Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →In one small codebase experiment, returning a whole document led to more correct answers overall than returning the top 10 retrieved chunks. But the advantage was concentrated in code queries: prose queries favored chunks slightly. The result is useful evidence about this setup—not a general rule that whole files are always better.
What did the experiment compare?
Mikhail’s September 2026 test asked whether the unit of context returned to a reader model affects its answers after retrieval: should it receive several chunks, an entire document, the known-correct file, or no context? The setup used 16 frozen queries—eight about code and eight about prose—and ran each of four arms for 10 trials per query, for 640 judged answers in total. A reader model produced each answer and a separate judge call evaluated it. The author reports pinned reader and judge model IDs, but the findings remain specific to this harness and its September 2026 setup. Source: Mikhail, “The Unit of Return,” published September 27, 2026.
As an Amazon Associate I earn from qualifying purchases.
| Arm | Context returned | Correct answers |
|---|---|---|
| A | Top 10 retrieved chunks | 26/160 (16.3%) |
| B | Whole top-1 retrieved document | 55/160 (34.4%) |
| C | Gold file supplied as oracle context | 156/160 (97.5%) |
| D | Empty context (closed book) | 0/160 (0%) |
These are answer-accuracy results reported by the author, not independent measurements. Each arm had 16 queries × 10 trials, hence a denominator of 160. The oracle is a useful upper-bound comparison: it shows that the reader usually answered correctly when given the known-relevant file, but it was not perfect. The empty-context arm shows how the reader did without supplied evidence in this test; it does not establish how other models or tasks would perform.
Did whole documents help equally for code and prose?
No. The aggregate advantage masks opposite patterns in the two query groups. Code answers improved substantially with whole-document context, while the prose results leaned modestly toward chunks.
#1 Best Overall
| Query group | Top-10 chunks | Whole document |
|---|---|---|
| Code (8 queries; 80 trials per arm) | 5/80 (6.3%) | 40/80 (50.0%) |
| Prose (8 queries; 80 trials per arm) | 21/80 (26.3%) | 15/80 (18.8%) |
The author reports Wilson 95% intervals for several rates: the code-arm intervals did not overlap in the reported comparison, whereas the prose intervals overlapped. With only eight queries in each group, that interval pattern should not be read as replication or a broad guarantee; the prose difference is best treated as directional. The query-level majority counts tell a similar but stricter story: whole documents produced a strict majority of correct trials on five of 16 queries, compared with two of 16 for chunks. A 5–5 split is a tie, not a majority.
What does retrieval accuracy say—and what does it not say?
The experiment reports two retrieval measures that happen to have the same hit-at-one count but describe different events:
Rank #2
- 🖥POWERFUL PROCESSOR and SUPERIOR STORAGE: Configured with top of the Intel Core i5 processor for lightning-fast, reliable and consistent performance to ensure an exceptional PC experience. 16GB RAM memory to smoothly run multiple applications and browser tabs all at once. 2TB HDD storage space to store apps, games, photos, music, and movies. Loaded with 16GB to zip through multiple tasks in a hurry without lag.
- 🖥️New 22 Inch Full HD (1920x1080) LED monitor: with 75hz, High-Quality panel with quick refresh rate and response time. With 1080p resolution, you can enjoy gaming or a modern computing experience. 22 Inch monitor has a Smart Contrast to provide optimized image quality. Bezel-less and sleek design with glossy finish, crisp edge-to-edge visuals. Wide Viewing Angles for clarity from any viewpoint. VESA Mountable and built-in tilt options allow for a variety of monitor configurations.
- ⌨️ +🖱️ RGB KEYBOARD AND MOUSE | RGB SPEAKER: 3 LED Colors - Blue, red, green, Backlight LED Lights for use at night time, looks amazing. The keyboard mouse and speaker are responsive, reliable, and probably plastered in RGB lights. It's important you pick the right one for your desktop.
- 💿 WINDOWS 10 Pro LATEST: A new installation of the latest Microsoft Windows 11 Professional 64 Bit Operating System software, free of bloatware commonly installed from other manufacturers. As Microsoft's latest and best OS to date, Windows 10 Pro 64 Bit will maximize the utility of each PC for years to come. Optional software such as Anti-Virus and Office 365 can also be easily downloaded through the Microsoft Windows App Store.
- For chunk retrieval, the correct target appeared as the top-ranked chunk in 5 of 16 queries (hit@1); it appeared within the top three and within the top ten in 6 of 16.
- For document retrieval, the top-ranked document was the gold file in 5 of 16 queries (top-1-is-gold).
Equal counts do not make the events equivalent: one checks a chunk ranking and the other a document ranking. Nor do retrieval hits determine whether the reader can use what was returned. Answer accuracy combines what the retriever found with how much relevant context the reader received and how well it reasoned over that context. So the higher whole-document answer rate cannot be attributed to retrieval quality alone.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsWhy might the returned unit matter?
The author’s practical framing is a familiar codebase question: “where is the real business logic here, and what can I safely throw away?” A chunking strategy can omit a needed relationship when a function, its caller, a type definition, or explanatory prose falls outside the selected snippets. A full file may preserve more of those connections, but it can also bring irrelevant material along. This is a plausible explanation for the code result, not a cause isolated by the test.
One clue is the oracle arm: it averaged fewer context characters than the whole-document arm while scoring much higher. That pattern supports the idea that having the relevant context complete may matter more than simply supplying more text. It does not prove completeness is the sole cause. The comparison between chunks and documents also changes the size and composition of the returned context; the experiment did not include a gold-span-sized excerpt control to separate the effect of context unit from context volume.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How strong is the evidence?
Read this as a pilot that can suggest a useful design question, not settle the chunk-versus-document choice for retrieval systems generally. The author reports these important limits:
Rank #4
- 【Ryzen 5 3500U Processor】The BOSGAME mini pc is driven by the Ryzen 5 3500U (4C/8T, up to 3.7GHz) , with integrated Radeon Vega 8 Graphics, delivering reliable power, 4K video streaming and multitasking. Handle daily workloads like spreadsheet calculations, web browsing, and HD video editing effortlessly.
- 【8GB DDR4 & 256GB SATA SSD】E4 Air mini computers with 8GB DDR4 RAM and a 256GB SATA SSD, this mini desktop ensures quick app launches and efficient multitasking. while the SSD accelerates file transfers—ideal for office documents, media storage, and everyday computing.
- 【4K Triple Display & USB-C & USB3.2】The mini desktop computer Drives three 4K monitors via HDMI, DisplayPort and USB-C for multi-window productivity or immersive home theater setups；USB 3.2 meets your multi-interface transfer needs.
- 【Dual RJ45 LAN & Wi-Fi 5 & BT5.0】Equipped with Dual Gigabit Ethernet, dual-band Wi-Fi 5, and Bluetooth 5.0, this ryzen mini pc ensure stable connections for 4K streaming, video calls, and file transfers. Wirelessly connect keyboards, headphones and speakers via BT5.0 ideal for office productivity and home entertainment.
- 【3-Year Reliable Customer Services】 All of our BOSGAME mini pc gaming have FCC, ROHS, CE certifications. BOSGAME enjoy a 1-year wa-rranty for the entire machine and a 3-year wa-rranty for parts, ensuring your long-term peace of mind. If you have any questions about your purchase, please let us know through Amazon.
- Small query set: 16 frozen queries total, including only eight prose queries. The author describes the percentages as directional rather than absolute effects.
- Snapshot, not hard freeze: the index was snapshotted, but drift between runs was documented and not eliminated.
- No token-cost accounting for F5: context character counts were recorded; the article does not report token costs for the four-arm test.
- No noise or frontier-model comparison: the article does not report a judged-noise measurement or a rerun with a frontier reader model.
- Unit and volume remain entangled: a comparison of whole documents with excerpts sized around the gold span would be needed to better distinguish those effects.
These qualifications matter because a result on a particular codebase, query set, index snapshot, and model pair may not carry over to a different repository or workload. The author-reported numbers establish what happened in this run; they do not establish a universal retrieval policy.
What does the separate NodeRAG test add?
It adds a different, limited comparison—not another arm of F5. In a separate no-LLM rig on 10 rule queries, chunked TF-IDF retrieval hit 8/10, while graph BFS hit 7/10. The reported token estimates were 301,981 for TF-IDF and 170,140 for graph BFS. Those estimates used a harness conversion of word count multiplied by 1.3, not measured model-token usage. Graph traversal therefore had the lower estimate in this run but missed more rule queries; these figures do not show graph traversal winning overall.
What should retrieval-system designers take from it?
The most defensible takeaway is to test the returned unit against the material and questions your system actually serves. In this pilot, whole-document context substantially outperformed top-10 chunks for code answers, while the prose results leaned the other way. That is a reason to measure code and prose separately and inspect failure cases—not to replace chunking everywhere with full-file returns.
The article mentions Anthropic’s Contextual Retrieval and Jina’s Late Chunking as related approaches, but neither was an arm in this four-way test. They are background, not evidence that either method would reproduce these results. The source is Mikhail’s author-reported September 2026 article; no independent replication is established here.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
