RAG Cost Calculator

Estimate what a Retrieval-Augmented Generation pipeline actually costs: the one-time job of embedding your corpus, the monthly bill to store those vectors, and the per-query embedding cost — with a self-hosted $0/token comparison. Every price is an editable input, so it works for any provider.

🔒 100% in-browser — nothing you enter is uploaded

Your corpus & usage

Total text chunks (or documents) to index.
Avg size of each chunk.
Searches served per month.
Avg length of each question embedded.
Presets fill the rate & dims — both editable.
Managed billing varies; self-host is $0/GB.
What your provider charges per 1M tokens.
Output dimensions of the embedding model.
Vector-DB storage rate; 0 for self-host.
One-time index build
$0
—
Ongoing / month
$0
—
Cost component Basis Rate Cost

Embedding rates and dimensions default to typical list values for reference only — they change often, so edit the fields to match your provider's current pricing. Confirm live rates at OpenAI, Cohere, and Voyage. Vector storage is modeled as a simple $/GB·month; real managed services (Pinecone, Weaviate, others) often bill by pod, serverless read/write units, or monthly minimums, so plug in your own effective rate. Raw vector size is chunks × dimensions × 4 bytes (float32). The Local (ABUZ8 OS) row is $0 per token and $0 per GB — marginal cost is electricity only; fixed hardware is covered in the self-host vs cloud calculator. This tool sizes the retrieval layer; the generation model that writes each answer is priced separately in the LLM price comparison.

Own your index, not just rent your answers.

Every row above is a bill that scales with your corpus and traffic. ABUZ8 OS embeds your documents with open models and stores the vectors on hardware you own — $0 per token, $0 per GB, and your data never leaves the machine. Sovereign RAG by default, cloud only when it genuinely wins.

✓ You're on the list. We'll be in touch.

Understanding what a RAG pipeline costs

Related: LLM API price comparison · Self-host vs cloud calculator · Local LLM VRAM calculator · Token counter

The three buckets: build, store, query

Retrieval cost splits cleanly. Build is the one-time job of embedding your entire corpus — it scales with total tokens and happens once (or again whenever you re-index). Store is the recurring vector-database bill for keeping those embeddings searchable. Query is the per-search cost of embedding each incoming question. This calculator shows all three, plus the raw size of the vectors so you can sanity-check any managed quote against what you're actually storing.

Why embedding is usually the cheap part

Embedding rates are a fraction of chat-completion rates, so for most corpora the one-time build lands in the single or double digits of dollars. The recurring spend is dominated by two things this tool deliberately keeps separate: the vector database (which rarely bills as pure per-GB storage) and the generation model that writes each answer. If your RAG bill feels high, it's almost always the hosting tier or the generation LLM — not the embeddings.

When self-hosting the retrieval layer wins

An open embedding model on a GPU you own has no per-token charge, and a self-hosted vector store (FAISS, pgvector, Qdrant, Chroma) has no per-GB charge — the marginal cost of both is electricity. Above a modest scale, owning the retrieval layer beats renting it, and the savings compound as your corpus and traffic grow. ABUZ8 OS runs this locally by default and only reaches for a paid API when a hosted model clearly beats your local one.

Is my data sent anywhere?

No. This calculator runs entirely in your browser. Nothing you enter is uploaded, logged, or stored.