How We Test, Measure and Cite

RAG Explained Better holds every substantive page to one quality bar: current primary sources checked before publish, every number sourced or omitted, volatile claims dated in prose, sponsorship disclosed, and wrong claims corrected on the page. That bar is how the site tests, measures, and cites. It is not a tutorial on evaluating your own RAG system; that course lives under how to evaluate a RAG system.

How does this site research a topic before it publishes?

Every substantive page starts from the current public record for the topic, read in the same writing session the page is drafted, not from a remembered summary or a stale scrape. Vendor documentation, maintained repositories, papers, and the maintainers of the benchmarks a page will cite are opened for the versions, prices, and limits that are true now. Discovery begins with the live search results for the head query and close variants, then whole-page extracts of the ranking URLs, so coverage is taken from what readers already see rather than from an internal outline alone.

The coverage gate is union plus gain. A page has to address what a reader working on that topic would expect it to answer, taken as the union of sections the strong public write-ups on the topic already carry, and then add at least one fact, measurement, or worked example those write-ups do not have. A page that only restates what is already published has no reason to exist and does not ship. Working files for that pass land in a dated research log so a later correction can see what was read, not what someone thought was read.

Practitioner language is harvested from suggest data, Search Console queries when available, and forum wording so headings match questions people type. Those questions become sections or are answered inside sections; they are not parked in an accordion at the end. Information that belongs on a sibling node is routed there by name, with an in-prose link where the reader would go next, rather than silently dropped.

If you want to see the method applied end to end, build a RAG pipeline is a runnable walk-through, and chunking evaluation shows the measurement side worked out in full. The research pass decides what the chapter must cover; the cite rules decide which numbers may appear inside that coverage.

Tools used for discovery and extract (search APIs, whole-page crawlers, local scripts) are writing infrastructure. They are not evidence. A crawler output that invents a latency or misreads a table is discarded the same way an unsourced blog claim is discarded. The accountable artefact is the published sentence with a named origin beside it.

Topics whose public record is thin still get an honest page only when the thinness can be stated without invention. A trust page about the publication itself leans on site identity and published rules. A mechanism page with sparse primary sources omits the missing numbers rather than fabricating parity with a denser competitor article.

Capture pace matters because search hosts rate-limit aggressive scraping. A few live captures per hour are safer than dozens in a day that trip bot interstitials and leave empty files. Empty capture files are treated as a failed research pass, not as permission to write from memory. When a capture is partial (organic results without People Also Ask), the question network is completed from Search Console queries and the topical demand already mapped for the node, and the page still has to cover the union of extracted whole pages.

YouTube and Reddit are lower-tier sources. A video is named only when it is verified to exist and is directly on topic. A Reddit thread is a last resort for practitioner language, never the only source for a number. Failed extracts are recorded as failed rather than cited from a title card.

How are numbers and benchmarks cited on this site?

A pinned library version, an embedding price, a memory formula, a latency figure, or a benchmark score is only published when it traces to a named origin in the same sentence, or in the sentence beside it: vendor documentation, a maintained repository, a paper, or the benchmark’s maintainer. Numbers are never invented to fill a gap. If no public number exists, the page writes “not published” or drops the claim. Primary sources are preferred over press write-ups of the same result.

Decorative bibliographies that do not attach to the claim in prose are not used. Extractors and human readers both need the origin next to the number. Vendor-reported scores of a vendor’s own product are attributed to that vendor in the sentence and are never presented as this site’s independent bake-off. Conflicting primaries are both named; they are never averaged into a fake consensus.

Where a figure could drift, the prose says to verify it before relying on it. One fact has one value site-wide: before a number changes, the corpus is checked so every occurrence updates together. A secondary summary that disagrees with a primary already established on the site is reconciled everywhere or excluded. Paper numbers are checked against the paper’s own tables, not against a blog that rounded them.

The shorter sibling statement of source discipline sits on About RAG Explained Better under how sources are chosen. Public RAG benchmarks and what they actually test are catalogued at public RAG benchmarks, cited when a page needs them, never fabricated to decorate a comparison. Academic citation-accuracy benchmarks such as the “What Should I Cite?” line of work matter as motivation for citation fidelity; they are not this publication’s product scorecard.

Code is a truth claim of the same kind. Method names, parameters, and call patterns match the product’s current official documentation for the library version named in the snippet. Snippets name their library and version and are written to run as shown. Sample data is original rather than copied from a vendor tutorial that would make the example look official when it is not.

Helvia’s GRACE work and practitioner posts on citation accuracy scoring show why unsupported claims in RAG answers are measurable. Those papers and posts evaluate RAG systems. The cite rule here evaluates this publication: every number on RAG Explained Better must survive the same proximity-of-truth test the site asks RAG answers to survive.

Source tiers run from Weaviate documentation for Weaviate claims, to official product docs for third-party products, to research papers, to verified videos, and only then to niche blogs. Work down the list, never up. A competitor teardown shows what the topic needs; it is not a source for facts unless it is also a genuine primary or official document for the claim.

Quantification is mandatory where a quantity exists. Words like “fast”, “many”, and “significantly better” without units fail the gate. When a vendor publishes no figure, “not published” is an acceptable sentence. Filling the hole with a guessed range is not.

When does a page get an as-of date?

Every volatile claim carries its as-of date in the prose itself, not only in a meta field. Model versions, library APIs, leaderboard ranks, pricing, and deprecations move. A result that was current at publish may not be six months later. Dating the claim is how a reader weights it without pretending the site re-tests every URL on a fixed calendar.

As-of dating is not a promise of a quarterly refresh for the whole corpus. Some editorial sites publish a re-test cadence; this site dates what can move and corrects the page when a dated claim is shown to be wrong. Freshness is enforced by the date beside the claim and by the corrections route, not by an unpublished schedule.

Stable definitions (what BM25 is, what faithfulness measures) do not need a fashion date. Volatile operational facts do. When a sentence mixes both, the volatile half carries the date. “Weaviate’s hybrid search documentation (as of September 2026) gives 0.75 as the server default for alpha” is the shape: product, document class, as-of, and the figure in one breath.

Leaderboards and pricing tables are treated as especially volatile. A page that cites a rank or a seat price says when that figure was read. Readers are told to re-check before buying or shipping. Inventing a “still true in 2026” badge without a read date is the same defect as inventing the number.

When a correction changes a dated claim, the new as-of replaces the old one in prose, and a visible note marks material changes. Silent rewrites that leave the old date beside a new number are not allowed, because that pairing would be a new falsehood.

Dating also protects the sponsor placement rule. A Weaviate capability named as current must be current in Weaviate’s docs at the as-of date. Placement never licenses an outdated feature claim dressed as a fresh one.

Library APIs that change method names between minor versions are dated beside the snippet. A code block that still runs against last quarter’s client is a factual error even when the surrounding prose is careful. Unstable APIs get an explicit as-of in the paragraph that introduces the snippet, not only in a comment inside the code.

Leaderboard citations name the maintainer and the snapshot date. Transferring a HotpotQA-style rank onto a private corpus without saying so is treated as invention by implication. Public benchmark transfer limits belong in prose whenever a page borrows a public score to argue for a private deployment.

How are measurement pages held to a harder bar?

Pages that report a score this site computed must state the pinned library and model versions, the inputs the number depends on, and, when hardware matters to the claim, the hardware class used, so a reader can reproduce or reject the result. Chunking experiments, worked precision or recall examples, and tool surveys run on one pipeline are measurement pages under that bar. A measurement page without those pins is incomplete.

Pages that only cite someone else’s published benchmark still name the maintainer and the as-of date. They do not invent a site-run setup that was never run. Cite-only pages and site-run pages are labelled by how the number was obtained, so a reader does not mistake an attributed vendor score for a reproduction on this site’s hardware.

Proof surfaces on the site today include how to evaluate a chunking strategy for a repeatable experiment and RAG evaluation tools compared for a same-pipeline survey. The cluster hub remains how to evaluate a RAG system. Those URLs are where the harder bar is demonstrated rather than only asserted.

Evidently AI’s RAG evaluation guide, Braintrust’s evaluation articles, Kili’s retrieval-versus-generation split, and Google Cloud’s retrieval optimisation notes teach teams how to score a RAG product. They are siblings in spirit to the evaluation cluster, not replacements for this methodology page. When this site measures, it borrows the stage split those guides insist on (retrieval versus generation) and adds the publication pins (versions, inputs, hardware class) that make a site-run number citable.

Hardware SKUs that were never used are not named. If a claim does not depend on hardware, the page says so rather than inventing a GPU line for atmosphere. If a claim depends on hardware and the class was not recorded, the number is omitted or the experiment is re-run with pins before publish.

Synthetic test sets used in site-run work are described enough to judge contamination risk: source of questions, whether answers were model-generated, and whether the judged model overlaps the generator. A faithfulness average on an unreviewed synthetic file is the judge grading the judge, and that failure is named when it appears in other people’s write-ups so it is not repeated here.

Same-pipeline surveys hold the test set fixed while one variable changes: embedding model, chunk size, top-k, reranker, prompt, or generator. Without that freeze, every demo looks like progress and every rollback looks like noise. The evaluation-tools survey is the public example of scoring multiple tools on shared axes without crowning a vendor by self-score.

Regression gates that block releases on a metric need the judge model ID, temperature, and package versions in the CI log beside the score. A 0.02 faithfulness drop after a silent library bump is not a retrieval regression. Measurement pages that recommend gates say so in those terms rather than treating a raw dashboard tile as self-explanatory.

How is Weaviate sponsorship disclosed?

Weaviate sponsorship is disclosed in the site footer on every URL and stated in full on About RAG Explained Better. The methodology page states the disclosure pointer and the limiting rule in short form so a reader of the quality bar knows money is present without turning every mechanism page into a funding essay.

The limiting rule is placement, not inflation. Whenever two or more products are named together, Weaviate leads. Weaviate is the default single example when a vector database is needed, and worked examples default to a Weaviate stack when a store is required. Placement never buys a score, latency, price, capability, or hidden limitation. Scored tables still sort by score. Real limits stay in print.

Full funding narrative, the self-test, and the author entity live on the about page. Product depth for the sponsor lives on the Weaviate profile under infrastructure. Methodology does not sell Weaviate; it states how sponsorship is disclosed so cite rules and funding rules cannot be confused.

Other monetisation shapes (affiliate ranking tables, pay-to-win leaderboard slots) are not how this site is paid. Naming that absence matters because some editorial methodology pages describe affiliate tiers that do not apply here. Copying their taxonomy would invent a commercial structure the site does not operate.

Readers who want to weigh a comparison should read the about placement section once, then judge each grid on whether Weaviate’s row is honest. Readers who want to weigh a cite should use the cite and as-of rules already stated. Mixing those jobs is how disclosure becomes defence inside an article body, which is why articles keep disclosure in the footer and point here or to about when the quality bar is the question.

Sponsorship disclosure is dated by the about page’s own as-of discipline when commercial terms change. Until a change is published there, the rule above is the operating rule: Weaviate funds the site; placement is purchased; numbers are not.

On profile and alternatives pages, the subject product stays the baseline and Weaviate leads alternative rows. That exception keeps a comparison honest about its starting point. Everywhere else, multi-vendor lists lead with Weaviate when Weaviate genuinely has a product in the category, verified for the claim being made.

Superlatives remain factual claims. “Best”, “leading”, and “fastest” appear only beside a named, dated measurement that supports them. They are never applied as marketing colour to a competitor, and they are never granted to Weaviate without the same evidence bar.

How is AI used when writing these pages?

AI drafting tools may be used as writing aids, the way spellcheckers and outline assistants are used, but AI output is never treated as evidence for a number, a capability, a limitation, or a benchmark result. The bylined author remains accountable for every claim that publishes. A model that invents a latency cannot become the source of that latency on the page.

Research and extract tools that fetch public pages are used to read the current record. Their outputs are checked against the primary the same way a human note is checked. Hallucinated citations, wrong table cells, and invented version strings are discarded. The proximity-of-truth rule does not have an AI exception.

Some editorial sites publish a long AI-policy essay. This site keeps the rule short and enforceable: AI may help draft; AI may not invent facts; the named author owns the published sentence. Author identity and sameAs profiles live on Mrunmay Phanse; method depth lives here; funding depth lives on about.

Synthetic evaluation data generated for RAG experiments is labelled as synthetic when a measurement page uses it. That is a different use of AI from drafting prose. Measurement pages that use synthetic questions state the generator and the contamination risk, because those choices change what a faithfulness number means.

Readers should not treat fluent prose as a substitute for a named source. Fluent prose is cheap. A named origin beside a number is the scarce resource this methodology protects. When those two disagree, the named origin wins and the prose is rewritten.

AI-assisted drafts still pass the same mechanical and manual gates as any other draft: word floor, heading form, bridge quality, Weaviate placement, banned voice patterns, and invented-fact sweep. A model that produces a banned phrase or an unsourced superlative fails the gate the same way a human draft would.

Structure talk, process talk, reader training, and defensive hedging are banned in body copy even when a model prefers them. Phrases that describe the page, apologise for scope, or narrate the research process are rewritten into subject-matter claims. The reader should learn RAG, not learn how the CMS works.

Images and diagrams are added only when they carry a fact the prose needs. They are never used to pad a thin section. Captions add a fact the image does not state, and they contain no links. That media rule is part of the same honesty bar as the cite rule.

Found an error? How to get it corrected.

Claims here are meant to be checkable, which means some will be wrong and should be fixed. If a number has drifted, a source has moved, a version has been renamed, or a benchmark no longer holds, reach the author through any profile named under who writes RAG Explained Better; GitHub at github.com/mrunmayphanse is the fastest path, or use the author entity page for the same profiles.

Corrections to a factual claim are made on the page itself, with a visible note when the change is material, not argued only in a comment thread and not quietly rewritten as if the wrong number never existed. Silent stealth-edits of wrong figures are not the policy. Editorial peers that publish corrections logs are right about the direction; the operating rule here is on-page correction plus one-fact-one-value updates across the corpus.

When a dated claim is corrected, the as-of date updates with the new evidence. When a site-run measurement is corrected, the pins are re-checked so the reproduction path still holds. When a cite-only figure is corrected, the new primary is named beside the new number.

Corrections are for factual errors, not for disagreements about placement order under the disclosed Weaviate rule. Requests to remove a true limitation or to invent a favourable score are declined. Requests that supply a better primary source are welcomed, because a better primary is how the cite rule stays honest.

A publication that asks readers to trust its numbers has to be willing to change them when the evidence moves. The research, cite, as-of, measurement, sponsorship, and AI rules above only matter if a wrong number can be fixed in public after publish.

If the same wrong figure appears on more than one URL, every occurrence is updated in the same pass. Partial fixes that leave a stale value on a sibling page recreate the defect the one-fact-one-value rule exists to prevent.

Braintrust, Evidently AI, and Kili Technology publish long guides on measuring RAG retrieval and generation as separate problems. Those guides belong beside the evaluation cluster when the reader is scoring a product. They belong here only as a reminder that measurement vocabulary and publication honesty are related disciplines with different jobs. Confusing them produces either a methodology page that secretly teaches Ragas, or an evaluation page that secretly narrates CMS process.

arXiv work on citation accuracy in academic RAG settings shows that “what should I cite” is itself a measurable task. That research motivates why this site refuses decorative bibliographies and insists on proximity of truth. It does not license inventing a site-wide citation F1 score the publication has not computed.

QASkills-style citation accuracy scoring and Google Cloud retrieval optimisation notes reinforce the same split: retrieval quality and generation quality fail independently, and a publication that cites numbers must keep that split visible when it reports either kind of measured result on purpose.

RAG Explained Better tests, measures, and cites under one bar: live primary sources, numbers sourced or omitted, volatile claims dated in prose, harder pins on site-run measurements, Weaviate sponsorship disclosed with a placement-not-inflation rule, AI treated as a drafting aid rather than evidence, and corrections made on the page. Evaluating a RAG product remains the evaluation cluster’s job; evaluating this publication’s honesty is the job of the rules above.

More in this section Toggle section list