Isifundo se-BookLeme: Hlanganisa incwadi ibe yi-Agent Skill efakekayo futhi eqinisekisekayo

Iyini i-BookLeme

I-BookLeme, ebizwa ngokuthi “读了么” ngesiShayina, iyithuluzi eliguqula izincwadi zibe ama-Agent Skill. Ayigxili ekukhiqizeni izifinyezo ezivamile, kodwa ikhipha indlela yokusebenza esencwadini, njengokulandelana kwemibuzo, izimo zokwahlulela, imithetho yobufakazi, izimo zenqubo kanye nemingcele lapho ingasebenzi khona, bese ikuhlela kube yimikhiqizo engafakwa, ibizwe futhi iqinisekiswe.

Le phrojekthi ihlukanisa ukucutshungulwa kwezincwadi kube yisikhathi sokuhlanganisa nesikhathi sokusebenza: ngesikhathi sokuhlanganisa kufundwa futhi kucutshungulwe incwadi yokuqala, kanti ngesikhathi sokusebenza kulayishwa kuphela izingcezu zolwazi ezikhiqiziwe lapho kudingeka. Lokhu kugwema ukufunda kabusha yonke incwadi njalo lapho kubizwa i-Skill.

Qaphela: “Ukukhipha isakhiwo esikhundleni sesifinyezo” kuyisinqumo sokwakheka kobunjiniyela esisetshenziswa yiphrojekthi, kodwa i-README ikumaka ngokusobala njengombono ongakaqinisekiswa yiphrojekthi, hhayi isiphetho esesifakazelwe.

Yini ekhiqizwa yi-BookLeme

I-BookLeme inikeza izinjini ezimbili zokukhiqiza. Zabelana ngokumelwa okuphakathi kwezigaba zokukhipha nokuqinisekisa, bese zihlukana kuphela esigabeni sokuhumusha; ngakho-ke uma kukhethwa ukukhiqiza izinhlobo ezimbili zemikhiqizo, akudingeki kucutshungulwe incwadi kabili.

Iphakethe lolwazi knowledge_pack

Iphakethe lolwazi liyisizindalwazi solwazi esihlelekile esingabhekwa, esiqukethe ikakhulukazi ithebula lokuhambisa le-Topic Index, izingcezu zezinhlaka ezilayishwa lapho kudingeka, ishidi lokubheka ngokushesha izinqumo kanye nohlu lwemithombo. Lifanele ukuphendula imibuzo efana nokuthi “isetshenziswa kanjani inhlaka ethile esencwadini”.

Behavior Skill behavior_skill

I-Behavior Skill isetshenziselwa ukuqhuba inqubo ephelele, futhi ingaqukatha izincazelo zezindima, imingcele eqinile, umshini wezimo, inkontileka yokuphumayo enegama, amaphethini angalungile kanye nezivikelo zemizwa. Ifanele ukwenza imisebenzi yezinyathelo eziningi, njengezingxoxo zokuqeqesha, hhayi ukubheka ulwazi kuphela.

Izici eziyinhloko

• Ukuqondisa okuzenzakalelayo kwezinjini ezimbili: Kusetshenziswa izici eziyi-12 ezinqumayo zohlobo ukukhetha phakathi kwephakethe lolwazi ne-Behavior Skill, futhi injini ingacaciswa ngesandla.
• Ama-plugin ohlobo axhumekayo: Lapho kwengezwa uhlobo olusha, kunganezelwa ifayela elilodwa le-Python ohlwini lwama-plugin omsebenzisi ngaphandle kokushintsha umugqa oyinhloko wokusebenza.
• Ukuqinisekiswa kokwethembeka kwezwi nezwi: Izingcaphuno kufanele zibe yizicucu eziqondile zombhalo wokuqala, futhi zihlolwe ngezinkomba eziningi ezingalinganiswa.
• Isango eliqinile lamalungelo obunikazi: Kwenziwa ukuhlolwa okuqinisekile kwamalungelo obunikazi ngaphambi kokufaka nokushicilela; ama-Skill akhiqizwe ngezincwadi zezinkampani zangaphandle agcinwa eyimfihlo ngokuzenzakalelayo.
• Akukho ukuncika kwabanye ngesikhathi sokusebenza: Ukuqaliswa okuyinhloko kusebenzisa kuphela umtapo ojwayelekile we-Python.
• Ukungaphindaphindeki ezingeni lesigaba: Umugqa wokusebenza usebenzisa i-SHA-256 ukusekela ukuqhubeka; uma imiphumela yesigaba esithile isuswa, kudingeka kuqhutshwe kabusha leso sigaba kuphela.

Ukufaka nokulungisa imvelo

I-BookLeme idinga i-Python 3.9 noma inguqulo ephakeme. Ngemva kokukopisha i-repository, ingaqhutshwa ngqo nge-entry point yemojuli ye-Python ngaphandle kokufaka kuqala okuncikile kwabanye ngesikhathi sokusebenza.

git clone https://github.com/bookleme/bookleme.git
cd bookleme
python3 -m bookleme --version

Uma umyalo ukwazi ukukhipha ulwazi lwenguqulo, kusho ukuthi imvelo eyisisekelo isiyatholakala. I-repository ithi isekela amafomethi okufaka afana ne-PDF, EPUB, DOCX, Markdown, HTML, RTF ne-MOBI; ngezansi kusetshenziswa umbhalo wencwadi we-Markdown ohlanganisiwe oza nephrojekthi ukubonisa yonke inqubo.

Ukuhlanganisa okokuqala: ukukhiqiza i-Skill kusuka kumbhalo wencwadi

Qalisa i-subcommand ethi run, usebenzise --input ukucacisa umbhalo wencwadi, bese usebenzisa --title ukusetha isihloko:

python3 -m bookleme run \
  --input evals/cases/synthetic/coach_book.md \
  --title "Incwadi yami yokuqala"

Ngokuzenzakalela, i-BookLeme izohlaziya izici zohlobo lombhalo bese ikhetha ngokuzenzakalelayo injini efanele yokukhiqiza. Inqubo yokuhlanganisa idlula ezigabeni zokubhalisa okokufaka, ukukhipha umbhalo, ukuhlukanisa umbhalo wesiShayina, ukuqondisa kohlobo, ukukhipha amakhono ayisisekelo, ukuqinisekisa ukwethembeka, ukuhumusha, ukuqinisekisa okuphelele, ukufaka nokuhlola kokuqala.

Isincomo: Lapho uzama okokuqala, sebenzisa kuqala umbhalo wokuhlola ohlanganisiwe onikezwe yi-repository. Ngaphambi kokucubungula izincwadi zezinkampani zangaphandle, qinisekisa ukuthi unamalungelo afanele okuzisebenzisa futhi uyayiqonda imingcele yokusabalalisa imikhiqizo ekhiqiziwe.

Ukukhetha injini yokukhiqiza

Sebenzisa --engine ukulawula uhlobo lokuphumayo:

python3 -m bookleme run \
  --input evals/cases/synthetic/coach_book.md \
  --title "Isibonelo sengxoxo yokuqeqesha" \
  --engine behavior_skill

Amanani atholakalayo afaka:

• auto: Vumela i-router ikhethe ngokuzenzakalelayo.
• knowledge_pack: Khiqiza iphakethe lolwazi elihlelekile kuphela.
• behavior_skill: Khiqiza i-Behavior Skill esebenzayo kuphela.
• both: Khiqiza zombili izinhlobo zemikhiqizo ngesikhathi esisodwa.

Uma ufuna ukuvala i-router bese uvumela zombili izinjini zikhiphe imiphumela, sebenzisa --no-router:

python3 -m bookleme run \
  --input evals/cases/synthetic/coach_book.md \
  --title "Isibonelo sezinjini ezimbili" \
  --no-router

Ukucacisa i-plugin yohlobo

I-plugin yohlobo inquma umshini wezimo womkhiqizo, inkontileka yokuphumayo, imingcele eqinile kanye nendlela yokuzwakalisa. Ama-plugin asevele asetshenzisiwe okwamanje afaka:

• coach_dialogue: Isetshenziselwa umqeqeshi noma ingxoxo yomuntu naye, isekela umshini wezimo wombuzo nempendulo, izivikelo zemizwa, ukuphazamiseka nokuqhubeka kwengxoxo.
• knowledge_reference: Isetshenziselwa ukubheka ngokushesha izinhlaka, ihambisa okuqukethwe nge-Topic Index; uma ihlangabezana nolwazi olungekho encwadini, kufanele yenqabe ukuphendula.

professional_analysis kanye decision_checklist kusengokuqukethwe komgwaqo wesikhathi esizayo we-M4, ngakho okwamanje akufanele kubhekwe njengezici eseziqediwe.

Ungasebenzisa --genre ukuphoqa ukukhetha i-plugin esivele yenziwe:

python3 -m bookleme run \
  --input evals/cases/synthetic/coach_book.md \
  --title "Ingxoxo yokuklama impilo" \
  --engine behavior_skill \
  --genre coach_dialogue

Ukuqinisekisa imikhiqizo ekhiqiziwe

Ngemva kokuqedwa kokukhiqiza, kufanele usebenzise umyalo wokuqinisekisa onikezwe yiphrojekthi ukuhlola imikhiqizo. I-README isebenzisa examples/coach njengesibonelo:

python3 -m bookleme validate examples/coach

Lesi sigaba sisebenzisa isiqinisekisi esisemthethweni, i-lint yangaphakathi kanye nesango eliqinile lamalungelo obunikazi. I-BookLeme ihlukanisa ukwethembeka kube yizinkomba ezingahlolwa, ezihlanganisa izinga lokuhlaziya izingcaphuno, izinga lokwephulwa kwezwi nezwi, izinga lokwenqaba izibonelo ezingezinhle, izinga lokutholwa kwamagama aqalayo, izinga lokuhambisana komkhiqizo, umkhawulo wokudlulelana kwama-shingle kanye nezinga lokumbozwa.

Le phrojekthi iphinde igcine inkomba ye-heuristic ethi groundedness, kodwa i-README ichaza ngokusobala ukuthi i-LLM judge ayikaxhunywa; ngakho okwamanje leli nani libikwa kuphela futhi alisetshenziswa njengesango eliqinile.

Ukuqalisa izivivinyo zokubuyela emuva

Ngemva kokushintsha ikhodi eyinhloko, ama-plugin noma izifanekiso, sebenzisa izivivinyo eziphelele:

python3 -m unittest discover -s tests -v

Uma uthuthukisa i-plugin, ungahlola ne-schema yezimo futhi uqinisekise ifayela le-plugin ngokuhlukile:

python3 -m bookleme schema --state
python3 -m bookleme plugin validate bookleme/plugins/builtin/coach_dialogue.py

Ukufaka imikhiqizo nokumemezela amalungelo okusebenzisa

Ngokuzenzakalelayo, umyalo wokukhiqiza awubhali imikhiqizo emthonjeni wama-Skill. Lapho kudingeka ukufakwa, kufanele ufake ngokusobala --install. Ungasebenzisa futhi --user-rights ukumemezela isimo samalungelo okuqukethwe okufakwayo:

python3 -m bookleme run \
  --input path/to/book.md \
  --title "I-Skill yolwazi lwangaphakathi" \
  --engine both \
  --install \
  --user-rights owned

--user-rights isekela amanani alandelayo:

• owned: Umsebenzisi unamalungelo okuqukethwe.
• licensed: Kutholwe imvume efanele.
• public_domain: Okuqukethwe kusemphakathini womphakathi.
• unknown: Isimo samalungelo asaziwa.

Isimemezelo samalungelo asithathi indawo yokwahlulela ngokomthetho. Ilayisensi ye-MIT ihlanganisa kuphela ikhodi ye-BookLeme converter nezincazelo zama-Skill, ayihlanganisi izincwadi ezifakwayo futhi ayinikezi ngokuzenzakalelayo umsebenzisi ilungelo lokusabalalisa imikhiqizo etholakala kuzo.

Ukuqonda isango eliqinile lamalungelo obunikazi

Imikhiqizo ye-BookLeme ibekwe njengamanothi ahlelekile atholakala kokwengeziwe, hhayi ukuphindaphindwa kombhalo wokuqala. Iphrojekthi ibeke amasango ayisithupha aqinile amalungelo obunikazi; ukwehluleka kwanoma yiliphi kuwo kuzovimba ukufakwa noma ukushicilelwa:

1. Ingcaphuno eyodwa akufanele idlule amagama angu-150, futhi kufanele ilandeleke izwi nezwi iye embhalweni wokuqala ofakiwe.
2. Isilinganiso esiphelele sezingcaphuno akufanele sidlule ku-0.03.
3. Okuqukethwe okukopishwe ngokulandelana akufanele kudlule amagama angu-150.
4. Ukudlulelana kwama-shingle angu-13 akufanele kudlule ku-0.10.
5. Umkhiqizo akufanele uqukathe umbhalo ogcwele noma izingcezu zokuqala.
6. contains_original_text kufanele ibe false.

Ama-Skill akhiqizwe ngezincwadi zezinkampani zangaphandle ayimfihlo ngokuzenzakalelayo futhi awavunyelwe ukusatshalaliswa. Uma unamalungelo okusabalalisa esidlangalaleni ngempela, kusadingeka unikeze ngokusobala amapharamitha amabili okuqinisekisa:

--public --confirm-public PUBLISH

Ipharamitha yomphakathi iyisango lokuqinisekisa lobunjiniyela, ayilingani nokugunyazwa kwamalungelo obunikazi. Umsebenzisi kusadingeka aziqinisekisele yena ububanzi belayisensi, imithetho esebenzayo kanye nesimo samalungelo ezinto ezifakwayo.

Amasu athuthukile: ukuqhubeka usebenzisa ukungaphindaphindeki kwezinga lesigaba

Umugqa wokusebenza we-BookLeme osuka ku-S0 uye ku-S9 uklanywe ngokungaphindaphindeki kwezinga lesigaba, futhi usebenzisa i-SHA-256 ukubona imiphumela esivele ikhona. Uma ukucutshungulwa kombhalo omude kuphazamiseka kungazelelwe, ungagcina imiphumela yezigaba eseziqediwe; uma ususa kuphela umphumela wesigaba esithile, ithuluzi lizophinda liqhube leso sigaba ngaphandle kokucubungula kabusha yonke incwadi.

Lo mklamo ulungele futhi ukulungisa ama-plugin noma indlela yokuhumusha. Njengoba zombili izinjini zabelana ngemiphumela yangaphambili yokukhipha, yokuhlukanisa, yokukhipha amakhono kanye neyokuqinisekisa ukwethembeka, ukushintsha ungqimba lokuhumusha akudingi ukufunda kabusha umbhalo wencwadi.

Amasu athuthukile: ukuthuthukisa i-plugin yohlobo yangokwezifiso

Ama-plugin angokwezifiso kufanele abekwe ku-bookleme/plugins/user/<genre_id>.py. Iphrojekthi idinga ukuthi i-plugin imemezele izinkambu ezibalulekile; uma kushoda izinkambu ezidingekayo, ukulayishwa kuzonqatshwa.

• state_machine: Ichaza izimo zenqubo nokuguquka kwazo.
• output_contract: Ichaza okuphumayo okufanele umkhiqizo ukuhlinzeke.
• hard_constraints: Ichaza imithetho engaphulwa.
• anti_patterns: Ibala ukuziphatha okuyiphutha okufanele kugwenywe.
• voice_calibration: Ichaza ithoni, indlela yokubiza abantu, amagama angavunyelwe kanye nokuthi iziphetho zabelwa bani.
• copyright_profile: Imemezela izilungiselelo ezihlobene nokulawula amalungelo obunikazi.
• trigger_words: Inikeza amagama okuqalisa esiShayina angu-3 kuya kwangu-5.
• eval_prompts: Inikeza imiyalelo yokuhlola, okungenani kube nesibonelo esisodwa esingesihle.

Ngaphambi kokubhala i-plugin, kunconywa ukufundaimininingwane ye-plugin; ngemva kokuqeda, sebenzisa plugin validate ukuze uqinisekise. Ungasebenzisi i-plugin ukweqa ukuhlolwa kokwethembeka, isango eliqinile lamalungelo obunikazi noma imfuneko yokwenqaba izibonelo ezingezinhle.

Indlela umugqa wokusebenza we-S0–S9 osebenza ngayo

1. S0–S2: Bhalisa okokufakwayo, ukhiphe ngendlela enqunyiwe, wenze ukujwayelanisa bese uhlukanisa umbhalo wesiShayina ube izigaba, ngaphandle kokuncika kumamaki angaguquki e-Chapter esiNgisini.
2. S3: Qondisa uhlobo lombhalo ngokusekelwe ezicini ezingu-12; lapho kunokungaqiniseki, ungakhetha ngokuzenzakalelayo ukusebenzisa izinjini zombili.
3. S4: Khipha amakhono ayisisekelo, usebenzisa isikhiphi esinqunyiwe kanye nesikhiphi se-LLM esihambisana ne-OpenAI uma kudingeka; uma ukucushwa kungekho, uhlelo lubuyela ngokuzenzakalelayo kwenye indlela.
4. S5: Qinisekisa izingcaphuno igama negama, gcina i-claim ledger, bese ubala izinkomba zokwethembeka.
5. S6: Nikeza iphakheji yolwazi noma i-Behaviour Skill ngokusebenzisa ama-plugin; kulapho izinjini zombili zihlukaniswa khona ngempela.
6. S7–S9: Yenza ukuqinisekiswa okuphelele, ukufakwa okungaphindi imiphumela ehlukile, kanye nokuhlolwa okuyisisekelo.

I-S3 ne-S4 zombili zisebenzisa imiphumela ye-S2 njengokufakwayo, futhi zingasebenza ngokufana. Izinjini zombili zabelana ngokumelwa okuphakathi kwe-S1, S2, S4 ne-S5, ukuze kugwenywe ukufunda nokukhipha ulwazi kabusha.

Izincomo Zokusebenza

• Qala ngemibhalo yezincwadi eyakhiwe ehambisana nephrojekthi ukuze ujwayele imiyalo, bese uqhubekela ezincwadini zangempela.
• Uma ungaqiniseki ngohlobo lomkhiqizo, sebenzisa --engine auto; uma udinga ukuqhathanisa imiphumela emibili ngesikhathi esisodwa, sebenzisa both noma --no-router.
• Ngemva kokushintsha i-plugin noma ithempulethi, sebenzisa njalo ukuqinisekiswa komkhiqizo kanye nezivivinyo zamayunithi.
• Gcina ikhono lokwenqaba imininingwane engatholakali encwadini; ungasebenzisi ulwazi olujwayelekile lwemodeli ukuvala izikhala emthonjeni.
• Ungafaki umbhalo wencwadi wokuqala, ikhava noma izithombe eziskeniwe endaweni yokugcina ikhodi, futhi ungashicileli ngokuqondile imikhiqizo yezincwadi zezinkampani zangaphandle ezizimele ngokuzenzakalelayo.
• Funda idokhumenti yokwethembeka kanye idokhumenti yamalungelo obunikazi ukuze uqonde izinkomba nemingcele yokushicilela.

Imingcele Yamanje

Njengamanje i-BookLeme isesigabeni se-M0–M3. I-coach_dialogue ne-knowledge_reference seziqalisiwe, kodwa ama-plugin okuhlaziya okukhethekile nawohlu lwezinqumo asesehlelweni lomgwaqo oluzayo. I-LLM judge ayikaxhunywa, ngakho ezinye izinkomba ezihlobene nazo zisengu-null. Uma isetshenziswa, kufanele kuhlukaniswe phakathi kwamakhono asevele enziwe, ahlolwe ngokoqobo, nalawo asadinga ukuqinisekiswa.

Isiphetho

I-BookLeme iguqula “ukufunda incwadi” kube uchungechunge lobunjiniyela olungahlolwa: kuqala kukhishwe futhi kuhlukaniswe okuqukethwe, bese kukhishwa amakhono, kuqinisekiswe ukwethembeka, ekugcineni kunikezwe njengephakheji yolwazi noma i-Behaviour Skill. Ngemva kokuqedela ukuhlanganiswa kokuqala, kunconywa ukuthi ukuqinisekiswa, izivivinyo, izimemezelo zamalungelo kanye nokuhlolwa kwamalungelo obunikazi kufakwe enqubweni evamile, ukuze kusetshenziswe kabusha izindlela ezisencwadini ngesikhathi kugcinwa imingcele yomthombo nokuba nokwenzeka kokuqinisekisa.