Parameters

These settings are on Report Analyst, under the analysis configuration. A stored result is tied to the combination you used, so changing one of them is a new run.

Setting	Default	Effect
Chunk size	500	Length of each passage, in tokens. Larger chunks keep more context in one passage and produce fewer of them.
Chunk overlap	20	Tokens repeated between neighbouring passages, so a sentence on a boundary is not cut in half.
Top K	5	How many passages are retrieved for each question before the model answers.
Model	First available model	The LLM that writes the answer. OpenAI models need OPENAI_API_KEY. Gemini models need GOOGLE_API_KEY.
LLM scoring	Off	A second model pass that scores retrieved passages for the question. Leave it off unless you want that extra call.
Batch scoring	Off	Score passages together instead of one at a time. Only applies when LLM scoring is on.

The processing step (Chunk, Embed, Map, or Answer) is separate from these numbers. See report analysis.

Retrieval for benchmarking uses the same chunk size, overlap, and top K. Those fields are on RetrievalConfig in the benchmark models.