Status: Phase 0 placeholder. Real measurements get filled in during Phase 3.
Retrieval quality is bounded by chunk quality. Bad chunks = bad retrieval = bad RAG, no matter how good the model is. Embedding similarity is computed per chunk, so a chunk that mixes two unrelated topics scores poorly on both. A chunk that splits a sentence mid-clause loses the answer.
Without overlap, a sentence that spans a chunk boundary is split — the answer to a question can sit half in chunk N and half in chunk N+1, and neither chunk alone has enough context to be retrieved or cited correctly. Overlap (typically 10–15% of chunk size) gives boundary content a second chance to be embedded as part of a coherent unit.
Chunk size: 800 tokens
Overlap: 100 tokens (~12.5%)
Tokenizer: cl100k_base (GPT-4 family)
Boundary: prefer paragraph > sentence > hard split
Rationale: 800 tokens is a common sweet spot — large enough to carry a full thought, small enough that retrieval scoring stays specific.
| Size | Overlap | Hypothesis |
|---|---|---|
| 400 | 50 | Tighter, more chunks, higher recall but more noise |
| 800 | 100 | Baseline |
| 1200 | 150 | Fewer chunks, richer context, may dilute scores |
Each configuration runs against the Phase 10 evaluation set. We compare retrieval recall, answer correctness, and groundedness — not vibes.
To be filled in during Phase 3 with side-by-side numbers and 1–2 concrete bad-chunk examples per configuration.
| Config | Recall@5 | Correctness | Groundedness | Notes |
|---|---|---|---|---|
| 400/50 | TBD | TBD | TBD | |
| 800/100 | TBD | TBD | TBD | |
| 1200/150 | TBD | TBD | TBD |
To be recorded once data exists. Format: "Chose X/Y because Z; tradeoff was W."