Invoice Parser (desktop)

Windows desktop replacement for the invoiceProject Classic ASP web app. One exe: WPF window hosting WebView2, an embedded local web server (loopback only), and a single background worker that parses queued invoices with the configured LLM. No IIS, no browser needed for parsing.

What changed vs the web app

• Upload -> prep -> parse is fully automatic. Drop files on the dashboard; the server extracts the PDF text layer (PdfPig) or renders pages (PDFium), queues the invoice, and the in-process worker calls the LLM. The UI gets live updates over SSE - no polling, no "open it and press Scan Invoice".
• Scanned PDFs no longer dead-end: pages are rendered server-side on demand.
• One worker owns the queue (replaces pending_process.asp, parse_queue.ps1 and the browser-side parser), so the claim/steal/stale logic is gone.
• All SQL is parameterized.
• pending/{id}/state.json keeps the same format - existing drafts can be migrated by copying invoiceProject\pending\* into this app's pending\.

Configuration (appsettings.json next to the exe)

Secrets live in appsettings.local.json (same shape, git-ignored, loaded on top of appsettings.json). Put ConnectionStrings:Pos and the two ApiKey values there; appsettings.json ships with those blank. Use a SQL login limited to the posbdat/posinv tables rather than sa.

• ConnectionStrings:Pos - SQL Server connection string.
• Llm:TextRoute / Llm:ImageRoute - { Url, Model, ApiKey } per customer.
  • Local llama.cpp: Url: "http://host:8080", empty ApiKey.
  • DeepSeek/OpenAI-compatible: full base URL + ApiKey (sends Bearer header).
  • Gemini: Url: "https://generativelanguage.googleapis.com", Model gemini-..., ApiKey (sends x-goog-api-key).
• Llm:Prompt - override the extraction prompt; empty = built-in default.
• Llm:TimeoutSeconds, Llm:MaxImageDim - LLM call timeout, vision image cap.
• Llm:AttachPageImages - text-layer PDFs: send each rendered page image with its extracted text (default true). Fixes vendors whose text layer is garbled by the form overlay. Needs a vision model on TextRoute; set false for text-only models (DeepSeek, a text-only local llama.cpp).
• Paths:Pending / Paths:Archive - relative to the exe or absolute.
• Port - local UI port (falls back to a random port if taken).
• Logging - Enabled, Directory (default logs), RetainDays (14), MaxFileMB (20), LogRequests. See "Log" below.
• Batch - Enabled, NamePrefix (INV_), CreatedBy (INVOICE), MinCostDelta (0.01), MarginThreshold (15). See "Cost batches" below.
• Tables:ProductBatchesHeaderTable / Tables:ProductBatchesTable - the posbdat batch tables a cost batch is written to.

Log

logs\invoiceparser-YYYYMMDD.log, one file per day, pruned after Logging:RetainDays. Copy the folder or read it over a share to see what a machine did without sitting at it. One line per event:

2026-09-11 14:22:31.104 INFO  [http] POST /api/save 200 412ms

Categories: startup, shutdown, crash (unhandled WPF/AppDomain/task exceptions), http (every request; 4xx logs WARN, 5xx ERROR), upload, prep, worker (queue and per-invoice stage changes), llm (endpoint, status, elapsed, bytes; failures carry the response body), match, db, catalog, save, batch, product, pending, archive. Connection-string passwords and API keys are stripped before anything is written. Set Logging:LogRequests to false to keep only warnings and errors.

Cost batches

Saving an invoice also writes a posbdat price/cost batch for the items whose cost moved. The invoice bills by the case, the product file costs by the unit, so the comparison is case cost / pack against Products.cost; a gap of at least Batch:MinCostDelta is a change, and an item with no cost on file counts as one. Rows with no invoiced cost, and UPCs that are not in the product file, are left out.

The batch is written inside the same transaction as InvoiceHeader and InvoiceDetail - if it cannot be written, nothing is saved - and InvoiceHeader.Price_Cost_Change is set to 1 when the invoice produced one. It is named INV_<invoice number> (_1, _2... if that name is taken) and carries the new cost only; retail is left alone, and items whose gross margin at the new cost falls under Batch:MarginThreshold are flagged on the archive screen for manual repricing. Apply it the usual way (sp_ApplyBatch).

The batch tables' column lists are read from INFORMATION_SCHEMA on first use and whatever Products and ProductBatches genuinely have in common is copied across, so a differently shaped install does not need a code change. If the batch tables are missing or have no cost column the invoice still saves and the reason is logged under batch.

Build / run (dev)

dotnet build
bin\Debug\net8.0-windows\InvoiceParser.exe

Deploy (copy-paste, no install)

dotnet publish -c Release -r win-x64 --self-contained -p:PublishSingleFile=true

Copy the publish folder (exe + appsettings.json + appsettings.local.json + wwwroot) to the target machine and run the exe. Requirements on the target: the WebView2 Runtime (preinstalled on updated Windows 10/11; otherwise the 2 MB Evergreen bootstrapper from Microsoft), network access to SQL Server and the LLM host.

Layout at runtime

InvoiceParser.exe
appsettings.json
appsettings.local.json   connection string + API keys (not in git)
wwwroot\              UI (HTML/JS/CSS, local jQuery)
pending\{id}\         state.json, source.<ext>, pages\page_N.jpg
archive\              {vendorId}_{invoice}.<ext> + .preview cache
logs\                 invoiceparser-YYYYMMDD.log
webview2_data\        WebView2 profile (created on first run)