Guide
Chunking strategies for RAG
Pick a strategy from the three FAQai implements. The right size depends on whether you need a short fact or a procedure that should stay together.
Start with Balanced
Balanced is heading-aware, about 400 tokens, with overlap so a sentence on a boundary can appear in the neighbouring chunk. Use it unless you already know the document needs shorter or longer units. The Free plan is fixed to this preset.
When to change it
- Fine-grained, about 200 tokens, when each answer is a short fact and a long chunk would mix unrelated topics.
- Large context, about 800 tokens, when a procedure or policy should stay in one chunk.
- Fixed size when headings are unreliable and you want uniform length.
- Paragraph when blank lines already mark the ideas you want to keep together.
What overlap does
Overlap repeats characters from the end of one chunk at the start of the next. The Balanced preset uses 250 characters. Custom overlap on a paid plan cannot exceed 25 percent of the target. Overlap does not create embeddings. The export still carries embedding_text for your model to embed.
The control surface is described on the document chunking feature.