# Method: chunk granularity vs. retrieval (agent-memory-chunk-granularity-recall-2026) Measured 2026-11-15T00:46:10Z on the task host. Localhost only, no network, no credentials, no paid API. ## Reproduce python3 chunk_probe2.py fixture.json results2.json `chunk_probe.py` is the earlier sweep (hits only, 66 s); `chunk_probe2.py` additionally records the rank of each needle sentence to depth 8 and reports `both_in_top8` (54 s). Both need Ollama serving the three models below at http://127.0.0.1:11434. Nothing else is required. ## Fixture 8 invented documents x 7 sentences (456-500 chars per document, 52-99 chars per sentence). 8 single-fact questions (one needle sentence per document) and 5 two-fact questions whose two required sentences sit on opposite sides of a 3-sentence window boundary. ## Granularities and embedders * whole_document: 8 vectors; three_sentence: 24 vectors (windows 1-3, 4-6, 7); single_sentence: 56 vectors. * all-minilm:latest 384-d digest 1b226e2802db; snowflake-arctic-embed:33m 384-d digest e8db018629b4; nomic-embed-text:latest 768-d digest 0a109f422b47. Ollama 0.20.3, F16, 2 vCPU / 11 GiB. * Scoring: cosine on L2-normalised /api/embed vectors, deterministic full sort, no sampling, no generation. ## Known deviations to carry into the article * No task-instruction prefixes were sent. nomic-embed-text-v1.5's model card requires `search_document:` / `search_query:`; Ollama's own library example omits them. nomic's absolute numbers are therefore not a fair model comparison. * `prompt_eval_count` was not recorded, so token counts are not measured; only characters are. * The 5 two-fact pairs straddle a window boundary by construction: fine-granularity two-fact hit@1 of 0 is a designed worst case, not an estimate.