Moving from theory to a tangible, working system that turns AI mistakes into high-quality training data. Ukusuka Ukubuyekezwa kwebhizinisi we-I.I. Ukusebenza kwe-Intelligent Artificial Intelligence (I.I.D.) kuyinto ingxubevange esithakazelisayo phakathi kwe-85% ye-prototype enhle kanye ne-99% ye-system eyenziwe ekukhiqizeni. Sihlanganisa ukuthi ingxubevange akuyona kuphela imodeli enhle, kodwa uhlelo olufunda kuzo zonke iziphumo. I-Part 1 ye-Series Usuku, thina uzothola izindandatho ethu futhi ukwakha isicelo web elula, esebenzayo okuyinto ibonise isikhwama esiyingqayizivele ye-data flywheel. Ngemuva kokugcina le nqaku, uzothola ukuguqulwa kwe-AI futhi ukuguqulwa i-dataset enhle, enhle-tuning-ready kusuka ekusebenzeni lakho. Thina usebenzisa i isibonelo kusuka ku-Open Source Framework yethu, Umehluko le-self-contained, ayidinga izinsizakalo ezingenalutho ezifana ne-Docker noma i-Redis, futhi ibonise ukuthi isampula se-core kungcono kakhulu. correction_deck_quickstart Ukuhlobisa Ukuhlobisa I-Scenario: A Flawed Invoice AI Ukukhetha ukuthi sinikeza i-AI ukuvikela idatha esakhiwo kusuka ku-invoices. Sitholela isithombe se-invoice, futhi sinikeze ukuba ivumela isikhwama se-JSON eluhlaza. Ngemuva kokuqala, i-AI inikeza umsebenzi enhle, kodwa akuyona enhle. It ikhiqiza lokhu imiphumela emibi: { "supplier_name": "Lone Star Provisins Inc.", // <-- TYPO! "invoice_number": "785670", "invoice_date": "2025-08-20", "inventory_items": [ { "item_name": "TAVERN HAM WH", "total_quantity": 15.82, "total_unit": "LB", "total_cost": 87.80 }, { "item_name": "ONIONS YELLOW JBO", "total_quantity": 5, // <-- WRONG QUANTITY! Should be 50. "total_unit": "LB", "total_cost": 35.50 } ] } Umphumela wethu kuyinto ukwakha uhlelo elawula umntu ukuguqulwa ngokushesha ngezinye iziphumo ezimbini, futhi, okungenani, ukuthatha iziphumo zayo zokulungisa. I-Three Core Components ye-Flywheel yethu Ukuze ukwakha lokhu, framework yethu ye-Foundry isekelwe eminyakeni ezintathu elula kodwa enhle ye-Python abstract: Job: Thola lokhu njenge-ticket e-tracking system. It is a database model that represents a single unit of work for the AI. It has the input_data (isithombe yekhwalithi), the initial_ai_output (i-JSON ebuthile emaphakeme), futhi indawo ukugcina i-corrected_output uma umntu ufakele. I-CorrectionRecord: Lesi yiphakheji yedolobha. Uma umntu ukugcina ukucubungula kwayo, akukwazi ukucubungula Job kuphela. Sinikeza i-CorrectionRecord eyahlukile, eyahlukile. Le ibhokisi yenzelwe ngokuzimela ukucubungula. Kuyinto ibhokisi eluhlaza ye-input yokuqala, isampula esihle we-AI, kanye ne-"ngokuqhathanisa" ye-human. Kuyinto isibonelo ephelele, esebenzayo sokucubungula. I-correctionHandler: Kuyinto isakhiwo se-business. Kuyinto isigaba elula esihlanganisa inqubo: ithatha idatha ye-form ebhaliswe kusuka ku-web UI, itholisa, update Job, yenza i-CorrectionRecord, futhi kusebenza ukuhambisa zonke imibuzo ku-training file. Lezi zinyathelo ezintathu zihlanganisa ukwakha isakhiwo se-flywheel yethu. Manje, thina ubona kubo ekuphenduleni. Let's Build It: I-Quickstart ku-Action Uma uxhumane, uxhumane Ukuhamba ku-The isixhumanisi, kanye nokufaka dependencies. I-Foundry Repository examples/correction_deck_quickstart Isinyathelo 1: Yenza i-Quickstart Script Kusukela ku-terminal yakho, kulula nje: python quickstart.py Ukubonisa umbhalo ukuthi i-webserver ye-local iyasungulwa ku . http://localhost:8000 --- Foundry Quickstart Server running at http://localhost:8000 --- --- Open the URL in your browser to use the Correction Deck. --- --- Press Ctrl+C to stop the server and complete the flywheel. --- Isinyathelo 2: Usebenzisa I-Correction Deck UI Ukukhula le-URL ku-browser yakho. Uya kuba i-Correction Deck UI elula. Kwi-left i-source invoice image. Kwi-right i-web form eyenziwe ngempumelelo nge idatha ebuthile ye-AI. Ukusebenza kwakho kuyinto umntu ku-loop. Yenza lezi zibe amayunithi: Ukuguqulwa kwe-Typo: Change Lone Star Provisins Inc. ku-Lone Star Provisions Inc. Ukuguqulwa Kwamanani: Qhagamshela inani ye-ONIONS YELLOW JBO kusuka ku-5 kuya ku-50. Waze Save Correction. Isinyathelo 3: Complete the Flywheel Sishayele ku-terminal yakho, futhi ukugcina i-server ngokuchofoza . I-script ikhiqiza ngokuzenzakalelayo isinyathelo sokugqibela se-flywheel: ukuhambisa umsebenzi wakho. Uyakwazi ukubona lokhu imiphumela: Ctrl+C --- Server stopped. --- --- Exporting approved corrections to fine-tuning format... --- --- Data successfully exported to 'corrected_data.jsonl' --- --- QUICKSTART COMPLETE --- Uyavuma. Uyavumela ukufinyelela okungenani okungenani we-data flywheel. I-Payoff: I-Training File ephelele Waze umbhalo. Uzothola ifayela entsha: Kuyinto umphakeli. Lesi isiphumo esifanelekayo sokusebenza kwakho, esithathwe futhi ifakwe ngokushesha ukucubungula imodeli ye-AI yesimanje. examples/correction_deck_quickstart corrected_data.jsonl Thola thwebula ngaphakathi. Kuqukethe umugqa omunye we-JSON eyakhelwe: {"contents": [{"role": "user", "parts": [{"fileData": {"mimeType": "image/jpeg", "fileUri": "/static/example_invoice.jpeg"}}, {"text": "Extract the key business data from the provided input."}]}, {"role": "model", "parts": [{"text": "{\"supplier_name\": \"Lone Star Provisions Inc.\", \"invoice_number\": \"785670\", \"invoice_date\": \"2025-08-20\", \"inventory_items\": [{\"item_name\": \"TAVERN HAM WH\", \"total_quantity\": 15.82, \"total_unit\": \"LB\", \"total_cost\": 87.8}, {\"item_name\": \"ONIONS YELLOW JBO\", \"total_quantity\": 50.0, \"total_unit\": \"LB\", \"total_cost\": 35.5}]}"}]}]}``` Kuyinto ingangena ngempumelelo, kodwa kuyinto ifomati esifanayo ukuthi amamodeli ezifana ne-Google's Gemino ne-OpenAI's GPT uchungechunge ukunambitheka. "i-role": "i-user": Kuyinto ifomu. Kuyinto ifomu yokufaka (fileUri) kanye ne-instruction etholakala ku-AI. "i-role": "model": Kuyinto impendulo ephelele. I-JSON string nge-corrections yakho asetshenziswe. Thina ngempumelelo ukuguqulela amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi amaminithi. Yini elandelayo? Thina siphinde i-core loop ye-flywheel: Correct -> Capture -> Format for Training. Kuyinto lokuqala enamandla, kodwa inqubo offline. Thina siphinde ukuba i-AI ufakele iphasiwedi yayo, bese siphinde sishintshile umsebenzi yayo. Kodwa ukuthi ungayifumana kakhulu? Yini uma i-pipeline ingasebenza, ukuhlangabezana ne-imeyile esidinga, futhi ngokuthembeka ngokuthembeka ukuze bafuna umntu ukwesiza isikhathi esifanayo? Ngo nqaku elilandelayo kulesi sethu, siza ukwakha i-Human-in-the-Loop pipe that knows when it’s in trouble and doesn’t fear to ask for clarification.
