Range observed across 892 data points spanning 6 public semantic domains, real SEC EDGAR filings and physical scale tests from 500K to 5M tokens. Savings vary significantly by workload: compact or low-density corpora see smaller reductions; large, multi-source, governed archives see the highest. All figures are evidence observations from our test programme, not guarantees.
Organisations — government, healthcare, finance, legal, defence — hold vast data repositories. When an LLM is asked a question, it shouldn't wade through millions of irrelevant tokens. It should receive only the precise, relevant data it needs.
Sending an entire large corpus may exceed model context, increase input cost and expose the model to avoidable noise. Conventional pipelines therefore have to retrieve, chunk or truncate before reasoning.
Your records go through the QCE first. It selects a budget-aware subset using declared importance, configuration and corpus signals, then returns full records for the downstream model. Realised reduction and quality are use-case dependent.
We built a reference application on top of the FDA's public adverse event database — millions of medical reports. A user queries: "aspirin bleeding elderly" and asks the LLM: "What are the risk factors?"
The reference application fetches an OpenFDA cohort and asks the selected LLM a user-supplied question. Large cohorts may require truncation or another retrieval layer before the model can process them.
The same model and question are run over QCE-selected, unmodified reports. The application displays both answers and token usage for inspection. This demonstrates integration and efficiency; it is not a clinical-validation claim.
Send documents, messages or structured records with token estimates and metadata. Current public requests are bounded; large-archive discovery and connectors remain a separate retrieval stage.
Configured variants use importance and corpus signals; advanced QCE variants also read the query, content, source, speaker and structural signals across the corpus. Quality remains task- and corpus-dependent.
We return the actual records — unmodified, unparaphrased. Your LLM gets the truth, not an interpretation of it.
Measure retained tokens, recall, ranking quality, latency and downstream answer behaviour. Results depend on the corpus, configuration, budget and task.
Matched 500K no-assistance calls: same query, same records, same 256K budget, no facets or retries. Models tested via OpenRouter gateway. Six-domain public datasets (science, health, finance, policy, climate, public-interest). Pair counts 6–8 per model.
Weavi-QCE is purpose-built for organisations with large, complex data where retrieval quality and model input cost can both be measured.
Bring together policy documents, adjuster notes, calls, medical evidence and investigation records, then select a budgeted evidence pack for a specific coverage, fraud or complaint question.
Prioritise relevant procedures across large technical-document corpora before sending the selected evidence to an engineer-facing model. Pilot outcomes should measure recall, lookup time and safety-critical omissions.
Use configured materiality signals to reduce a large administrative corpus before semantic analysis and LLM review. Value depends on recall, analyst throughput and the cost of missed cases.
Combine declared importance and source metadata across heterogeneous records, selecting a reviewable subset for downstream analysis. Secure deployment and independent mission testing would be required.
Reduce large email and document sets to a budgeted review corpus while preserving full records. A pilot should compare QCE with sparse, dense and hybrid retrieval under the same relevance labels.
Select configured high-value records for authorised clinical review without summarising the source data. Healthcare deployment would require customer validation, privacy review and appropriate compliance controls.
Data, cloud and model platforms could use a selection layer to manage long enterprise histories before inference. Distribution economics require partner-side volume and pricing validation.
Prepare a traceable subset of telemetry, inspections, historical events and weather evidence for expert and model review. Any safety claim requires domain-owner testing and governance.
Select full passages from large case-law corpora under a fixed model-input budget. Quality should be judged against independently labelled authorities and citation accuracy.
A mid-tier bank processing 50,000 applications/year holds 100M+ tokens: applications, valuations, credit reports and compliance documents. QCE selects only what matters for each risk query — fraud signals, income evidence, covenant breaches — before any model sees it.
Tax authorities, central banks, regulatory bodies and public-sector AI programmes all hold massive, governed corpora. QCE enables LLM-powered policy review, compliance monitoring and citizen-service AI within strict data-minimisation constraints.
Weavi-QCE has been exercised from compact API requests through a 5M physical-token direct-selector study and 500M represented pressure. Physical, declared and represented quantities are reported separately.
| Token scale | Current route | Illustrative workload | Evidence status |
|---|---|---|---|
| 2 Ki – 32 Ki | live API | Chats, focused documents, agent memory | Contract and internal testing |
| 32 Ki – 128 Ki | live API | Cases, claims, meetings, repositories | Internal testing |
| 128 Ki – 500 Ki | six-domain data | Regulatory, medical, scientific and policy records | 96.5% latest labelled recall at 500K→256K |
| 500 Ki – 5 Mi | physical direct selector | Large supplied candidate corpora | 91.5% latest labelled recall at physical 5M→1M |
| 5 Mi – 500M | represented | Extreme compression-ratio and budget-pressure studies | 62.5% recall at represented 500M→5M; not physical 500M |
Test date: 15 Sep 2026. The physical-5M arm used 5,000 complete records and 20M content characters; neutral padding contributed parsing volume and is not independent semantic diversity. Represented scale exercises ranking and budget pressure on bounded physical text. Genuine massive archives require streaming/indexed shards and measured reduction equivalence.
"The best context window is not the biggest one — it is the most truthful one."
Weavi-QCE does not summarise or paraphrase. It selects. What reaches your model is drawn directly from your data — scored, ranked and preserved.
We believe confidence should come from inspection: testing routes, methodology, seeds and outputs are available for technical diligence. We welcome independent review and intend to publish a stronger reproduction package. Request access at hello@weavi.ai.
Define the corpus, budget and success criteria first—then compare the result with your existing pipeline.