IngestCore manual

Everything the product does, in the order you will use it. The screens below come from the real product, with fictitious data — no customer document appears here.

System flow

IngestCore is document plumbing: a file goes in, structured data comes out. The diagram below is the full path — each box matches a screen or routine described in this manual.

Watched folderXML · PDF · JPG · PNGImportautomatic or manualClassificationtype + pipelineSecurity analysisdeterministicQuarantinenever leaves the machineReading pipelineOCR + local AICanonical JSONfields + confidenceReviewonly what the rule asks forERP outputJSON · CSV · layoutARCHITECTURE — three processes on YOUR machine, supervised by the applicationApplication (Electron)Local API · 127.0.0.1Python worker · OCR + AI

A document held by the security analysis stays readable inside the product — what it cannot do is leave: export, integration and cloud stay blocked until someone releases it with a justification.

Installation

Download XCODEF-IngestCore-1.0.0-Setup.exe from the product page or the Microsoft Store and run it. The installation:

  • Does not ask for administrator — it installs in your user profile;
  • Ships everything inside: local API, reading worker, OCR (Portuguese and English) and the local AI manager;
  • Creates shortcuts in the Start Menu and on the Desktop;
  • Keeps your data outside the program folder, in %LOCALAPPDATA%\XCODEF\IngestCore — updating or reinstalling never takes your documents with it.

Requirements: Windows 10/11 64-bit, 8 GB of RAM (16 GB recommended for local AI), 1 GB of disk for the product plus room for your documents. A graphics card is optional — with one, local AI runs faster and the product itself recommends better models.

First run

On first launch, a nine-step wizard sets up the essentials in about two minutes: input folder, OCR language, local AI (optional) and a summary of the security that stays on. “Not now” is a valid answer at every step — everything can be configured later, and the wizard does not come back on its own.

  1. Welcome and preparation — where the data lives;
  2. Component check — database, OCR and data area verified on the spot;
  3. Input folder — pick a folder of yours (e.g. Documents\IngestCore\Inbox). The wizard refuses the product's internal folders;
  4. OCR — main language of your documents;
  5. Local AI — “Analyse this computer” measures processor, memory and graphics card and recommends the right model, with the download size visible before any click;
  6. Cloud (optional) — off by default; it takes three decisions of yours to exist;
  7. Security — summary of what is active;
  8. Finish — the product is already working.

Management cockpit

The screen that answers what the operation's owner asks: how many documents arrived, how many went through with no human touch, the average confidence, where time is being spent and what needs attention today.

Management cockpit: volumes, health, alerts and trends of the operation.

The indicators can be cut by today, 7 days, 30 days or the whole period. Clicking a document type filters the entire cockpit by it — that is how you find out the delay is only in one supplier's invoices.

Watched folder and import

The watched folder is the heart of the operation: every file dropped there is imported automatically (XML, PDF, JPG, PNG). Once read, the original is moved to the read\<TYPE> subfolder — organised by document type — and whatever fails goes to _error with the reason in a file next to it.

You can also import manually with the Import button and reprocess any document without it counting again against the licence.

Input automation: the watched folder, where read files go and the check interval.

Operation and review

The Operations screen is the workbench: processing queue, documents waiting for review and the full detail of each one — extracted data, text read, results per stage and audit trail.

Operations: queue on the left, open document on the right, with classification and actions.

Documents below the type's confidence rule go to review: you see the document and the fields side by side, correct what needs correcting and approve. The rule is configurable per document type — a supplier invoice may demand 95%, an internal receipt may accept 80%.

Consolidated data: header, line items and totals in the format that goes to the integration.

The document tabs show each stage separately, which makes it easy to find where something went wrong: if the OCR text is fine but the field came out empty, the problem is extraction, not reading.

Text read by OCR, with the engine used and the processing time.
Extraction by local AI, with the confidence level of what was understood.
Automatic checks: line items against the total, required fields, well-formed tax IDs.
Audit trail: every step of the document, from arrival to who confirmed the type.
Full history: every document, with filters by status and type.

Document types

Each type (invoice, service invoice, bank slip, receipt, contract…) has its own pipeline: classification keywords, extraction prompt, validation schema, review rule and confidentiality level. The 19 factory types cover day-to-day tax documents — and new types are data, not code: your installation can gain a type specific to your operation without updating the product (XCODEF service, R$ 150/h).

Document types: keywords, expected fields and the review rule of each one.

Local AI

The Local AI tab in the AI Center is the smart manager: it measures your computer (processor, memory, graphics card, disk) and recommends the best setup — you do not need to know what VRAM or quantisation means.

  • Assisted engine installation (Ollama): official source, integrity checked by SHA-256, real progress and cancellation;
  • Approved catalogue: only models tested by XCODEF, with the size visible before the download;
  • One-click test: the model is approved on YOUR machine before going into production;
  • It also works without AI: the embedded OCR keeps reading everything; AI improves extraction, it is not a prerequisite.
Local AI: analysis of the computer and the models recommended for it.

The remaining tabs cover the full intelligence cycle: test extraction on an isolated bench, teach the AI from real corrections, choose providers and models, and track cost when the cloud is used.

Test bench: paste a text, adjust the prompt and see the JSON that would come out.
Teach the AI: today's correction becomes tomorrow's hit on the same supplier's document.

Security and quarantine

Every document goes through a deterministic analysis (no AI — there is nothing to “convince”) that looks for content trying to manipulate the reading: hidden instructions, invisible text, divergence between what the PDF shows and what it contains. High risk means automatic quarantine.

  • A quarantined document does not export, does not integrate and does not go to the cloud — but stays readable so you can judge it;
  • Releasing it requires a justification and is audited;
  • Confidentiality matrix: each document type has a level (Public / Internal / Confidential / Restricted) and the consequence is visible — Restricted never leaves the machine, even with the cloud configured;
  • Every block becomes an auditable event.
Policies: what may leave the machine, by document type and by provider.
Protections that are on by default, including blocking URLs found inside the document.

Licence and activation

The trial includes 1,000 documents or 120 days, whichever comes first — with no sign-up. From 80% usage on, a notice shows the consumption bar, the machine code (XCM-…) and the way to buy.

  1. Copy the machine code on the Licence tab (one click);
  2. Place the order at xcodef.com/ingestcore/pedido informing that code;
  3. Pay by Pix, bank slip or card;
  4. The key arrives by e-mail — paste it on the Licence tab and click Activate. Activation is offline: nothing from your machine is sent.
Licence tab: status, trial consumption, machine code and the activation field.

When the trial ends, ingestion of new documents pauses — everything already processed stays accessible, searchable and exportable. Your data is never held hostage. Changed computers? XCODEF reissues the key for the new code.

Export and ERP

The result of each document is a canonical JSON — fields, line items, totals and confidence level in a stable format. Export produces JSON and CSV ready to import; layouts specific to your ERP are built to order and become part of your installation.

Quarantined documents are held back with a warning — the export states how many were left out and why, so reconciliation never closes wrong in silence.

Settings Center

Everything configurable lives here, grouped by subject and with built-in diagnostics: component status, automation, reading, AI and OCR, operation, folders, backup, security, system and licence. Profiles let you export and import the whole configuration — useful to standardise several machines.

Services: everything running underneath, in plain names, with live state and the command for each part. This is where you restart what failed.
Status: real-time diagnosis of the local service, database, OCR and AI engine.
Operation: worker pace, attempts per task and the limits that raise an alert.
Backup and logs: automatic routine with rotation, with nobody having to remember.

Deployment guide (IT)

PrivilegesPer-user installation (%LOCALAPPDATA%). No administrator or UAC required in normal use.
NetworkThe local API listens only on 127.0.0.1 (preferred port 3333; if taken, it uses another and logs it). No document leaves the machine without a configured cloud provider + a policy allowing it + a budget above zero.
Possible outbound trafficOnly what the customer enables: AI model downloads (official source) and a cloud provider using the customer's own key. Verifiable by firewall.
Internal authenticationRandom per-session token between interface and API — no fixed secret on disk.
Silent installationSetup.exe /S (NSIS). Same artefact as the website and the Microsoft Store.
Data%LOCALAPPDATA%\XCODEF\IngestCore (SQLite database, logs, backups, exports). Uninstalling preserves the data.
UpdatingInstall the new version on top; migrations run by themselves on first launch; data preserved.
ProcessesThree processes: IngestCore.exe (interface), local API (embedded Node) and worker (embedded Python). Supervised by the application; closing the window keeps the watcher in the tray; Exit shuts everything down.
Local AI requirementsOptional. 8 GB of RAM minimum (16 GB recommended); a graphics card accelerates it; the product measures and recommends — with no card it picks light models or OCR only.
Technical logs%LOCALAPPDATA%\XCODEF\IngestCore\logs (desktop.log, api-local.log, worker-ingestcore.log).

Troubleshooting

  • The application opens but says “local service did not start”
    Check %LOCALAPPDATA%\XCODEF\IngestCore\logs\desktop.log — the cause is on the last [ERROR] line. Close it from the tray icon (Exit) and open it again.
  • I dropped a file in the folder and nothing happened
    Confirm in Settings Center → Automation that the watcher is on and pointing at the right folder. Failed files go to the _error subfolder with the reason next to them.
  • A document went to quarantine
    That is the security analysis doing its job. Open the document → Security tab → look at the findings; if it is a false positive, release it with a justification.
  • The licence key does not activate
    Check that you copied the whole key (it starts with XCODEF.) and that it was issued for this machine — the Licence tab shows the local code. Different? Reply to the key e-mail asking for a reissue.
  • Local AI became slow
    AI Center → Local AI → Test. If the test suggests a lighter model, accept it — it was chosen for your actual machine.
  • I need to move the product to another computer
    Install it on the new one, copy the %LOCALAPPDATA%\XCODEF\IngestCore folder (data) and ask for the key to be reissued with the new machine code.

Still stuck? contato@xcodef.com — attach the diagnostic package (Settings Center → Status → Export diagnostics).

Prefer to read offline?
The complete manual, with these screens, as a PDF.
Download the manual as PDFThe PDF is still only available in Portuguese; this page is in English.