Reliable event facts and workflow transitions

Index

Updated 5 October 2026. Local application changes are not deployed. See the complete fix record and final report.

Why these changes

Broad semantic search can return only part of a FAQ, a stale page, or an ordinary speaker session when the visitor asks for keynote speakers. Extra model reasoning does not repair incomplete evidence. These changes give the existing assistant narrower source contracts and preserve the existing workflows.

Exact event facts

The read-only eventFacts tool now covers 23 topic identifiers: free-ticket pathways, onsite registration, stand plans/manual access, volunteering, refunds, transfers, upgrades, invoices/payment, floorplans, registration, startup eligibility/options/inclusions/Award, attendance, industries, recordings, discounts, HR Leader eligibility, Explorer applications, parking, opening hours and sponsorship opportunities. It reads complete uploaded FAQ documents in the general and current-event Pinecone namespaces, reconstructs overlapping text chunks, and retains complete matching question/answer entries. It also reads complete event-page bodies from the existing Pinecone sitemap namespace. The existing publish/refresh pipeline remains responsible for keeping that namespace aligned with published Sanity pages.

For these narrow questions the API can return the exact FAQ answer without asking a model to paraphrase it. Conditional wording and eligibility requirements stay intact. Exhibitor logistics uses customersuccess@unleash.ai; the other topics use support@unleash.ai. Links to unpublished internal pages are removed. Missing volunteer information produces a support fallback rather than an invented team list. When a volunteer page is published, the tool returns its complete body to the model.

This tool follows the existing Sanity capability switch and is available without permission to submit email. It does not publish content, mutate Pinecone, register visitors or send enquiries. FAQ reads do not add an application cache; underlying publishing/indexing propagation remains a separate operational measurement.

Code: topic selection and answer policy, source loading, Pinecone document reconstruction, tool registry.

Startup and withdrawn facts

Startup badge checkout, reviewed exhibiting applications and the separate Startup Award application have distinct answers and links. Published audience descriptions are not eligibility criteria; applying does not guarantee approval or a prize. A support-package value must not become a promised cash prize. Recording terms for separately purchased recordings do not establish pass inclusion. Missing pricing, eligibility or logistical details remain unavailable.

For recognized named room/document questions, the application requires current source evidence for the same entity and requested field. A similar room/document or stale conversation detail cannot fill a removed answer. This conservative check covers a narrow pattern; it is not a universal semantic fact checker and adds no Jev answer-validation call.

Code: source-only policy responses, named fact check, streaming link normalization.

Keynote speakers

The canonical session Format field determines keynote status, ignoring casing and contradictory older Keynote session markers. The shared agenda summary applies the same rule. A narrow explicit request to list keynote speakers reads the complete agenda and returns only speakers attached to sessions with that format. Other agenda questions continue through existing agenda tools.

The agenda store retains its existing ten-minute warm-instance cache, cleared in the uploading process on upload; this change does not promise immediate upstream event-management sync. Missing keynote sessions produce a clear unavailable answer. Stale event-management formats are corrected by refreshing the existing agenda namespace, rather than changing the assistant’s lookup source. Personalised plans retain the existing interest and pillar collection.

Code: agenda parsing and structured lookup, agenda upload formatter, API tool and response handling.

Switching workflows

Explicit requests for registration help or to proceed with a named pass switch into the existing ticket workflow even after an agenda conversation. An email provided in that flow no longer gets diverted into agenda delivery just because an earlier plan exists. A confirmed send containing “looking forward to the event” is not treated as a malformed agenda-email recipient.

Profile extraction preserves saved booking fields through null values; explicit personal singular-ticket requests supply one ticket, quantity ranges remain estimates, and event/new-booking changes clear old details. A requested group quote can retain no selected pass. Information-only pass/pricing questions do not initiate company/role collection. The ticket-loop replay verifies that these fields reach the consent summary.

The selector still routes workflow destinations. These corrections do not make classification a consent decision or add an answer-validation model. The existing ticket workflow continues to collect missing details, summarise them, obtain confirmation and verify the submission result.

Code: ticket intent rules, workflow orchestration and email guards.

Safety and verification

Tests cover event isolation, complete FAQ aliases, conditional registration policy, published-link filtering, contact routing, workflow transitions, contradictory keynote markers and the read-only capability switch. The targeted end-to-end retest uses isolated OpenAI and Claude apps and loopback-only submission receivers. Raw evaluator verdicts remain unchanged; source gaps and obsolete ground truth are recorded separately in the failure-only report.

The earlier one-day-to-two-day agenda defect is fixed and retested, together with source-year, recipient/profile reuse and selected-plan validation. The final reviewed cohort has no verified open application defect; raw evaluator non-passes remain in the report with their reviewed causes. Final local tests are 264 web library plus 39 shared AI passes.

Seven live Pinecone mechanical update/delete scenarios passed, followed by four verified response probes across OpenAI and Claude for withdrawn FAQ/file facts. Initial response defects and the failed prompt-only attempt remain archived. These observations are not a Sanity publish-to-answer SLA; actual publish-hook configuration, sustained timing, PDF extraction and real CRM/email delivery remain unverified. Application changes remain local; five published Paris pages were refreshed into 22 vectors through the existing pipeline.

No new external plugin or runtime “skill” is required. The useful additions here are a narrow read-only tool, source rules, and deterministic workflow guards in the existing stack.