below are our synthesis of published Australian AI ranges: Team 400 prices a chatbot with knowledge base integration at $30,000 to $80,000, and Quanton AI puts a narrowly focused generative AI project at $20,000 to $50,000 and a production deployment at $50,000 to $150,000.

Scope	What’s included	Typical range (AUD, ex GST)	Typical timeline
Pilot	One document collection (for example a policy library), web chat interface, cited answers, a test set of 50 to 100 questions	$25,000 to $50,000	4 to 8 weeks
Department production system	Two to four sources (SharePoint, a document system, a wiki), scheduled re-indexing, single sign-on, feedback capture, monitoring	$60,000 to $150,000	3 to 5 months
Enterprise system	Many sources, document-level permissions mirrored from source systems, audit logging, multiple user groups, embedding into other applications	$150,000 to $350,000+	5 to 9 months

Where does the build budget go?

A RAG system has five parts, and ingestion and evaluation usually take more effort than the chat interface. Buyers tend to picture the chat window; most of the cost sits behind it.

Component	What it involves	What makes it expensive
Ingestion	Connecting to sources, extracting text from PDFs, Word, scans and web pages, splitting into chunks, attaching metadata	Scanned documents needing OCR, tables and forms, many source systems, frequent changes
Retrieval	Embeddings, vector and keyword search, re-ranking, filters by date, product or jurisdiction	Specialist terminology, similar-looking documents, questions needing several sources at once
Generation	Prompts, citation formatting, refusing when the answer isn’t in the documents	Strict tone or regulatory requirements, multi-step answers
Access control	Making sure each user only retrieves documents they’re allowed to see	Permissions stored in several systems, frequent staff changes, sensitive matters
Evaluation and monitoring	Test questions with expected answers, automated scoring, logging, feedback review	Needing subject experts to write and check answers; see how to evaluate an LLM application

Access control is the cost line most often missing from cheap quotes. A system that indexes the whole shared drive and lets everyone ask questions of it will happily surface HR files and board papers. Mirroring permissions from source systems is careful engineering and adds meaningfully to the budget, but in most organisations it isn’t optional.

How much does it cost to embed your documents?

Embedding is cheap: a large document collection costs a few US dollars to embed at current prices. This surprises buyers, because the word “AI” suggests a big compute bill.

Here is the arithmetic for a collection of 20,000 documents averaging five pages:

1. Estimate tokens. At about 500 tokens a page: 20,000 × 5 × 500 = 50 million tokens.
2. Price the embedding. At the time of writing, OpenAI lists text-embedding-3-large at US$0.13 per million tokens and text-embedding-3-small at US$0.02. So 50 × US$0.13 = US$6.50, or 50 × US$0.02 = US$1.00.
3. Estimate storage. With 500-token chunks, that’s about 100,000 chunks. Each large embedding holds 3,072 numbers at 4 bytes, about 12 KB, so the vectors take roughly 100,000 × 12 KB = 1.2 GB before indexes and text.

At that size, PostgreSQL with pgvector on a managed service is enough. Amazon RDS for PostgreSQL supports pgvector, and a managed Postgres plan such as Supabase Pro starts at US$25 a month at the time of writing. Dedicated vector databases such as Pinecone have their own pricing (its Standard plan has a US$50 monthly minimum at the time of writing). Our vector database comparison covers when each is worth it, including Australian hosting options.

The expensive part is the engineering around embedding: clean extraction, sensible chunking, metadata, and re-indexing only what changed.

What does a RAG system cost per 1,000 questions?

At the time of writing, model usage is roughly US$9 to US$36 per 1,000 questions depending on the model tier, before hosting. This is illustrative, and the method matters more than the result, because prices change often.

The assumptions for one question:

• System instructions: 1,500 tokens
• Retrieved passages: 8 chunks × 500 tokens = 4,000 tokens
• The question plus a little conversation history: 500 tokens
• Total input: 6,000 tokens. Answer with citations: 600 output tokens.
• Embedding the question itself: about 50 tokens, which costs almost nothing.

Per 1,000 questions, that’s 6 million input tokens and 0.6 million output tokens. Using Anthropic’s published prices at the time of writing:

Model tier (example)	Price per million tokens (input / output)	Input cost	Output cost	Per 1,000 questions (USD)	Approx. AUD
Fast, low cost (Claude Haiku 4.5)	US$1 / US$5	6 × $1 = $6	0.6 × $5 = $3	$9	$13
Mid-tier (Claude Sonnet 5)	US$2 / US$10	6 × $2 = $12	0.6 × $10 = $6	$18	$26
Premium (Claude Opus 5.5)	US$4 / US$20	6 × $4 = $24	0.6 × $20 = $12	$36	$51

AUD conversions use the Reserve Bank’s 25 September 2026 rate of about US$0.70 per A$1.

Prompt caching lowers this further. The 1,500-token system prompt is identical on every question, so it can be cached. On the mid-tier model, a cache hit costs 10% of the normal input price: 1.5 million cached tokens × (US$2 − US$0.20) saves about US$2.70 per 1,000 questions, roughly 15%.

What raises it: sending more chunks, long conversation history, agent-style multi-step retrieval (several model calls per question), premium models, and regional endpoints. At the time of writing, Anthropic notes a 10% premium for regional endpoints on AWS Bedrock and Google Cloud for recent models, and OpenAI a 10% uplift for regional processing on eligible models. The LLM running costs guide goes further on these levers.

How much does a RAG knowledge base cost per month? A 300-person firm

For a typical mid-sized organisation, the model bill is the smallest part of running a knowledge base. Illustrative figures, AUD ex GST:

Monthly item	Assumption	Cost
Model usage	300 staff × 4 questions a working day × 21 days = 25,200 questions; × $26 per 1,000 on a mid-tier model	$655
Hosting	Application, Postgres with pgvector, storage and backups in an Australian region	$600
Monitoring and logging	Error tracking, usage dashboards, log retention	$150
Maintenance	15% of a $110,000 build a year, ÷ 12	$1,375
Content governance	4 hours a week of a subject expert’s time at an internal cost of $90 an hour × 4.33 weeks	$1,559
Total		$4,339

That’s about $14.50 per staff member a month, or about 17 cents per question. The biggest line items are people: maintenance and content governance. Cutting the model bill in half would save about $330 a month; a bad answer about a policy could cost far more.

What are the hidden costs of a RAG system?

The hidden costs sit in what quotes leave out, so check whether a quote includes access control, evaluation and content clean-up, not just the chat interface. Also check for:

• GST at 10% on ex GST quotes.
• Model and embedding usage, billed in US dollars by the provider.
• Hosting, including separate development and production environments.
• OCR for scanned documents, which some services charge per page.
• Licences for source systems’ APIs or connectors.
• Your experts’ time writing test questions and reviewing answers.
• Content remediation. RAG makes contradictions in your documents visible. Fixing them is business work.

Is a RAG knowledge base worth the cost?

A RAG knowledge base is worth it when many people spend real time searching for answers in documents that already exist, and a wrong or slow answer has a cost. Policy libraries, contracts, technical manuals and procedures are the classic fits. It is not worth building when the documents are few, rarely consulted, or so out of date that the first job is rewriting them.

The break-even arithmetic is simple. In the example above, the system costs $4,339 a month to run. At the same internal cost of $90 an hour, it covers its running costs if it saves about 48 hours a month across the firm: roughly 10 minutes per person per month for 300 staff. Recovering the build cost as well takes more, so estimate honestly how often people search today and how long it takes them. A pilot on one collection, with usage logging, gives you real numbers before you commit to production.

How do you reduce RAG costs without hurting answer quality?

Spend on retrieval quality and evaluation, save on everything else.

1. Start with one high-value collection where people waste time searching today.
2. Remove duplicates and superseded versions before indexing. Fewer, better documents improve answers and reduce cost.
3. Use pgvector in your existing Postgres unless scale demands otherwise.
4. Route by difficulty. A low-cost model can handle simple lookups; reserve premium models for complex questions.
5. Cache the stable parts of the prompt.
6. Build the test set first. It stops you paying for tuning that doesn’t help.
7. Consider buying. If your content lives in Microsoft 365 and generic answers are acceptable, an off-the-shelf assistant may cost less. See RAG vs fine-tuning and our other guides before committing.

How All Webbed Labs builds RAG knowledge bases

We build RAG systems on PostgreSQL with pgvector by default, hosted in Australian cloud regions, with document-level permissions carried over from your source systems. Every build starts with a test set of real questions and expected answers written with your subject experts, and we report accuracy against it before launch and after every significant change. We quote a fixed price after paid discovery, and the code, indexes and test set stay in your accounts. See our RAG knowledge base development service.