AION LUMEN
My SAP certificate was due to expire that day. Access to recertification was blocked, even though my learning subscription was paid.
Folio had already collected my emails and records locally. One AI assistant searched that collection directly for the relevant documents. Another prepared the support request and assigned the licence with my approval. I sent the request myself.
The valid licence had not been assigned to my account.
Emails and documents collected locally.
Searched Folio’s collection directly.
Prepared the request; I sent it. Assigned the licence with my approval.
Release preview.6 builds on this foundation: optional vector search retrieves knowledge phrased differently. The memory extensions organise claims by topic and submit them for independent review.
A separate test with fictional records demonstrates the new search and memory features using local models.
One email contains a debit and a blocked assessment. The payment record identifies neither the product nor the licence period.
Actual Folio screenshot · fictional data · open to enlarge
Qwen separates financial and certification information. This is the access and deadline claim before independent review.
Actual Folio screenshot · fictional data · open to enlarge
Gemma accepts the access and deadline claim. The finance bundle is held because of its event entry. This rejection does not establish that the payment is false.
Actual Folio screenshot · fictional data · open to enlarge
The German question uses Wissenstest; the source says Assessment. In this small memory-search case, lexical search finds nothing, while vector retrieval adds the reviewed access and deadline claim.
Actual Folio screenshot · fictional data · open to enlarge
The claim, confirmed by the second model, retains its source reference. It helps prepare the next question: which licence is assigned to which account?
Actual Folio screenshot · fictional data · open to enlarge
These screenshots come from an isolated Folio instance running memory-quality-20261014-rc7. Both emails are fictional. Import and initial topic routing were prepared. Extraction, independent review, indexing and retrieval used unmodified product code. The images show saved states, not a continuous recording of a click-through workflow. The product screenshots remain in German.
| 15 Oct · Demo | Two emails. The first produces finance and certification proposals. Qwen3.8 27B extracts; Gemma 4 31B reviews. One claim is accepted. The finance bundle is held because the reviewer flags its event entry; this may include an overly conservative judgement. For the second email, the correction attempt discarded supported details; Folio therefore accepted nothing. |
|---|---|
| 15 Oct · Retrieval | For this paraphrase in memory search, lexical search returns 0 results and hybrid search returns 1. The corpus contains only one confirmed claim. No unrelated control question was tested in the same domain, so this does not establish selectivity or general retrieval performance. The companion also uses a vocabulary bridge between Wissenstest and Assessment. Finance is excluded from vector retrieval. |
| 14 Oct · Software | 520 automated tests passed. Type checking found no errors and 16 existing warnings. These checks test software behaviour, not model accuracy. |
| 14 Oct · Local reviewer | 20 valid synthetic cases matched expectations: 12 rejections and 8 acceptances. Three malformed fixtures were corrected and rerun; their original outputs are preserved and excluded. |
A separate 14 October pilot using the same fictional problem did not establish autonomous support handling. Qwen produced no final draft within the output limit. Gemma produced a draft that assumed an unresolved payment-to-product association. This local-model test does not demonstrate autonomous support handling. The personal story above involved separate cloud AI sessions.
The new build is active locally. Automatic processing of the historical file backlog remains paused while work on longer documents undergoes further testing.
Synthetic sources and recorded outcomes (JSON) · Earlier model comparison, 7 October