bhavikmangla/mdkeychunker-study

MDKeyChunker confirmatory study: chunk and enrichment caches

Every chunk and every LLM-generated field used in the confirmatory study of Upgrade the Retriever First: What an LLM Call per Chunk Buys in RAG (Mangla, arXiv:2603.23533, v4).

With these files, every retrieval result in the paper can be re-scored without a single LLM call. The harness, analysis plans, amendments and per-question results are in the code repository: https://github.com/bhavik-mangla/MDKeyChunker (benchmarks/chunking_study).

What is inside

677 source documents: 300 Qasper papers (files named by arXiv id, e.g. 1603.01514.json) and 377 FreshStack documentation files (named <corpus>__<path>.json, e.g. laravel__controllers.md.json; corpora Laravel, Angular and Ultralytics YOLO). Each variant folder holds one JSON file per document:

{"chunks": [{"text": "...", "section": "Title > Section", "title": "...",
             "summary": "...", "keywords": "[...]", "questions": "[...]",
             "key": "...", "related_keys": "[...]", "entities": "[...]",
             "context": "...", "llm_failed": "False"}],
 "stats": {"spec": "ollama:qwen2.5:7b", "calls": 26, "in_tokens": 149543,
           "out_tokens": 1622, "parse_retries": 0, "failed_chunks": 0}}

Fields an arm does not produce are empty. stats records calls, tokens, retries and spend for the document.

Folder	Paper arm	Generator	What it is
base/struct	S, S+T	none	Header-led Markdown chunks (Stage 1); S+T prefixes the section path at indexing time
base/fixed512	F512c	none	512-character windows
base/fixedtok	F256t	none	256-token windows with 32-token overlap
base/fixedlen	FL	none	Windows matched to the median S chunk length
base/parapack	P	none	Header-blind paragraph packer
local/enr	E	qwen2.5:7b	One enrichment call per chunk (title, summary, keywords, questions, key, related keys, entities), rolling key dictionary on
local/enr_nork	E−K	qwen2.5:7b	The same call without the rolling key dictionary
local/cr	CR	qwen2.5:7b	Contextual retrieval: one situating sentence per chunk (context)
local/perfield	PF	qwen2.5:7b	One call per field instead of one per chunk (100 Qasper papers)
local/merged	M	none (from E)	Key-based merges (Stage 3)
local/merged_nork	M−K	none (from E−K)	Key-based merges without the dictionary
local/enr_tp, local/merged_tp	M−TP	qwen2.5:7b	E−K keys after one TopicGPT-style consolidation call per document, and merges on them
local/rmerge_s0–s4	RM	none	Random adjacent merges, same count as M, seeds 0–4
local/rmerge_tp_s0–s4	RM (TP)	none	The same, matched to M−TP
local/smerge, local/smerge_tp	SM	none	Embedding-similarity merges, same count as M / M−TP
local/h4_merge_counts_*.json			Merge counts per document used to match the controls
strong/enr	Estrong	gpt-oss-120b	Enrichment with the strong generator
strong/cr	CRstrong	gpt-oss-120b	Contextual retrieval with the strong generator (Qasper only)
qwen235b/enr	scaling arm	Qwen3-235B-A22B-Instruct-2507 (DashScope)	Small exploratory scaling arm (23 documents)

The query-time arms (E+KX, E+AX, E+SX, late chunking) and the retriever variants need no extra caches: they are computed from the folders above.

How it was made

Local generation used qwen2.5:7b (Ollama, 4-bit) at temperature 0, seed 0; the strong generator was openai/gpt-oss-120b through OpenRouter, pinned to one provider with data collection disabled. Prompts, decoding settings and model digests are fixed in results/v2/frozen_config.json in the code repository, and the analysis plan was made public before the first build. Chunks whose output could not be parsed after three retries keep empty fields and are flagged llm_failed.

Licence and source data

The chunk text is taken from the source datasets, so this dataset keeps their terms and is released under CC BY-SA 4.0:

• Qasper (Dasigi et al., 2021), CC BY 4.0.
• FreshStack (Thakur et al., 2025), CC BY-SA 4.0; its corpora come from the Laravel, Angular and Ultralytics documentation repositories, under their own open-source licences (MIT, BSD, Apache-2.0, CC BY 4.0, GPL-3.0, AGPL-3.0).

The generated fields are model output and carry no extra restriction beyond the above. Use for research evaluation.

Citation

@article{mangla2026mdkeychunker,
  title   = {Upgrade the Retriever First: What an {LLM} Call per Chunk Buys in {RAG}},
  author  = {Mangla, Bhavik},
  journal = {arXiv preprint arXiv:2603.23533},
  year    = {2026}
}