Yadda Wakili na AI na Tattaunawa Ke Aiki a Ciki

Yadda Wakili na AI na Tattaunawa Yake Aiki a Ciki (Tsarin OpenClaw)

Yadda wakili na AI na tattaunawa yake aiki a aikace, juyi-juyi? Wannan rubutu yana buɗe akwatin baƙar fata na OpenClaw: daga lokacin da saƙon abokin ciniki ya isa WhatsApp har zuwa rubutun da wakili ya rubuta a mayarwa. Zai zama na fasaha. Yana da daraja idan kana yanke shawarar tsarin samfur, idan za ka sayi bayani kuma kana son kimanta zurfin, ko kuma kana son sanin abin da ke faruwa a bayan tattaunawar.

TL;DR: kowane juyi yana wucewa ta matakai 6 — ingest, warware mahallin, zaɓi ƙwarewa, yanke shawarar mataki na gaba, aiwatar da guard-rails, adana ƙwaƙwalwa. Duk zagayowar tana gudana a cikin <2 daƙiƙa a gefen Cloudflare, ba tare da sabar da aka kafa ba.

---

Me yasa tsarin yake da mahimmanci

Wakilin tattaunawa wanda yake kama da yana aiki a cikin demo amma ya karye a cikin samarwa yawanci yana da ɗaya daga cikin waɗannan matsaloli 4:

1. Latency mai girma — abokin ciniki yana jira daƙiƙa 8 don amsa, tattaunawar ta mutu.
2. Ruɗi da ba a sarrafa ba — wakili yana ƙirƙira farashi, lokaci, manufa.
3. Mahallin da aka rasa — abokin ciniki ya dawo bayan kwanaki 2 kuma wakili ya "manta" da komai.
4. Farashi da ba a sarrafa ba — kowace tattaunawa mai tsawo tana cika tambaya kuma kana biyan dukiya a cikin token.

Duk 4 zaɓuɓɓukan tsarin ne, ba iyakoki na samfurin ba. An gina OpenClaw don guje wa 4 — kuma hanyar fahimta ita ce kallon zagayowar juyi.

---

Zagayowar juyi (matakai 6)

Ka yi tunanin cewa abokin ciniki ya aika saƙon "ina son yin alƙawari don Asabar da safe". Me ke faruwa tsakanin "an karɓa" da amsar wakili?

Mataki na 1 — Ingest (edge worker, <50ms)

Saƙon WhatsApp yana shigowa ta hanyar webhook na Meta kai tsaye a cikin Cloudflare Worker a wurin kasancewa (PoP) mafi kusa da yanki. A Brazil, wannan yana nufin São Paulo ko Rio, latency na hanyar sadarwa < 20ms.

Ma'aikacin yana yin abubuwa uku:

1. Yana tabbatar da sa hannun webhook (HMAC akan asirin WABA).
2. Yana gano mai haya ta lambar wayar mai karɓa (multi-tenant ta to_number).
3. Yana daidaita payload — sauti ya zama rubutu, hoto ya zama bayanin, wuri ya zama {lat,lng}, rubutu yana kasancewa kamar yadda yake.

A ƙarshen mataki na 1 kana da abu {tenant_id, conversation_id, user_message} shirye don mataki na gaba.

Mataki na 2 — Warware mahallin (D1 + KV, ~80ms)

Wakili yana buƙatar guda 3 na mahallin kafin ya yanke shawara:

• Tarihin kwanan nan na tattaunawa (N na ƙarshe masu dacewa).
• Ƙwaƙwalwar dogon lokaci na abokin ciniki (abubuwan da ya fi so, tarihin sayayya, bayanin kula).
• Matsayin wakili (siffa, ƙwarewa da aka kunna, ƙa'idodi).

Duk sun fito ne daga D1 (SQLite da aka rarraba na Cloudflare). D1 ya maye gurbin Postgres/Mongo na al'ada — babu sabar bayanan da za a kula da ita, samun dama cikin 'yan ms daga ma'aikaci, multi-tenant ta hanyar tenant_id.

Mahimmin batu: ba ma ɗaukar dukkan tattaunawar a cikin faɗakarwa. Memory Manager v2 na OpenClaw (wanda aka bayyana a cikin takardun mu na ciki) yana zaɓar kawai jujjuyawar da suka dace da jujjuyawar yanzu (N na ƙarshe + N masu girman dacewa ta ma'ana). Wannan yana kiyaye farashin alama a kan tsinkaye ko da a cikin tattaunawa sama da 100+ jujjuyoyi.

Mataki na 3 — Zaɓin ƙwarewa (injin manufa, ~20ms)

Kowane wakili yana da saitin ƙwarewa da ake samu — ayyuka da zai iya kira. Misalai: consultar_calendario, criar_evento, gerar_link_pagamento, consultar_pedido, chamar_humano.

Idan aka ba da saƙo "quero marcar pra sábado de manhã", injin manufa yana tace:

• Ƙwarewar da ta dace da nufin da aka gano (tsara lokaci).
• Ƙwarewar da aka yarda da ita ga wannan matakin tattaunawa (ba duk ƙwarewa ake samunta koyaushe ba).
• Ƙwarewar da wannan mai haya ya kunna (kalanda kawai yana bayyana idan mai haya ya haɗa).

A ƙarshe kana da ƙaramin rukunin ƙwarewa da aka wuce ga ƙirar — ba 50 da zasu yiwu ba, kawai 4 da suka dace a nan. Wannan yana rage yuwuwar ƙirar ta kira ƙwarewar da ba daidai ba sosai.

Mataki na 4 — Yanke shawara (kiran LLM, 400-1200ms)

Yanzu ƙirar ta shigo. OpenClaw yana yin kira guda ɗaya zuwa LLM na gaba (Anthropic Claude, OpenAI GPT, Google Gemini — mai saituwa ta mai haya) tare da:

• System prompt = siffar wakili + ƙa'idodi + ƙwarewar da ake samu.
• History = jujjuyoyin da aka zaɓa a mataki na 2.
• User message = saƙon jujjuyar yanzu.

Ƙirar tana amsa ɗaya daga cikin abubuwa biyu:

• Amsa ta ƙarshe (rubutu kai tsaye ga abokin ciniki).
• Kiran kayan aiki (buƙata don aiwatar da takamaiman ƙwarewa tare da sigogi).

A misalin "quero marcar pra sábado de manhã", ƙirar yawanci tana dawo da:

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

Mataki na 5 — Aiwatarwa tare da tsare-tsare (mai canzawa, ~100-500ms)

Ƙwarewar ba ta gudana a cikin ƙirar ba. Tana gudana a cikin lambar mu, wanda:

1. Yana tabbatar da sigogin (date_range yana da tsarin da ya dace? yana cikin ƙa'idodin tenant?).
2. Yana duba izini (wannan wakili yana da 'yancin duba wannan kalanda?).
3. Yana aiwatar da kira (Google Calendar API a wannan yanayin).
4. Yana dawo da sakamako mai tsari ga ƙirar.

Me yasa wannan yana da mahimmanci? Domin ƙirar ba ta taɓa ƙirƙira sakamakon ba. Idan kalandar ta dawo da [10h, 11h], ainihin wannan ne zai je kira na gaba. Idan fasaha ta gaza, ƙirar ta san cewa ta gaza. Babu haɗarin wakili ya "ƙirƙira" cewa akwai lokaci da ƙarfe 9 lokacin da babu.

Ga yanayin da suka shafi bayanai masu mahimmanci (farashi, lokaci, sunan abokin ciniki), bututun yana tilasta tool call — ba ya barin ƙirar ta amsa daga "iliminta" na kanta. Wannan yana kawar da nau'in ruɗi mafi yawa a cikin wakilan kasuwanci.

Mataki na 6 — Amsa da adanawa (~50ms)

Tare da sakamakon fasaha a hannunsa, ƙirar tana yin kira na biyu — yanzu don samar da amsa ta ƙarshe ga abokin ciniki. Misali:

"Ina da Asabar da ƙarfe 10 da 11. Wanne ka fi so?"

A lokaci guda, ma'aikaci:

1. Yana aika saƙon baya ta hanyar API na WhatsApp.
2. Yana adana cikakken juyawa (mai amfani + mataimaki + kiran kayan aiki + tsawon lokaci) a cikin D1.
3. Yana sabunta ƙwaƙwalwar dogon lokaci idan juyawar ta samar da sabon gaskiya (misali: "abokin ciniki ya fi son Asabar").
4. Yana fitar da abin lura na gani (ma'aunin jinkiri, farashin alama, ƙimar haɓakawa).

Duk wannan yana gudana a lokaci guda. Adanawa ba ya toshe aika saƙon — abokin ciniki ba ya jiran D1.

---

Ina kariyar ruɗi take

Wakili da ke ruɗi a cikin aiki yana rasa amana da sauri. OpenClaw yana da layukan kariya 4:

1. Tushen-gaskiya da aka tilasta. Bayanan gaskiya (farashi, lokaci, suna) koyaushe suna fitowa daga fasaha, ba daga ƙirar kaɗai ba.
2. Tabbatarwa sau biyu akan bayanai masu mahimmanci. Ana tabbatar da tsari tare da abokin ciniki kafin a adana. Ana tabbatar da biyan kuɗi kafin a ba da damar shiga.
3. Ƙa'idodi masu ƙin yarda a sarari. Halin kowane wakili ya haɗa da "kada ka taɓa ƙirƙira X, Y, Z" — ƙirar tana biyayya.
4. Komawa ga ɗan adam. Lokacin da babu wata fasaha da ta rufe tambayar, wakili yana cewa "bari in duba tare da ƙungiyar" kuma yana buɗe tikiti — ba ya yin zato.

A cikin binciken da muka yi a cikin watanni 6 da suka gabata (tattaunawar gaske da aka sake duba da hannu), ƙimar ruɗin gaskiya ta kasance ƙasa da 0.3% na juyawa — kuma kusan dukkan shari'o'in sun kasance saboda tsari (tenant ya manta kunna fasaha mai dacewa), ba kuskuren ƙirar ba.

---

Farashin kowace tattaunawa

Tsarin gine-gine mai kyau ba a gani ba har sai ka duba takardar kuɗi. Ganin cewa kowane juyawa yana yin kira 1-2 na LLM + bincike a cikin D1, farashin da aka saba don cikakken tattaunawa (juyawa 10-15) yana kasancewa a: