Riba:
- Bayyana manufar alama, bambancin shigarwa/fitarwa da alamar alama.
- Za a iya ƙididdige farashin buƙatu da aikin aiki na wata-wata daga adadin alamun da farashin naúra
- Za a iya kwatanta tasirin zaɓin samfuri da tsayin gaggawa akan farashi
Ba za ku iya gina mafita a sikeli ba tare da fahimtar tattalin arziƙin APIs na LLM ba. A demo yana gudana sau ɗaya; Babban abu shine iya yin hasashen abin da lissafin zai kasance lokacin da aka yi dubban kira kowane wata. A cikin wannan naúrar mun tsara bangaren kuɗi: menene alamar, dalilin da yasa aka sanya farashin shigarwa da fitarwa daban-daban, yadda ake ƙididdige farashin buƙatun, da yadda ake tsara kasafin kuɗin aiki na wata. Wannan bayanin yana ba ku damar auna dawo da dabarun ingantawa (caching, zaɓin samfuri, tsari) a cikin raka'a masu zuwa.
Menene Token?
Token ita ce mafi ƙanƙanta naúrar da ƙirar ke sarrafa rubutu. Kalma ba koyaushe alama ce ba; Alama yawanci sashi ne na kalma. Kusan magana, a cikin Ingilishi, alamar 1 shine ≈ haruffa 4 ≈ 0.75 kalmomi. A Turkanci da lamba, rabon ya bambanta: galibi ana raba kalmomin Turkanci zuwa ƙarin alamu fiye da na Ingilishi saboda ƙaƙƙarfan tsari da haruffa. Sabili da haka, ya zama dole a auna adadin alamun tare da kayan aikin ƙidayar alamar mai bayarwa maimakon zato ta ido.
Tokenization (tsarin rarraba rubutu zuwa alamomi) na iya bambanta ga kowane samfuri. Wannan yana da sakamako mai amfani guda biyu: (1) Rubutu ɗaya na iya ba da lambobi daban-daban na alamu a cikin nau'i daban-daban; (2) Hasashen da aka yi tare da tokenizers daga wasu masu samarwa (misali ɗakin karatu na tiktoken na OpenAI) ba zai zama daidai ba ga Claude - yi amfani da tip ɗin kirga alama don ƙirar da kuke amfani da ita.
Alamu: "Game da alamu nawa?" Kar a amsa tambayar a makance. Ƙaddamar da rubutun wakilci ta hanyar API ɗin ƙidayar alama; Tushen yanke shawara na kasafin kuɗi akan aunawa.
Alamomin shigarwa da fitarwa
Daftar ta ƙunshi abubuwa biyu:
- Alamu na shigarwa: Duk abin da kuka aika zuwa samfurin - faɗakarwar tsarin, balaguron da ya wuce, saƙon mai amfani, takaddun shaida idan akwai. Ana sarrafa waɗannan gaba ɗaya.
- Alamar fitarwa: Amsar da samfurin ya samar. Ga kowane alamar fitarwa, ƙirar tana yin lissafin mataki-mataki.
Tare da yawancin masu samarwa, fitarwa yana da tsada sau da yawa fiye da shigarwar. Dalilin yana da sauƙi: karanta shigarwar gaba ɗaya yana da rahusa fiye da samar da alamar fitarwa. Sanin wannan asymmetry yana bayyana dalilin da yasa ingantawa kamar "na buƙatar taƙaitaccen amsoshi" suna da tasiri sosai.
Samfuran farashin (kowane alamar miliyan 1, USD)
Teburin da ke ƙasa nuni ne; Farashi na iya canzawa akan lokaci, da fatan za a tabbatar da jerin abubuwan mai ba ku na yanzu.
samfurin aji
samfurin samfurin
Shiga ($/1M)
Fitowa ($/1M)
Yawan amfani
sauri/arha
Haiku 4.5
1.00
5.00
Rabewa, lakabi, taƙaitaccen bayani
daidaitacce
sonnet 5
3.00
15.00
Babban manufa, coding, aikin wakili
mai karfi
Shafin 4.8
5.00
25.00
Maɗaukakiyar tunani, ayyuka masu tsayi
A kowane aji, fitarwa shine sau 5 na shigarwa; Bugu da ƙari, har ma da shigar da samfurin mai ƙarfi shine sau 5 da shigar da samfurin arha. Waɗannan gatura guda biyu (shigarwar ↔ fitarwa da ajin samfuri) sun tsara tsarin yanke shawarar farashin ku.
Yadda za a lissafta Kudin?
Tsarin tsari mai sauƙi ne:
farashi = (input_token / 1,000,000) × farashin shigarwa + (fitarwa_token / 1,000,000) × farashin_price
Misali lissafi. Buƙata tare da Sonnet 5: Alamu na shigarwa 1,500, alamun fitarwa 400.
shigarwar = 1,500 / 1,000,000 × 3.00 = fitarwa $ 0.0045 = 400 / 1,000,000 × 15.00 = $ 0.0060 duka = $ 0.0105 (kimanin 1 cent)
Kira ɗaya yayi kama da arha. Amma ninka da ƙara: 20,000 kira kowace rana → $210 kowace rana, ~ $6,300 kowane wata. Anan ne ma'auni ya shigo cikin wasa.
Samfurin Kasafin Kudi na wata-wata
Don cire farashin kayan aiki na wata-wata, yi amfani da wannan samfuri:
1) Matsakaicin alamar shigar da buƙatun: ......2) Matsakaicin alamar fitarwa ta kowace buƙata: ......3) Yawan buƙatun kowace rana: ......4) Kwanakin aiki a kowane wata: ......5) Farashin kowace buƙata = (1) / 1M × farashin shigarwa = (2) / 1M × fitarwa_price6) Farashin kowane wata = (5) × (3) × (4) × (4)
Zuba wannan ƙirar a cikin maƙunsar rubutu da ganin yadda jimlar ke gudana lokacin da ka canza ƙirar ya ƙunshi zaɓin ƙirar (raka'a 5) da cache (naúrar 6) yanke shawara.
Gajarta Gaggawa tare da Samfura masu Kwafi
Yawancin farashi yana zuwa daga dogon lokaci da ba dole ba da kuma fitar da kayan aiki da ba dole ba. Samfuran da ke ƙasa suna ba da tanadi kai tsaye.
# Iyakance tsayin fitarwa. Amsa tare da iyakar abubuwa 3. Ƙara dalili ko jumlar gabatarwa.
# Koma filin da aka nema kawai Mayar da JSON mai zuwa kawai, kar a ƙara kowane rubutu:{"categori": "...", "gaggawa": "ƙananan|matsakaici|high"}
# Cire mahallin da ba dole ba Cire kwanan wata da adadin kawai daga rubutu mai zuwa. Kar a maimaita dukkan rubutun.Text: """"{{text}}"""
# Takaita dogon jawabi (ajiya na shigar da bayanai) Takaita wannan jawabin cikin abubuwa 5. Zan yi amfani da wannan taƙaitawar maimakon cikakken abin da ya gabata a zagaye na gaba. Jawabi: """{{ past}}"""
Rauni mai ƙarfi / ƙarfi mai ƙarfi (dangane da farashi)
# WEAK (sakin fitarwa, tsada) Yi nazarin wannan buƙatar tallafin kuma rubuta mani cikakken bita.
# KARFI (ƙaddara fitarwa, arha da abin iya faɗi) Rarraba wannan buƙatar tallafin. Kawai mayar da JSON mai zuwa:{"category":" daftari | fasaha | maidawa | sauran"," gaggawa":"ƙananan|matsakaici|high"}Kada a rubuta kwatance.
Sigar mai rauni tana samar da alamun fitarwa mai yiwuwa 500; karfi version ~ 15. Saboda fitarwa yana da tsada, wannan babban bambanci ne ga kowane kira kuma yana ƙaruwa tare da ƙara.
Mini Cases guda uku
Case 1 - Kudin ɓoye na dogon lokaci. Kamar yadda tsarin lissafin kuɗi ya tsara kowane daftari, ya ƙara "littafin doka" mai shafuka 40 a matsayin shigarwa ga kowace buƙatu: ~ 12,000 alamun shigarwa a kowace buƙata. Tare da Sonnet 5 12,000/1M × 3 = $0.036 an shigar dashi. Biyan kuɗi 5,000 kowace rana → $180 kowace rana. Ta hanyar adana littafin ƙa'ida (raka'a 6) farashin shigarwa ya ragu da ~90%.
Case 2 - Sakamakon rage girman samfurin. Ƙungiya ɗaya tana yin alama mai sauƙi "tabbatacce / mara kyau" tare da Opus 4.8: shigarwar 300 + 10 alamun fitarwa. Kudin Opus 300/1M×5 + 10/1M×25 = $0.00175. Canjawa zuwa Haiku, 300/1M × 1 + 10/1M × 5 = $0.00035 — 5x mai rahusa, bambamcin daidaito ba shi da iyaka. A kan kira miliyan 3 kowane wata, bambancin shine $5,250 → $1,050.
Case 3 - Sakin fitarwa. Lokacin da ƙungiyar tallace-tallace ta samar da bayanin samfur, ba ta kafa iyaka akan fitarwa ba; samfurin wani lokaci ya ce alamun 1,500. Lokacin da na ƙara umarnin "mafi girman kalmomi 60", matsakaicin fitarwa ya ragu daga alamun 900 zuwa 90. Tun da bugu yana da tsada, an rage lissafin wata-wata da kashi uku, kuma rubutun ya zama da amfani sosai.
Kuskuren gama gari
- Yin hasashen alamar da ido: Kuna iya yin kuskure musamman a cikin Turanci da lamba. Auna.
- Zaton shigarwa da fitarwa iri ɗaya ne: Fitowa yawanci ya fi tsada; Yawancin haɓakawa yana zuwa daga rage fitarwa.
- Kar a yaudare ku da arha kira guda: An yanke shawarar da girma. $0.01 × miliyan = $10,000.
- Hasashen tare da alamar mai bada sabis: Yana ba da sakamakon da ba daidai ba; Yi amfani da kayan aikin ƙidayar alamar samfurin.
- Unlimited fadada tarihin tattaunawa: Ana ƙara kowane zagaye zuwa shigarwa; takaita cikin doguwar tattaunawa.
- Tsayar da 'max_tokens' ba dole ba: Yana ɓoye tsarin kasafin kuɗi da haɗarin yankewa; Ba da ƙima na gaske.
Zurfafa: Tagar Magana da Dogon Shigar Kuɗi
Yana da mahimmanci don ganin yadda farashin ke tattarawa "a cikin tattaunawar" ba kawai "kowace bukata ba." Adadin adadin rubutun da samfurin zai iya aiwatarwa ana kiransa taga mahallin; Jimlar shigarwa da fitarwa dole ne su dace cikin wannan taga. Samfuran zamani suna ba da manyan tagogi (daruruwan dubbai, har ma da miliyoyin alamu), amma wannan ba yana nufin za ku iya “cika shi marar iyaka” - duk abin da kuka saka a cikin taga ana cajin shi azaman shigarwa.
Tarko a cikin doguwar tattaunawa shine: tare da kowane sabon zagaye za ku sake aika tarihin gaba ɗaya (rashin ƙasa a cikin raka'a 1). A cikin zance na zagaye 20, buƙatu ta 20 tana ɗaukar duka zagaye 19 na farko azaman shigarwa. Don haka, yayin da tattaunawar ta yi tsayi, farashin kowane buƙatu yana girma gabaɗaya maimakon layi. Tattaunawa na zagaye 50 tare da mataimakin wakili na iya haifar da farashin shigarwa sau da yawa a zagaye na farko.
Akwai hanyoyi guda biyu don sarrafa wannan. Na farko shine maimaitawa: matsar da tsofaffin zagayen zuwa cikin toshe maimaituwa guda ɗaya, tare da kiyaye ƴan zagayen ƙarshe danye. Na biyu shine saurin caching (raka'a 6): karanta ƙayyadaddun mahallin a kashi goma na farashin maimakon maimaita sarrafa shi akan cikakken farashi. Tare, suna da matukar muhimmanci rage lissafin akan dogon aiki mai tsananin mahalli. Don haka token Economics shine game da ƙirar gabaɗayan zaman, ba buƙatu ɗaya ba.
A takaice
Token ita ce mafi ƙanƙanta naúrar da ake sarrafa rubutu; shigarwa da fitarwa ana farashi daban-daban, kuma kayan fitarwa galibi ya fi tsada sosai. Farashin shine adadin alamun da aka ninka ta farashin naúrar, kuma an yanke shawarar gaske ta ƙarar. Gajarta hanzari, iyakance fitarwa, da zabar samfurin mafi sauƙi wanda ke aiwatar da aikin shine mafi girman levers kai tsaye waɗanda ke rage farashin sau da yawa.
Aikin aikace-aikace
Zaɓi wani aiki na kanku. (1) Ƙayyade adadin shigarwar da ƙididdiga alamun fitarwa don saurin wakilci (auna tare da kayan aikin ƙidayar alama idan zai yiwu). (2) Yi ƙididdige farashin kowane buƙatu don azuzuwan ƙirar uku. (3) Yi ƙididdige adadin buƙatun ku na yau da kullun kuma ku fitar da kasafin kuɗi na wata-wata don ƙirar uku. (4) Ƙara umarni don taƙaita fitarwa kuma lura da tanadin da ake sa ran.
jerin abubuwan dubawa
- [ ] Zan iya bayyana ma'anar alamomi kuma wannan alamar ta bambanta dangane da samfurin.
- [ ] Na san dalilin da yasa aka sanya farashin shigarwa da fitarwa daban-daban.
- [ ] Zan iya lissafta farashin buƙatu tare da dabarar.
- [ ] Zan iya ƙirƙirar kasafin kuɗi na wata-wata don nauyin aiki ta amfani da samfuri.
- [ ] Zan iya nuna tare da misali fa'idar rage fitarwa da raguwar ƙira.