{"resourceId":"aws-naswa-nlx-nova-lite-job-posting-extraction-20261101","versions":[{"version":"external-79b9e8ca94317b4ad52dcd3fa3f8c837a0b711a59863d91d5d0e677a712b32c9","resource":{"id":"aws-naswa-nlx-nova-lite-job-posting-extraction-20261101","title":"AWS says redesigning NASWA's job-posting extraction pipeline around Amazon Nova Lite and prompt decomposition cut processing costs 80 percent","organization":"National Association of State Workforce Agencies (NASWA) NLx Research Hub, with Amazon Web Services","sector":"State workforce agencies; labor market information","geography":"United States (national; serves all state workforce agencies)","publishedAt":"2026-11-01","publicationDate":"2026-11-01","eventDate":null,"sourceName":"AWS Public Sector Blog","sourceLabel":"How NASWA scales real-time labor market insights with AWS generative AI","sourceUrl":"https://aws.amazon.com/blogs/publicsector/how-naswa-scales-real-time-labor-market-insights-with-aws-generative-ai/","evidenceClass":"vendor-claim","outcomeClass":"emerging","topics":["infrastructure","operating-model","governance-procurement"],"finding":"An AWS solutions architect writes (Nov 1) that NASWA's National Labor Exchange (NLx) Research Hub, holding almost 200 million job postings, first used Amazon Bedrock batch processing to extract structured fields and hit performance and cost problems before launch. AWS and NASWA rebuilt the pipeline to combine batch and on-demand processing with Amazon Nova Lite and to split one large prompt into separate prompts for education requirements, pay, benefits, remote status and full-time status. AWS reports an 80 percent cost reduction against the original approach, says bottlenecks that threatened the launch timeline were resolved, and says data quality was maintained. The parsed pay, benefits, degree, remote and full/part-time fields reach Research Hub users starting in 2027. AWS's own lessons include separating explicit extraction (a stated salary) from inference (an occupation code).","sledRelevance":"Cross-tagged to State Government because the Research Hub serves state workforce agencies and their labor market information units, which use posting data for training investment and policy. This is a rare AWS public-sector example of production-scale batch extraction where cost, not chat, is the design driver, and it transfers to state agencies that process large document backlogs such as licensing files or case notes. Relevance is bounded: the inputs are public job postings without personal data, so privacy constraints that govern most state agency records do not apply here.","evidence":"Vendor blog by an AWS solutions architect, with an unnamed 'project team lead' quoted. The 80 percent figure has no dollar baseline, token volume, model comparison or time period. 'Maintained data quality' is not quantified, and no accuracy, error-rate or human-review figures are given. The only published validation of the underlying method is the 2025 ACM paper filed separately in this edition, which used a different model and a 6,800-posting sample.","architectureImplications":"The design pairs a low-cost model (Nova Lite) with several narrow prompts, one per field, and mixes batch with on-demand calls instead of relying on Bedrock batch jobs alone. More prompts per record raise call counts but shorten each output, which the 2025 paper says drives cost. Agencies copying the pattern need orchestration, retries and per-field result storage.","governanceImplications":"Labor market statistics derived from model output can inform policy and funding, so per-field validation against external benchmarks and documented model versions matter. The post does not say how NASWA will publish error rates or method notes when fields go live in 2027.","securityPrivacyImplications":"Job postings are public and contain no individual records, so exposure is low. The pattern changes if an agency applies it to case files, claims or resumes, where data classification, retention and access logging would have to be designed in.","caveats":"Vendor-authored; cost and quality claims are unaudited and lack baselines. Data reaches users only in 2027, so there is no user-side evidence yet. Transferability to sensitive state records is untested.","streamIds":["aws","state-government"],"roles":{"sales":"Sales framing: this is a cost story for document-heavy batch work, not a chatbot story. State labor market information units and agencies with large text backlogs can use it to open a conversation about extraction pipelines priced per token. Lead with the redesign lesson, not the 80 percent figure, because AWS gives no baseline. Ask what records they would process, how they would judge accuracy, and who would fund ongoing runs. Interpretation: value depends on validating each field against a trusted benchmark, which the post does not report.","engineering":"Engineering framing: the reusable idea is per-field prompt decomposition on a small model with mixed batch and on-demand calls. Before proposing it, size token volumes, test whether narrow prompts actually beat one schema prompt on the customer's documents, and plan orchestration and retries. Treat extraction and inference fields differently, as AWS itself advises. The post gives no accuracy data, so a proof of value must measure per-field precision and recall against labeled samples and compare cost with the agency's current method.","delivery":"Delivery framing: plan a bounded pilot on a representative document sample with a labeled answer set. Define per-field accuracy targets, a cost ceiling and a documented model version before scaling. Assign an agency data owner for field definitions and a technical owner for the pipeline after vendor handoff. Proposed acceptance steps are measurable checks, not observed results: NASWA's fields are not yet released, so the pilot creates the agency's first evidence on quality and cost."},"roleDetails":{"sales":{"customerProblem":"Agencies hold large volumes of unstructured text, such as job postings, that are too costly to structure with one large prompt or manual coding.","stakeholders":["State labor market information director","Workforce agency CIO","Agency data governance lead","Budget officer"],"discoveryQuestions":["Which document sets would you structure first, and how many records per day?","What trusted benchmark could validate each extracted field?","Do the documents contain personal or confidential data?","Who pays for ongoing per-token processing after the pilot?"],"valueHypothesis":"Splitting extraction into narrow prompts on a low-cost model may lower processing cost for high-volume, low-sensitivity text; the size of any saving is unknown until measured against the agency's current method.","potentialEngagement":"A six-to-eight-week extraction pilot on one document set with a labeled validation sample and a cost model, ending in a scale or stop recommendation.","unsupportedClaims":["A guaranteed 80 percent cost reduction for other agencies","Accuracy or data quality at any stated level","That the pattern suits records containing personal data"]},"engineering":{"fit":"Strong fit for high-volume, public or low-sensitivity text with clearly stated fields; weaker where fields must be inferred.","architectureIntegration":"Per-field prompts to Amazon Nova Lite on Bedrock, orchestrated across batch and on-demand calls, writing structured results to the agency's data store.","prerequisites":["Bedrock access with Nova Lite in the chosen Region","Field definitions agreed with data owners","Labeled sample for validation"],"constraints":["No accuracy figures published for Nova Lite in this use","More calls per record raise orchestration complexity","Sensitive records need separate classification and controls"],"security":"Low exposure for public postings; for agency records, apply data classification, encryption, access logging and retention rules before any prompt sees the data.","proofOfValue":["Measure per-field precision and recall against a labeled sample","Compare total cost per thousand records with the current method","Test single-prompt versus decomposed prompts on the same documents"]},"delivery":{"work":"Build and validate a field-by-field extraction pipeline on a bounded document set, then decide on scale-out.","dependencies":["Agency field definitions and benchmark data","Bedrock model access and quotas","Vendor or partner architects during build"],"ownership":"Agency labor market information or data office owns field definitions and publication; agency IT owns the pipeline after handoff.","skillsAdoption":"Train analysts to review extraction errors and engineers to version prompts and models.","governanceCheckpoints":["Data classification sign-off before ingestion","Method and error-rate documentation before publishing statistics","Model version change review"],"acceptanceCriteria":["Each released field meets an agreed accuracy threshold on a held-out sample","Cost per thousand records falls below an agreed ceiling","Pipeline reruns reproduce results with the same model version"],"risks":["Unmeasured extraction errors flow into policy statistics","Model updates change outputs silently","Dependence on vendor architects after launch"]}},"retrievedAt":"2026-11-02T14:33:00Z","enrichedAt":"2026-11-02T14:40:06Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Workforce agencies could get regional trend data without buying commercial labor market feeds; no staff-time effects are reported.","procurementImplications":"Shows a co-development model where an AWS account team redesigned a member-association pipeline; agencies should clarify who owns prompts, code and ongoing model costs before similar work.","operatingModelImplications":"Weekly AWS-NASWA touchpoints and an AWS-led redesign suggest the agency relied on vendor architects; sustaining the pipeline after handoff is not described.","sourceVerification":{"openedUrl":"https://aws.amazon.com/blogs/publicsector/how-naswa-scales-real-time-labor-market-insights-with-aws-generative-ai/","referenceExcerpt":"Realized an 80 percent cost reduction compared to the original processing approach","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}