Moving from theory to a tangible, working system that turns AI mistakes into high-quality training data. Kwimeko Kwakhona, thina siphile malunga neengxaki ezisetyenziswa kwe-AI, nto leyo ingxaki ebandayo phakathi kwe-85% ye-prototype ebandayo kunye ne-99% ye-system ebandayo ekukhiqizeni. Sifunyenwe ukuba ingxaki yokuba i-model engcono kuphela, kodwa i-system ebandayo ukufundisa kuzo zonke iingxaki. I-Part 1 ye-Series Kwiintsuku, siza kuzakulungisa izixhobo zethu kunye nokwakha isicelo web elula, esebenzayo esibonisa i-core loop ye-data flywheel. Kwixesha lokugqibela le nqaku, uya kubuyekeza ingxaki ye-AI kunye nokukwazi ukuvelisa i-dataset epheleleyo, ezisetyenziswa kakuhle ukusuka kwimveliso yakho. Ukusetyenziswa kwe- isibonelo ukusuka kwinkqubo yethu open-source, Le nqakraza i-self-contained, ayidinga iinkonzo ze-external njenge-Docker okanye i-Redis, kwaye ibonise ukuba i-pattern ye-core kunokwenzeka. correction_deck_quickstart Ukucinga Ukucinga I-The Scenario: A Flawed Invoice AI Ukukhangisa ukuba sisakhiwa i-AI ukufumana idatha eyenziwe kwi-invoices. Sinikezela umfanekiso we-invoice, kwaye sinikezela ukuguqula i-object ye-JSON eluhlaza. Kwi-pass yayo yokuqala, i-AI ikwenza umsebenzi elifanelekileyo, kodwa ayikho epheleleyo. It ikhiqiza lo mveliso olungaphakathi: { "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 } ] } Indawo yethu kuyinto ukuvelisa inkqubo enokukwazi umntu ukuguqulela ngempumelelo ezimbini ezimbini, kwaye, ngokufanelekileyo, ukwamkela izibuyekezo ze-retraining. Iimpawu ezintathu eziphambili ze-Flywheel yethu Ukwenza oku, framework yethu ye-Foundry ibonelela kwiintlobo ezintathu ezincinane kodwa ezinamandla ze-Python: Job: Qinisekisa oku njenge-ticket kwi-tracking system. It is a database model that represents a single unit of work for the AI. It holds the input_data (i-image invoice), the initial_ai_output (i-flawed JSON phezulu), kwaye indawo yokugcina i-corrected_output xa umntu ufakelo. I-CorrectionRecord: Yinto i-ticket yedolobha. Xa umntu ushiye ukuguqulwa kwayo, siya kuthetha nje ukuhlaziywa kweJob. Siza kuvelisa i-CorrectionRecord eyahlukileyo, eyahlukileyo. Le nqakraza yenzelwe ngempumelelo. Izixhobo i-copy eluhlaza yeengxaki yokuqala, i-inthanethi ye-AI, kunye ne-"ukuguqulwa kwe-ground truth" ye-umntu. I-example epheleleyo ye-training ye-portable. I-correctionHandler: Oku yinkqubo yebhizinisi. I-class efanelekileyo enikezela inqubo: ithatha idatha ze-form eyenziwa kwi-web UI, ibhalisele, ukuhlaziywa kweJob, yenza i-CorrectionRecord, kwaye ibhalisele ukuhambisa zonke iinkcukacha kwifayile ye-training. Zonke iintsuku ezintathu zihlanganisa kunye nokwenza isakhiwo se-flywheel yethu. Ngoko ke, thina ukhangela kwi-action. Let's Build It: I-Quickstart kwi-Action Ukuba ulandele, uqhagamshelane Ukuhamba kwi isixhobo, kunye nokufaka iingxaki. I-Foundry ye-Repository examples/correction_deck_quickstart Isinyathelo 1: Yenza i-Quickstart Script Ukusuka kwi-terminal yakho, nqakraza nje: python quickstart.py Ukubonisa umyalezo ukuba i-webserver ye-local iye yasungulwa kwi . 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: Ukusebenzisa i-Correction Deck UI Ukuvula le URL kwi-browser yakho. Uya kuba UI elula ye-Correction Deck. Kwi-left i-source invoice image. Kwi-right i-web form eyenziwe ngexabiso ze-AI. Umxholo wakho kufuneka uye umntu kwi-lock. Yenza iiyure ezimbini: Ukuguqulwa kweTypo: Ukuguqulwa kwe-Lone Star Provisins Inc. kwi-Lone Star Provisions Inc. Ukuguqulwa kwamanani: Qhagamshela inani kwi-ONIONS YELLOW JBO ukusuka ku-5 kuya ku-50. Ukucinga Save Correction. Isinyathelo 3: Complete the Flywheel Ngoku, qhagamshelane kwi-terminal yakho, kwaye uqhagamshelane i-server ngokucindezela . I-script ikhawuleza ngokuzenzakalelayo i-step ye-flywheel: ukuhambisa umsebenzi wakho. Uya kuba le mveliso: Ctrl+C --- Server stopped. --- --- Exporting approved corrections to fine-tuning format... --- --- Data successfully exported to 'corrected_data.jsonl' --- --- QUICKSTART COMPLETE --- Uyavuma. Uyavuma nje umzila olupheleleyo we-data flywheel. I-Payoff: I-Training File epheleleyo Ukuvula ifayile. Uya kufumana ifayile entsha: Yinto i-prize. Yinto imiphumo esebenzayo yokusebenza kwakho, ifakwe kwaye ifakwe ngokufanelekileyo ukucaciswa kwimodeli ye-AI esebenzayo. examples/correction_deck_quickstart corrected_data.jsonl Thina ukhangela ngaphakathi. Iziquka umgca omnye 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}]}"}]}]}``` Yinto ingangena ngempumelelo, kodwa yinto elifanelekileyo ye-conversational iimodeli ezifana ne-Google Gemino kunye ne-OpenAI GPT series zithathanda ukucacisa. "Role": "user": Le nkxaso. Oku kuquka umfanekiso yokufaka (fileUri) kunye neengxelo ethetha i-AI. "i-role": "model": Lo mfuneko epheleleyo. Iziquka i-string ye-JSON kunye neengxaki zakho ezisetyenziswa. Thina siye kwenziwe ngempumelelo iiveki eziliqela zokusebenza lwezilwanyana kwi-example yokulungiselela kwizinga eliphezulu kunye ne-machine-readable. Ngoku, ndicinga ukwenza oku ngexesha le-100 iingcebiso. Okanye i-1000. Unxibelelanisa kuphela iingcebiso; unxibelelanisa ngokukhawuleza kwaye ngokukhawuleza i-dataset ebonakalisa le nkqubo epheleleyo yeengcebiso kwi-version elandelayo yeemveliso yakho. Yintoni elandelayo? Thina siphinde i-core loop ye-flywheel: Correct -> Capture -> Format for Training. Yinto yokuqala enamandla, kodwa inqubo engaphandle. Sifuna ukuba i-AI ufakele i-batch yayo, yaye ngoko siqhalisa umsebenzi yayo. Kodwa ukuba ungayifumana ngakumbi? Yintoni ukuba i-pipeline ingasebenza, ukufumana into engaziwayo, kwaye ngokufanelekileyo i-pause ukuba bafuna umntu ukunceda ixesha elifanelekileyo? Kule nqaku elandelayo kuleli mveliso, siya kuqhuba i-Human-in-the-Loop pipeline enokuthi uyazi xa kuxhomekeke kwimibuzo kwaye ayikwazanga ukufumana iinkcukacha.
