{"resourceId":"nvidia-nemo-french-public-service-assistant-repo-scope-20261027","versions":[{"version":"external-746933e6694b1700656200a0ec34642fb8dda1c78962dda27581be376307d648","resource":{"id":"nvidia-nemo-french-public-service-assistant-repo-scope-20261027","title":"NVIDIA's French public-service assistant code ships six sourced procedures and ten test prompts, and states tuning does not add facts","organization":"NVIDIA (NVIDIA-NeMo/Nemotron repository)","sector":"Government citizen services (open-source reference code)","geography":"France; Global","publishedAt":"Undated repository README (retrieved 2026-10-27)","publicationDate":null,"eventDate":null,"sourceName":"GitHub: NVIDIA-NeMo/Nemotron","sourceLabel":"use-case-examples/french-public-service-assistant README","sourceUrl":"https://github.com/NVIDIA-NeMo/Nemotron/tree/main/use-case-examples/french-public-service-assistant","evidenceClass":"vendor-claim","outcomeClass":"cautionary","topics":["developers-agents","governance-procurement","knowledge-work"],"finding":"The reference implementation contains six procedures whose facts come from published pages, each with source URL and verification date, plus ten test prompts with expected behavior. The README says fine-tuning teaches behavior such as staying inside supplied facts but does not add facts the model lacks, and that production assistants should read fees, documents and deadlines from the agency's own data or retrieval at answer time. It says the workflow has been run end to end on a DGX Spark.","sledRelevance":"Gives agencies a concrete template for sourcing and dating facts and an explicit vendor statement that fine-tuning is not a substitute for authoritative retrieval, which is central to any public-service assistant procurement.","evidence":"Repository documentation and code structure; ten test prompts is a smoke-test scale, not an evaluation.","architectureImplications":"Separate behavior tuning from fact delivery: tuned model plus retrieval over agency-owned, dated sources.","governanceImplications":"Each fact should carry a source URL and verification date; agencies need a review cadence for re-verifying facts.","securityPrivacyImplications":"Retrieval sources should be official published content; avoid placing resident records in the fact store without privacy review.","caveats":"Six procedures and ten prompts is far smaller than a production service catalog; README is undated and may change.","streamIds":["nvidia"],"roles":{"sales":"Use the repository's own wording to steer customers away from 'train the model on our rules' expectations. NVIDIA's code says tuning does not add facts and that fees, documents and deadlines should come from the agency's data at answer time. That opens a broader engagement: building a sourced, dated fact store and retrieval layer, with tuning as a behavior layer. It is a strong procurement-readiness message for state and county buyers drafting assistant RFPs. Be clear that the sample covers only six procedures and ten prompts.","engineering":"Engineering should adopt the repository's structure as a starting point: facts drawn from published pages with source URL and verification date, test prompts with expected behavior, and a tuning step limited to behavior. Implement retrieval for every volatile fact and ensure the tuned model answers only within retrieved facts. Expand the ten prompts into a much larger holdout set covering each service and edge case. Note that the workflow was run end to end on a DGX Spark, so a proof of value can run on modest local hardware.","delivery":"Delivery should stand up a content-operations process alongside the model work: assign owners for each procedure, record source and verification date, and set a re-verification cadence. Expand from the sample's six procedures to the agency's priority services in phases. Acceptance should require every answer's facts to trace to a dated source and the holdout set to pass agreed thresholds. Track fact-store staleness as an operational metric and plan staffing for content maintenance after go-live."},"roleDetails":{"sales":{"customerProblem":"Agencies expect fine-tuning to teach the model their rules, risking outdated or invented facts.","stakeholders":["Program office owner","Web/content manager","CIO","Procurement officer"],"discoveryQuestions":["Where are your authoritative fees and deadlines published?","Who verifies them and how often?","Will your RFP require sourced answers?"],"valueHypothesis":"A sourced, dated fact store plus behavior tuning yields assistants that can be audited and updated without retraining.","potentialEngagement":"Fact-store and retrieval design engagement for priority services.","unsupportedClaims":["Ten test prompts is not a production evaluation.","No measured accuracy results in the README."]},"engineering":{"fit":"Any citizen-service assistant with volatile facts.","architectureIntegration":"Tuned model for behavior; retrieval over agency-owned dated sources for facts.","prerequisites":["Published procedure pages","Source URL and verification date per fact","Retrieval pipeline"],"constraints":["Tuning does not add facts","Sample scope is six procedures"],"security":"Limit fact store to official content; privacy review for any non-public data.","proofOfValue":["Every answer traceable to a dated source","Expanded holdout pass rate"]},"delivery":{"work":"Build fact store with owners and verification dates; expand procedures in phases.","dependencies":["Content owners","Retrieval infrastructure"],"ownership":"Program office content owners with IT support.","skillsAdoption":"Train content owners on verification workflow.","governanceCheckpoints":["Fact provenance review","Re-verification cadence agreed"],"acceptanceCriteria":["All facts sourced and dated","Holdout thresholds met"],"risks":["Stale facts","Under-resourced content maintenance"]}},"retrievedAt":"2026-10-28T03:05:27Z","enrichedAt":"2026-10-28T03:12:33Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Program staff must maintain the fact store; content owners need time allocated for verification.","procurementImplications":"RFPs for citizen-service assistants can require dated, sourced fact stores and retrieval-at-answer-time for volatile facts.","operatingModelImplications":"Introduces a content-operations role for fact verification alongside model operations.","sourceVerification":{"openedUrl":"https://github.com/NVIDIA-NeMo/Nemotron/tree/main/use-case-examples/french-public-service-assistant","referenceExcerpt":"Six procedures with facts taken from published pages, each with its source URL and verification date","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}