Indlela Umenzeli we-AI Wengxoxo Asebenza Ngayo Ngaphakathi

Indlela Umenzeli we-AI Wengxoxo Osebenza Ngayo Ngaphakathi (I-Architecture ye-OpenClaw)

Indlela umenzeli we-AI wengxoxo osebenza ngayo empeleni, itheni ngetheni? Lesi sihloko sivula ibhokisi elimnyama le-OpenClaw: kusukela ngesikhathi umlayezo weklayenti ufika ku-WhatsApp kuze kube umbhalo umenzeli awubhalayo ebuyela emuva. Kuzoba yizobuchwepheshe. Kufanele uma unquma i-architecture yomkhiqizo, uma uzothengisa isisombululo futhi ufuna ukuhlola ukujula, noma uma uthanda ukwazi okwenzekayo ngemuva kwengxoxo.

TL;DR: itheni ngalinye lidlula ezigabeni ezi-6 — ukufaka, ukuxazulula umongo, ukukhetha amakhono, ukunquma isenzo esilandelayo, ukusebenza nge-guard-rails, ukugcina inkumbulo. Umjikelezo wonke usebenza ku-<2 amasekhondi ku-edge ye-Cloudflare, ngaphandle kweseva eqinile.

---

Kungani i-architecture ibalulekile

Umenzeli wengxoxo obonakala esebenza ku-demo kodwa ephuka ekukhiqizeni ngokuvamile une**-nye yalezi zinkinga ezi-4**:

1. I-Latency ephezulu — iklayenti lilinde amasekhondi ayi-8 liphendule, ingxoxo iyafa.
2. Ukuphaphazela okungalawuleki — umenzeli uqamba intengo, isikhathi, inqubomgomo.
3. Umongo olahlekile — iklayenti libuyela ngemva kwezinsuku ezi-2 futhi umenzeli "uyakhohlwa" konke.
4. Izindleko ezingalawuleki — ingxoxo ngayinye ende igcwalisa i-prompt futhi ukhokha imali enkulu nge-token.

Ezi-4 ziyizinqumo ze-architecture, hhayi imikhawulo yemodeli. I-OpenClaw yakhiwa ukuze igweme ezi-4 — futhi indlela yokuqonda ukubheka umjikelezo wetheni.

---

Umjikelezo wetheni (izigaba ezi-6)

Cabanga ukuthi iklayenti lisanda kuthumela umlayezo othi "ngifuna ukubhukha ngoMgqibelo ekuseni". Yini eyenzekayo phakathi kwe-"received" nempendulo yomenzeli?

Isigaba 1 — Ukufaka (edge worker, <50ms)

Umlayezo we-WhatsApp ufika nge-webhook ye-Meta uqonde ku-Cloudflare Worker endaweni yokufinyelela (PoP) eseduze kakhulu ngokwejografi. EBrazil, lokhu kusho iSão Paulo noma iRio, i-latency yenethiwekhi < 20ms.

I-worker yenza izinto ezintathu:

1. Iqinisekisa ukusayina kwe-webhook (i-HMAC ngokumelene nemfihlo ye-WABA).
2. Ihlonza i-tenant ngenombolo yocingo lomamukeli (i-multi-tenant nge-to_number).
3. Iyalungisa i-payload — umsindo uba ukubhalwa, isithombe siba incazelo, indawo iba {lat,lng}, umbhalo uhlala unjalo.

Ekupheleni kwesigaba 1 unento ethi {tenant_id, conversation_id, user_message} elungele isinyathelo esilandelayo.

Isigaba 2 — Xazulula umongo (D1 + KV, ~80ms)

Umenzeli udinga izingcezu ezi-3 zomongo ngaphambi kokunquma:

• Umlando wakamuva wengxoxo (ama-N turnos afanelekayo akamuva).
• Inkumbulo yesikhathi eside yekhasimende (okuthandwayo, umlando wokuthenga, amanothi).
• Isimo se-agent (persona, amakhono anikwe amandla, imithetho).

Konke kuvela ku-D1 (i-SQLite esatshalaliswayo ye-Cloudflare). I-D1 ithatha indawo ye-Postgres/Mongo yendabuko — ayikho iseva yedatha okufanele uyigcine, ukufinyelela ngaphakathi kwama-ms ambalwa kusukela ku-worker, i-multi-tenant nge-tenant_id.

Iphuzu elibalulekile: asifaki ingxoxo yonke ku-prompt. I-Memory Manager v2 ye-OpenClaw (echazwe emibhalweni yethu yangaphakathi) ikhetha kuphela ama-turnos afanelekayo we-turno yamanje (ama-N akamuva + N wokufaneleka okuphezulu kwe-semantic). Lokhu kugcina izindleko ze-token zibikezeleka ngisho nezingxoxo ezingaphezu kuka-100+ turnos.

Isigaba 3 — Ukukhetha amakhono (policy engine, ~20ms)

I-agent ngayinye inayo isethi yamakhono atholakalayo — imisebenzi engayibiza. Izibonelo: consultar_calendario, criar_evento, gerar_link_pagamento, consultar_pedido, chamar_humano.

Njengoba kunikezwe umlayezo othi "quero marcar pra sábado de manhã", i-policy engine ihlunga:

• Amakhono ahambisanayo nenhloso etholiwe (ukuhlela).
• Amakhono avunyelwe kulesi sigaba sengxoxo (akuwona wonke amakhono atholakalayo njalo).
• Amakhono le tenant ewanikile amandla (ikhalenda ivela kuphela uma i-tenant ihlanganisile).

Ekugcineni unayo isethi encane yamakhono edluliselwa kumodeli — hhayi wonke ama-50 angaba khona, kuphela ama-4 anengqondo lapha. Lokhu kunciphisa kakhulu amathuba okuthi imodeli ibize ikhono elingalungile.

Isigaba 4 — Isinqumo (LLM call, 400-1200ms)

Manje imodeli ingena. I-OpenClaw yenza ikholi elilodwa ku-LLM yomngcele (Anthropic Claude, OpenAI GPT, Google Gemini — ilungiselelwa nge-tenant) nge:

• System prompt = persona ye-agent + imithetho + amakhono atholakalayo.
• History = ama-turnos akhethwe esigabeni 2.
• User message = umlayezo we-turno yamanje.

Imodeli iphendula okunye kwalezi zinto ezimbili:

• Impendulo yokugcina (umbhalo oqonde kuleyo khasimende).
• Tool call (isicelo sokusebenzisa ikhono elithile ngama-parameters).

Esibonelweni esithi "quero marcar pra sábado de manhã", imodeli ngokuvamile ibuyisela:

{
  "tool": "consultar_calendario",
  "args": { "date_range": "2026-04-19 06:00 to 12:00" }
}

Isigaba 5 — Ukusebenza nge-guard-rails (kuyahlukahluka, ~100-500ms)

Ikhono aligijimi kumodeli. Ligijima kukhodi wethu, okungu:

1. Qinisekisa amapharamitha (i-date_range inefomethi elilungile? ingaphakathi kwemithetho ye-tenant?).
2. Hlola imvume (leli-agent linelungelo lokubuza leli khalenda?).
3. Yenza ikholi (Google Calendar API kulokhu).
4. Buyisela umphumela ohleliwe kumodeli.

Kungani lokhu kubalulekile? Ngoba imodeli ayikaze yenze umphumela. Uma ikhalenda libuyisela [10h, 11h], yilokho kanye okuhamba kukholi elandelayo. Uma i-skill ihluleka, imodeli iyazi ukuthi ihlulekile. Awukho ubungozi bokuthi i-agent "iziqambe" ukuthi kunesikhathi ngo-9h lapho kungekho.

Ezimweni ezibandakanya ulwazi olubucayi (intengo, isikhathi, igama leklayenti), i-pipeline iphoqa tool call — ayivumeli imodeli ukuthi iphendule ngolwazi lwayo "lwangaphakathi". Lokhu kuqeda uhlobo lwe-hallucination oluvamile kakhulu kuma-agent ezohwebo.

Isigaba 6 — Impendulo nokugcinwa (~50ms)

Ngomphumela we-skill ezandleni, imodeli yenza ikholi lesibili — manje ukwakha impendulo yokugcina yeklayenti. Isib:

"Nginomgqibelo ngo-10h no-11h. Uyithanda eyiphi?"

Ngesikhathi esifanayo, i-worker:

1. Ithumela umlayezo emuva nge-API ye-WhatsApp.
2. Igcina ijika eliphelele (user + assistant + tool calls + ubude besikhathi) ku-D1.
3. Ibuyekeza inkumbulo yesikhathi eside uma ijika likhiqize iqiniso elisha (isib: "iklayenti likhetha umgqibelo").
4. Ikhipha isenzakalo sokuqaphela (isilinganiso se-latency, izindleko ze-token, izinga lokunyuswa).

Konke lokhu kugijima ngesikhathi esifanayo. Ukugcinwa akuvimbeli ukuthunyelwa komlayezo — iklayenti alilindi i-D1.

---

Kuphi ukuvikela ngokumelene ne-hallucination

I-agent eyi-hallucinate ekukhiqizeni ilahlekelwa ukwethembeka ngokushesha. I-OpenClaw inemigqa emine yokuvikela:

1. Umthombo weqiniso ophoqelelwe. Idatha yamaqiniso (intengo, isikhathi, igama) njalo ivela ku-skill, ayikaze ivele kumodeli yodwa.
2. Ukuqinisekisa kabili kudatha ebucayi. Ukubhukha kuyaqinisekiswa neklayenti ngaphambi kokugcinwa. Inkokhelo iyaqinisekiswa ngaphambi kokukhulula ukufinyelela.
3. Imithetho engemihle ecacile. I-persona ye-agent ngayinye ihlanganisa "ungakaze wenze X, Y, Z" — imodeli iyalalela.
4. Ukubuyela kumuntu. Lapho kungekho skill emboza umbuzo, i-agent ithi "ake ngihlole nethimba" futhi ivula ithikithi — ayiqagelisi.

Ezinhlolweni esizenze ezinyangeni eziyisithupha ezedlule (izingxoxo zangempela ezibuyekeziwe mathupha), izinga le-hallucination yeqiniso lahlala ngaphansi kwe-0,3% yamajika — futhi cishe zonke izimo zazingenxa ye-config (i-tenant yakhohlwa ukunika amandla i-skill efanele), hhayi iphutha lemodeli.

---

Izindleko ngengxoxo

Ukwakheka okuhle akubonakali uze ubheke i-invoice. Njengoba ukuphendula ngakunye kwenza izingcingo ze-LLM ezingu-1-2 + ama-lookups ku-D1, izindleko ezijwayelekile zengxoxo ephelele (imijikelezo engu-10-15) ziba: