Topic 2: Strategies
14 min read·21 Sept 2026
The strategies at a glance. Read top to bottom: each one is smarter and usually more expensive than the one before.
| # | Strategy | Splits on | Cost | Best for |
|---|---|---|---|---|
| 1 | Fixed-size (characters or tokens) | A fixed length | Lowest | Uniform text, quick baselines |
| 2 | Sentence / paragraph | Natural language boundaries | Low | Prose |
| 3 | Recursive with separators | Biggest boundary that fits | Low | General-purpose default |
| 4 | Structure-aware | Headings, sections, lists, tables | Low | Manuals, policies, wikis |
| 5 | Format-specific | Markdown headers, HTML tags, code symbols | Low | Docs sites, HTML, code |
| 6 | Semantic | Shifts in meaning (embedding distance) | Medium (embeddings) | Long unstructured text, transcripts |
| 7 | Proposition-based | Atomic facts written by an LLM | High (LLM per passage) | Dense factual text |
| 8 | Agentic / LLM-driven | Boundaries chosen by an LLM | Highest | Small, high-value, messy documents |