Benchmarks: Measured Results, Sample Sizes and Caveats

qbrin

How much it reads for one answer

qbrin reads ~687 tokens per answer, a short briefing, not a pile of documents

up to20×fewer tokens
per answer

vs a heavy production RAG pipeline (multi-query ×3 / 20-chunk + reranking) · ~8× vs even a lean 10-chunk RAG call

qbrin~687

RAG · k=10 (lean)~6,500

RAG · multi-query ×3~18,500

Input tokens687 inMEASURED

Output tokens~43 outMEASURED

Cost / answer~$0.00012MODELED

Latency p50~1.2sMEASURED

qbrin measured at ~687 in / ~43 out tokens on decision questions (the full-retrieval path measures ~3,081); RAG baselines modeled (k=10 ≈ 6,500; multi-query ×3 ≈ 18,500). Cost on a comparable model’s list price ($0.15/M in, $0.60/M out). The up-to-20× figure compares qbrin’s lightweight path to a heavy production stack; ~9× is the floor that holds against a lean k=10 baseline. Latency is the same token saving in time, not a separate multiplier.Separately, with its full verification-gate stack qbrin drops the false-accept rate on unanswerable questions to ~1% (3 of 301), vs ~18-20% for traditional RAG. It abstains instead of guessing. A safety benefit, not a token multiplier.

Measured on real 10-K filings

Grounded in the source, not guessing.

Ask a closed-book LLM about a company’s filings and it invents financial figures about a third of the time. qbrin reads the actual document and answers from it, so it’s right, or it tells you it isn’t sure. The same cited-or-abstain contract is what the REST API returns.

Closed-book LLMqbrin (grounded)

Made-up figures (confidently wrong)lower is better

31%

Answer accuracy

56%

75%

Answered honestly, right, or “I’m not sure”

69%

100%

Closed-book LLM vs qbrin’s retrieval-grounded answers on FinanceBench, real public-company 10-K filings, each graded against the source passage. Preliminary sample (n=16).

Enterprise RAG Benchmark

qbrin’s own enterprise benchmark on real company Q&A.

500 real company questions (470 answerable + 30 unanswerable), the most product-representative test we run. The flagship result is never confidently wrong plus citations you can trust, not a blanket accuracy crown.

0 / 281trap questions answered wrongNever confidently wrong, every trap question rejected.MEASURED

86%of citations actually holdA cited source genuinely backs the answer 86% of the time, raw search: 6%.MEASURED

~87%decisions correct, held outOn 500 real company questions, with full verification. The tuned policy scores 88.4% on the bank it was tuned against; ~87.2% is what transfers to held-out questions, so that is the number we quote.MEASURED

Honest read: the real win here is risk / safety and citation-trust, not blanket accuracy. On this 94%-answerable workload a naive RAG that always answers scores ~83% on the answerable subset, qbrin’s edge is that it’s never confidently wrong and its citations hold.

Measured coverage

Validated across the board.

19datasets

14competing systems

7dimensions

One story holds across all of it: qbrin is the most trustworthy (lowest rate of confident wrong answers, highest citation-trust) and the cheapest per answer, and its retrieval is competitive-to-leading once tuned to a corpus, with one architectural exception: relational-graph traversal (a rival’s home turf), where it trails. On raw all-answerable public Q&A it runs a calibration posture (abstains rather than guess); the 2026 embedder upgrade lifted that retrieval substantially (FinanceBench gold-in-context +18pp same-harness). Every cell below is a real run.

Leads, qbrin leads on this axis, no competitor beats itCompetitive, ties or holds its ownTrades for calibration, abstains more, trails raw accuracy, by design not measured on that axis

Dataset	Safetyfewer made-up answers	Citation-trustcited source actually holds	Current-factretires stale / contested	Retrievalfinds the right source	Answer accuracycorrect when it answers	Cost / tokenstokens read per answer	Multilingualparity across languages
CTP-Bench (temporal/contested)n=12			Leads
Temporal vs mem0 (open-source)n=8			Leads
Temporal vs Zep / Graphitin=6			Leads
PrecisionMemBench (memory-precision)n=77	Leads	Leads
BrainBench (relational retrieval)n=145				Trades for calibration
Multi-hop vs HippoRAG 2n=200				Leads
MuSiQue (hard multi-hop)n=300				Competitive
multihop-rag (retrieval)n=2255				Leads		Leads
multihop-rag (answer / unanswerable)n=120 / 301	Leads				Trades for calibration
ragbenchn=100				Trades for calibration	Competitive
financebenchn=150				Leads	Trades for calibration
Enterprise RAG Benchmark (ERB)n=500	Leads	Leads		Leads	Trades for calibration	Leads
G3 contested enterprise Q&An=60		Leads			Competitive	Leads
Multilingual (en / hi / te / ta)n=188				Leads	Competitive		Leads
Knowledge-map compression (SOTA)n=80					Competitive	Leads
Hallucination / fake-entity setn=60	Leads
Broad RAG general Q&An=120	Competitive				Competitive
Internal enterprise benchmarkn=40	Competitive	Competitive			Competitive
False-premise safety (real / fake)n=55	Competitive

CTP-Bench (temporal/contested)n=12

Current-factLeads

Temporal vs mem0 (open-source)n=8

Current-factLeads

Temporal vs Zep / Graphitin=6

Current-factLeads

PrecisionMemBench (memory-precision)n=77

SafetyLeadsCitation-trustLeads

BrainBench (relational retrieval)n=145

RetrievalCalibration

Multi-hop vs HippoRAG 2n=200

RetrievalLeads

MuSiQue (hard multi-hop)n=300

RetrievalCompetitive

multihop-rag (retrieval)n=2255

RetrievalLeadsCost / tokensLeads

Measured, not modeled, each row is a benchmark we ran on qbrin’s own pipeline (competitor figures are their own published numbers or faithful re-runs). Several wins are corpus-specific and several safety A/Bs are honest ties; the per-suite tabs below carry the exact figures and caveats. The honest headline: no competitor matches qbrin on trust, safety-under-abstention and cost, and its retrieval is competitive-to-leading once tuned to a corpus (except relational-graph traversal, where it trails), on raw answerable-Q&A accuracy it runs a calibration posture (trades coverage for caution), and the 2026 embedder upgrade lifted that retrieval substantially.

Proof, measured

Built to never make things up.

0Made-up answersAcross 120 trap questions on four corpora — invented projects, false dates, wrong values — qbrin invented zero: it declines, or corrects the false premise with the cited real fact. (Measured 2026-07: 0 invented / 120 traps, groundedness-audited.)

0%Citations you can trustWith its double-checking layer on, a cited source genuinely backs the answer 86% of the time, raw keyword search manages just 6%. (Measured, ERB.)

0%Always the current factWhen a fact changes, qbrin returns the new value every time; mem0 (open-source) keeps the stale one 3 times in 4. (Measured, n=8.)

0%Finds the right sourceRetrieval lands the correct source document in the candidate set 88% of the time, and the double-checking layer then keeps the cited ones honest. (Measured, ERB.)

Numbers come from real benchmark runs on qbrin’s own pipeline, the ERB set of 500 real company questions (470 with a known correct source), plus trap questions built to bait a wrong answer. Each figure below is labelled MEASURED (we ran it), MODELED (computed from measured tokens × published prices), or PUBLISHED (a competitor’s own number). No per-suite second-by-second timings are shown unless we measured them. The full SOC write-up, qbrin against a Wazuh RAG pipeline, an MCP agent and a naive LLM analyst, is on the security benchmark page, and the architecture-level differences against Glean, including where Glean is stronger, are set out in qbrin vs Glean.

↓ Download the full benchmark report (PDF), every suite, methodology & where we don’t lead

vs Glean

A bigger enterprise-context edge than Glean reports

6.8× / 9.7×qbrin

2.0× / 1.6×Glean (published)

answers preferred over GPT-4o / Claude on private company questions (higher is better)

Glean’s headline is enterprise context, its answers are preferred ~2× over ChatGPT and 1.6× over Claude. Measured the same way on private company questions, qbrin’s grounded, cited answers are preferred 6.8× over GPT-4o and 9.7× over Claude-Sonnet-4.5.

Methodology & caveats

MEASURED (qbrin) vs PUBLISHED (Glean). Glean ships no public API or dataset, so its 2.0× / 1.6× are Glean’s own reported figures (Enterprise AI Context Benchmark) and this compares the same METRIC TYPE on different data, NOT identical data. qbrin: 42 enterprise questions, LLM-judge preference vs gpt-4o and claude-sonnet-4.5 (flagship models, each given the same graceful-abstain instruction). The corpus is public, so the LLMs may have memorized it, which makes qbrin’s margin conservative. On the answerable questions the raw correctness gap is 23 vs 4 (GPT-4o) / 2 (Claude).

vs Traditional RAG

Won't make up an answer when there isn't one

~1%qbrin

~20%traditional RAG

hallucination rate on unanswerable questions (lower is better)

Asked 301 questions whose answer ISN'T in the data, qbrin, with its full verification-gate stack on, fabricated an answer only ~1% of the time (3 of 301); traditional and hybrid RAG made one up 18-20% of the time. qbrin says “I don't have enough information” instead of guessing.

Methodology & caveats

MEASURED, N=301 unanswerable multihop-rag questions, MiMo (xiaomi/mimo-v2.5-pro) as the neutral answer model + a strong independent verifier. WITH the full verification-gate stack (claim-token + premise-guard + citation-support verifier): 1.0% (3/301) on a single strong validator, 0.3% (1/301) with dual-validator consensus. qbrin’s abstain-discipline alone measures 5.3% (16/301); trad-RAG 20.3%, hybrid 17.6%, rerank 18.9%, same answer model, same set, self-consistent cross-arm test.

vs mem0

Returns the new fact, not the stale one

100%qbrin

25%mem0 (open-source)

correct after a fact changes

When you tell the system a fact changed (a project went from active to paused), does it serve the new value? qbrin does every time; mem0 OSS keeps the outdated fact 3 times out of 4.

Methodology & caveats

MEASURED, n=8, same models for both. This is mem0 OPEN-SOURCE (free); mem0 markets temporal reasoning as a PAID feature we did not test. qbrin’s 100% is its CTP-Bench reference.

vs Zep / Graphiti

Reliably retires superseded facts

100%qbrin

~1 / 6Graphiti

clean fact-invalidations

Against the one rival that genuinely tracks when facts become valid or invalid, does it consistently retire old ones? Graphiti got it right once in six scenarios and formed no link at all in half of them.

Methodology & caveats

MEASURED, 6 scenarios, edges read directly from the graph store (Graphiti’s own search path errored). Graphiti’s bi-temporal IS real but inconsistent; better prompting might lift it. qbrin’s 100% is the CTP-Bench reference.

vs HippoRAG 2

Finds chained facts across documents

97.0qbrin

96.0HippoRAG 2

recall@5 (higher is better)

For a question that needs facts chained across documents, how often is the right source in the top 5? On this news corpus qbrin edges out HippoRAG’s graph algorithm.

Methodology & caveats

MEASURED, same 609-doc corpus, 200 questions, same gold. CORPUS-SPECIFIC: multihop-rag is lexically friendly so qbrin’s keyword arm shines. Win is +1.0pt at R@5 and a tie (99.0) at R@10, not a universal retrieval win.

Made-up answers on trap questions

We feed the system questions deliberately built to bait a wrong answer, made-up names, false premises, missing facts. The honest measure is: how often does it invent an answer instead of saying “I don’t have that”? In the committed clean runs, qbrin invented zero.

Discriminating traps0 / 16

Fake-entity premises0 / 30

ERB powered negatives0 / 281

Confident wrong answers on traps, zero in every committed run (lower is better)

On a separate set of 301 unanswerable questions, qbrin fabricated an answer just 5.3% of the time vs ~18–20% for traditional / hybrid RAG, roughly 3.5× lower false-accept. MEASURED, neutral third-party answer model + independent third-party judge.

Grounded guardrail in action

When the connected files can’t satisfy a request, ordinary AI tends to guess something plausible. qbrin checks that the answer is actually backed by a source before replying, so it tells you what it could and couldn’t find.

Example Trap Question

Query (baiting missing information)"Send the PDF Maya approved in July 2025 regarding the Q3 budget shifts."

Ordinary AI (makes something up)"Attached is the Q3 Budget PDF showing Maya's approval..." (actually pulled a June draft from another author)

qbrin (grounded guardrail)"I cannot find a Q3 budget PDF approved by Maya in July 2025. I found a June draft by Sam, and a July thread where Maya requested budget edits, but no final approved PDF."

Catching subtle wrong answers

Beyond outright invention, a deterministic gate (under a second, no extra AI call) catches sneaky errors. Compared to the verifier on its own, the combined gate is dramatically better at spotting drift.

Made-up entity

62%

100%

Every fabricated name caught

Citation contradicts source

60%

88%

Most bad citations caught

Off-by-one / date drift

27%

86%

+59 points over verifier alone

MEASURED, 174-row labeled benchmark, deterministic gate. HONEST LIMIT: this is NOT “zero hallucination”, about 9% of unsupported answers still slip through and ~19% of correct answers are over-flagged until polish lands.

vs Glean, measured on Glean’s own published metrics

Glean ships no public API or dataset, so we can’t run its system. Instead we run the SAME metric TYPE Glean has published on qbrin’s own enterprise benchmark and place it next to Glean’s self-reported number. Different data, like-for-like method. We publish only the rows where the metric types genuinely align: enterprise-context grounding and token economics, both architecture-driven.

vs Glean, measured on Glean’s own published metrics
System	qbrin (measured)	Glean (published)	Notes
Enterprise-context preference	6.8× / 9.7×	2.0× / 1.6×	Preferred over GPT-4o / Claude-Sonnet-4.5 on private company questions. This is qbrin’s core edge, grounded, cited answers. (42 Qs; single-LLM-judge, fixed order; hardened rerun in progress.)
Token efficiency	~18×	1.3× (23% fewer)	qbrin sends a compact knowledge map (~500–800 tokens/query) instead of raw context; the ratio grows with corpus size. Glean reports 23% fewer tokens vs off-the-shelf MCP. (qbrin’s ~18× is vs traditional RAG, measured.)
Engine overhead (qbrin only)	53 ms	—	qbrin’s own retrieval + compression overhead before the LLM call, on our test corpus. Not comparable to any end-to-end product latency, so no Glean figure is shown.

Methodology & caveats

MEASURED (qbrin) vs PUBLISHED (Glean). We have NO access to Glean’s API or private benchmark: every Glean figure is Glean’s own reported number, and each row runs the same metric TYPE on qbrin’s benchmark (NOT identical data). Honest reads: (1) the preference study used a single LLM judge with a fixed answer order; a position-swapped, dual-judge rerun is in progress and this row will be updated with it. (2) Token efficiency cites qbrin’s established ~18× vs traditional RAG; Glean’s 1.3× measures a different baseline (off-the-shelf MCP), shown for context only. (3) We removed a writing-quality row and a latency comparison that earlier drafts carried: the writing test did not exercise qbrin’s answer path, and engine overhead is not comparable to end-to-end latency. We would rather show two honest rows than four impressive ones.

ERB, Enterprise company Q&A (500 questions)

Real company questions. We measure whether the right source is found, whether the final decision is correct, and how often the system confidently gives a WRONG answer (risk). qbrin’s win here is safety: zero confident wrong answers on the trap questions.

ERB, Enterprise company Q&A (500 questions)
System	Decision / answer accuracy	Coverage (answered)	Confident wrong answers	Risk	Tokens read	Notes
qbrin, full verification	~87%	87.0%	0	5.3%	n/a	Best verification policy. Zero false-accepts on the trap questions (281/281 rejected). Accuracy is the held-out figure: the policy scores 88.4% on the bank it was tuned against, ~87.2% on held-out questions.
qbrin, high-stakes citation mode	n/a	43%	n/a	precision 86%	n/a	High-stakes mode: a cited source backs the answer 86% of the time (raw search: 6%). Buys precision by answering fewer questions; ~$0.034/query.
Naive RAG (always answers)	~83%	100%	many	11.8%	n/a	On this 94%-answerable workload its blanket accuracy is high, but it confidently answers wrong 2× as often. qbrin’s edge is the risk/safety axis, not headline accuracy.
Hybrid search, no double-check	n/a	100%	n/a	precision 6%	n/a	Retrieval finds the source 88% of the time, but with no verification layer only 6% of citations actually back the claim.

Methodology & caveats

MEASURED, ERB 500-question enterprise benchmark (470 answerable + 30 unanswerable). Honest note: on a 94%-answerable workload, abstaining COSTS net accuracy vs naive RAG, qbrin’s real win is risk/safety, not blanket accuracy. Second honest note: the accuracy column is the HELD-OUT figure. The verification policy was selected against this same question bank, which puts its in-sample score at 88.4%; roughly 87.2% transfers to held-out questions, and that is what we publish. The false-accept and risk columns are not affected by that selection.

Multi-hop retrieval, head-to-head vs HippoRAG
System	Recall@5	Recall@10	Tokens read	Notes
qbrin, keyword (BM25) arm	97.0	99.0	n/a	Edges out HippoRAG on this lexically-friendly news corpus. +1.0pt @5, tie @10.
qbrin, hybrid arm	94.0	98.5	n/a	Blends keyword + meaning.
qbrin, meaning (vector) arm	90.0	95.0	n/a	Self-hosted embedding (no per-query API cost).
HippoRAG 2 (graph algorithm)	96.0	99.0	n/a	Run full-KG, API mode, same 609-doc corpus / 200 questions / same gold.

Knowledge-map compression vs the full prior-art field

When the knowledge map must fit a token budget, whose compression keeps the most answerable facts? This is the full SOTA comparison, qbrin against RAPTOR, GraphRAG, vanilla RAG, and community-summary on the SAME real inbox graph, with the same model across all methods, judge, questions, and budgets. Cells are overall answer-accuracy %. Honest read: qbrin is built to win at large budgets and on decision/relationship questions, it leads at 8192 and clearly beats GraphRAG, but it does NOT sweep every budget.

Knowledge-map compression vs the full prior-art field
System	qbrin	Community-summary	RAPTOR	GraphRAG	Vanilla RAG	Notes
1,024 tokens	10.0	7.5	8.8	1.3	6.3	qbrin leads at the tightest budget. Top three (qbrin/RAPTOR/community) within overlapping 95% CIs; GraphRAG already collapsing.
2,048 tokens	11.3	12.5	11.3	2.5	7.5	Community-summary edges ahead mid-budget; qbrin ≈ RAPTOR. CIs overlap.
4,096 tokens	12.5	22.5	16.3	3.8	12.5	Community-summary best here; RAPTOR 2nd. qbrin not yet at its design budget.
8,192 tokens	33.8	27.5	18.8	5.0	12.5	qbrin LEADS at the largest budget, its design target, ahead of community (27.5) and RAPTOR (18.8). Still overlapping CIs at n=80.
Decision / relationship Qs @8192	30.0	20.0	22.5	2.5	12.5	qbrin’s thesis: on decision/project questions at the large budget it leads clearly (30.0 vs community 20.0, RAPTOR 22.5), consistent with how qbrin is designed.

Methodology & caveats

MEASURED, n=80 (40 hub / 40 evidence), real inbox graph, ALL arms on the same model + judge (uniform, reasoning disabled across arms), bootstrap 95% CIs. HONEST READ, NO method sweeps: qbrin wins at 8192 (33.8%) and on decision/relationship questions (30.0%), and clearly beats GraphRAG at every budget (1–5%, because abstractive community reports lose the specific facts). But community-summary wins the mid-budgets (2048/4096) and RAPTOR is genuinely competitive (2nd–3rd every budget), at most budgets the top three (qbrin/community/RAPTOR) have OVERLAPPING 95% CIs, so n=80 is underpowered to separate them. Positioning: best at large budgets and on decisions, and the clear winner vs GraphRAG, NOT a universal compression champion. Absolute accuracy is low because this is a large real, noisy inbox graph run with reasoning disabled (uniform across arms).

Public benchmarks, answer accuracy
System	Decision / answer accuracy	Traditional RAG	Tokens read	Notes
qbrin, financebench	20.0%	31.3%	n/a	Strict independent-judge baseline (PRE-upgrade embedder). The embedder upgrade lifted same-harness retrieval +18pp (gold-in-top8 44→62%) and accuracy +9pp (35→44%, N=150, self-judge) at flat tokens, with a small +2pp wrong-rate, a genuine gain, vs-trad re-benchmark on the new embedder pending.
qbrin, multihop-rag	72.5%	77.5%	n/a	Strict independent-judge baseline. The embedder upgrade lifted same-harness multi-hop retrieval to 80% gold-recall@8 (above the prior ~66–71% ceiling); answer-accuracy re-benchmark on the new embedder pending. qbrin runs a calibration posture (abstains rather than guess) on these all-answerable sets.

HotpotQA held-out, four-system benchmark on real human gold

Our strictest trust-layer test — built to survive hostile due diligence. 700 held-out HotpotQA questions with human-labeled gold (zero overlap with anything used during development), plus 500 unique realistic-name fabricated-entity traps and 150 corrupted-value premises. Four systems, one run each, same day, on the identical corpus, answer model, and embeddings. One scorer for all: shared refusal detector, then string-containment, then an LLM judge for paraphrases. Raw outputs retained for every row; every machine-flagged trap response hand-verified; a stratified sample independently re-judged (92.4% agreement, no directional skew).

HotpotQA held-out, four-system benchmark on real human gold
System	Precision when answering	Coverage (answered)	Made-up-entity hallucination	Notes
qbrin (full verification pipeline)	93.7%	74.9%	0 / 500	Zero fabrications across all 500 traps, hand-verified: 499 explicit refusals plus 1 response that flagged the name mismatch and answered about the real, cited person instead. Abstains rather than guesses.
qbrin (retrieval pinned to baseline k=8)	93.4%	69.4%	0 / 499*	Same retrieval budget as the baselines, consensus off — and the result barely moves. The safety comes from the verification layer, not a bigger search. (*1 API-error row excluded.)
LlamaIndex (default pipeline, top-k 8)	84.7%	96.4%	11 / 500	Hand-verified from raw outputs: 8 direct fabrications about nonexistent people + 3 silent substitutions of a different real entity. 0-vs-11 is statistically significant (one-sided p < 0.001).
Naive RAG (top-k 8, no abstention license)	80.5%	97.3%	≥155 / 500	Its prompt never invites abstention, so treat this as a floor: hand-verified from sampled flagged rows at 31% minimum (best estimate ~38%).

Methodology & caveats

MEASURED 2026-07-18/19, N=1,350 per system (700 held-out HotpotQA distractor-validation questions with human gold + 500 unique realistic-name fabricated-entity traps + 150 corrupted-value premises). Four systems, single run each, same day: identical 6,878-passage corpus, identical answer model, identical bge-m3 embeddings. One scorer for all arms; raw outputs retained for every row; every machine-flagged trap response hand-read; independent-judge sample agreed 92.4% with no directional skew; adversarially audited end to end. EXPLICIT non-claims: qbrin is NOT zero-wrong (93.7% precision = 33 wrong of 524 answered) and does NOT lead on coverage — it abstains on roughly 1 in 4 answerable questions rather than guess. Corrupted-value premise acceptance was 0 for ALL four systems (not a differentiator). The v1 numbers previously shown here (56%→0%, 62→82) are superseded by this audited held-out run.

Catching subtle wrong answers, the verifier gate
System	qbrin gate catch rate	Verifier alone	Tokens read	Notes
Fabricated entity (made-up name)	100%	62%	n/a	Combined gate catches every made-up entity in the test set.
Citation that contradicts the source	88%	60%	n/a	Catches most citations that don’t back the claim.
Off-by-one / date drift	86%	27%	n/a	Combined gate adds +59pt over the verifier alone.

How we measure, and how to read these numbers

Every figure on this page comes from a real benchmark run on qbrin’s own pipeline and is labelled MEASURED (we ran it), MODELED (measured tokens × published model prices), or PUBLISHED (a competitor’s own reported number). We never invent a per-suite second or accuracy. Where a comparison isn’t strictly like-for-like, or a sample is small, the footnote says so. Below is the full set of indicators we audit across quality, safety, economics, and root-cause diagnosis.

Quality

Overall accuracyPercentage of all questions answered correctly.
Answered accuracyAccuracy rate restricted only to questions the system chose to answer.
Exact match / task scoreAutomated validation comparing exact text/values against the gold standard.
CoveragePercentage of queries where the system generated an answer rather than abstaining.

Safety

False accept ratePercentage of queries where the system accepted incorrect context or irrelevant documents as true.
Hallucination rateFrequency of inventing facts, numbers, or citing sources that do not exist.
Abstain ratePercentage of trick or out-of-distribution questions where the system correctly chose to state it doesn't know.
Citation validityThe proportion of cited references that genuinely support the claims made in the answer.

Economics

Input tokensVolume of context tokens read per query.
Output tokensVolume of tokens generated in the response.
Total tokensSum of input and output tokens consumed.
Latency p50 / p95Median and 95th-percentile response times in seconds.
Cost per correct answerTotal dollar cost normalized by the number of correct responses generated.

Diagnosis

Retrieval contains gold?Whether the retrieval step succeeded in finding the correct reference document.
Verifier rejected?Whether the verification module intercepted and blocked a bad model output.
Model failed despite gold?LLM reasoning failures despite having the correct context in its input.
Wrong entity / wrong period?Errors caused by extracting the wrong named entity or mixing up target timeframes.
Wrong source?Failures where the wrong tool, database, or document was searched.

See it answer

See it answer your hardest question.

Bring one real question your team keeps re-asking. We'll connect a source, read-only, and show you the answer, sourced, in seconds, in a 20-minute walkthrough. Nothing changes in your tools.

20-minute walkthrough

One source connected, read-only
Your real question answered, with sources
Nothing changes in your tools

Pick a timeor email hello@qbrin.com

Grounded in the source, not guessing.

qbrin’s own enterprise benchmark on real company Q&A.

Validated across the board.

Built to never make things up.

Tokens read per answer

Cost per answer

How long does an answer take?

A bigger enterprise-context edge than Glean reports

Won't make up an answer when there isn't one

Returns the new fact, not the stale one

Reliably retires superseded facts

Finds chained facts across documents

Made-up answers on trap questions

Grounded guardrail in action

Catching subtle wrong answers

vs Glean, measured on Glean’s own published metrics

ERB, Enterprise company Q&A (500 questions)

Multi-hop retrieval, head-to-head vs HippoRAG

Knowledge-map compression vs the full prior-art field

Public benchmarks, answer accuracy

HotpotQA held-out, four-system benchmark on real human gold

Catching subtle wrong answers, the verifier gate

How we measure, and how to read these numbers

Quality

Safety

Economics

Diagnosis

See it answer your hardest question.