Inspiration

Receipts, invoices, warranty PDFs, and photos of paper receipts end up in different folders on a NAS. When a product fails, the important questions are simple but hard to answer: What did I buy? Where is the proof? Is it still covered? Which document matches which purchase? I built Warranty Receipt Graph to turn that scattered evidence into an auditable lifecycle view instead of another list of search hits.

What it does

Warranty Receipt Graph is a report-only Uliya Skill for UGREEN AI NAS, built for the Open Innovation track. Given one user-authorized folder, it links:

Purchase → Receipt/Invoice → Line item/Product → Warranty

It reports active, expiring-soon, expired, and unknown warranty coverage; flags missing purchase proof, amount/date mismatches, ambiguous matches, and unlinked warranty documents; and shows the evidence behind every claim. It never moves, deletes, renames, tags, uploads, or saves anything. Proposed follow-ups are listed but not executed.

How I built it

The Skill delegates read-only work to three registered experts:

  • file-agent discovers bounded document and image candidates.
  • document-agent extracts receipt, invoice, purchase-order, and warranty facts, runs warranty QA, and reconciles supported document pairs.
  • photo-agent runs OCR and capture-metadata lookup on receipt photos.

The parent assembles the graph from verified results only. I wrote an ontology for purchases, receipts, invoices, line items, products, warranties, findings, and coverage; provenance rules for every node and edge; explicit matching rules; and a report format with a synthetic example. Matching uses exact identifiers first, then one exact merchant/date/total/currency composite. Ambiguity is reported, not merged. Warranty expiry is either explicit or computed from a stated term plus a verified start date and labeled inferred. The package passes the official Uliya Skill Kit 0.7.7 static checker with zero errors and zero warnings.

Challenges

The hard part was not OCR; it was avoiding false certainty. Two receipts can share a merchant, date, and total but be different purchases. A warranty document can name a product without proving the purchase. A missing warranty document is not the same as missing purchase proof. I handled these as explicit findings and kept unresolved candidates separate instead of guessing. I also had to keep the parent inside the delegation boundary: it can normalize verified results but must never call NAS CLIs directly.

What I learned

A useful knowledge graph needs provenance and ambiguity handling more than visual complexity. Read-only workflows also need clear boundaries: saying "no writes" is not enough unless the orchestration, report, and proposed actions all reflect it.

What's next

The next step is NAS validation: real folder discovery, document extraction schemas, OCR quality, reconciliation, permissions, large-folder limits, and failure handling. Runtime UI and any state-changing actions are intentionally out of scope for this report-only Round 1 version.

Built With

  • ai
  • document
  • graph
  • json
  • knowledge
  • markdown
  • mermaid
  • nas
  • ocr
  • provenance
  • ugos
  • ugreen
  • uliya
Share this project:

Updates

Submission history