RAG Chunk Size & Overlap Calculator

Fixed-size chunking slides a window of chunk_size tokens across your document, advancing by a strideeach step. The stride is the chunk size minus the overlap. Every calculation below is deterministic sliding-window arithmetic — the same logic LangChain's RecursiveCharacterTextSplitter and LlamaIndex's SentenceSplitter use to advance the window.

Let N be the document length in tokens, w the chunk size, o the overlap and s = w − o the stride.

1. Token estimate.Tokens are used as entered. Characters are converted at about 4 characters per token, and words at about 1.33 tokens per word, following OpenAI's tokenizer guidance. These heuristics are close for English prose but not exact.
2. Chunk count. If N ≤ w the whole document is a single chunk. Otherwise the count is ceil((N − o) / s). This is algebraically identical to ceil((N − w) / s) + 1; the tool computes both and they always agree, which is how you can trust the number.
3. Duplicated tokens. Every chunk after the first repeats exactly o tokens from its predecessor, so the overlap adds (chunks − 1) × o tokens on top of the document.
4. Total tokens embedded = N + (chunks − 1) × o. You pay to embed every one of them, including the duplicates.
5. Embedding cost = total_tokens ÷ 1,000,000 × price_per_1M, with the price taken from the selected model (OpenAI text-embedding-3-small at $0.02/1M, -3-large at $0.13/1M, ada-002 at $0.10/1M, Cohere embed-english-v3.0 at $0.10/1M, Voyage voyage-3 at $0.06/1M) or your custom figure. This is a one-time cost per index build.
6. Vector store size = chunks × dimensions × bytes_per_value, where a float32 value is 4 bytes, float16 is 2 and int8 is 1. Dimensions come from the model (1536, 3072 or 1024). This is the raw vector payload; database index structures and metadata sit on top of it.
7. Retrieval prompt budget = top_k × wtokens. This is the upper bound of context your retriever injects into the prompt each query — keep it well under the generation model's context window.

Two guards protect you from silent mistakes: overlap must be smaller than the chunk size (otherwise the window never advances), and if a chunk is larger than the embedding model's maximum input the provider truncates it, so the tool warns you.

• OpenAI — Embeddings guide (dimensions & max input tokens)
• OpenAI — API pricing (embedding price per 1M tokens)
• OpenAI — What are tokens (the ~4-characters-per-token heuristic)
• Cohere — Embed docs (embed-english-v3.0 dimensions & limits)
• Voyage AI — Embeddings docs (voyage-3 context length & dimensions)
• LangChain — RecursiveCharacterTextSplitter (chunk_size / chunk_overlap semantics)
• LlamaIndex — Node parsers (SentenceSplitter defaults)

Model prices, dimensions and input limits were last cross-checked against these sources on 2026-07-08. Providers change pricing without notice — use the Custom model option to enter current figures, and email me if a default looks stale.