I-Cisco AI Yethula I-FAPO: Ukuthuthukiswa Okusheshayo Kwepayipi Ngesigaba Sokuhluleka Kwezinga Lesinyathelo Ne-Claude Code Orchestration

0 0 5 minutes read

I-Cisco AI Yethula I-FAPO: Ukuthuthukiswa Okusheshayo Kwepayipi Ngesigaba Sokuhluleka Kwezinga Lesinyathelo Ne-Claude Code Orchestration

Ukuthola ukwaziswa okufanele kuseyingxenye enzima kakhulu yokuthumela izinhlelo zokusebenza ezithembekile ze-LLM. Izinguquko ezincane zamagama zingashintsha ukunemba ngamaphesenti angama-20. Okusebenza ezibonelweni ezimbalwa kuvame ukwephula esikalini. Uma ipayipi lezinyathelo eziningi libuyisela impendulo engalungile, ukuthola isinyathelo esihlulekayo kusho ukuhlola okuphumayo okuphakathi ngesandla.

I-Cisco AI yethulwa I-FAPO ukubhekana nalelo phuzu. I-FAPO imele i-Fully Automated Prompt Optimization. Kuwuhlelo oluqhutshwa yi-Claude Code oluthuthukisa amapayipi e-LLM kusukela ekwazisweni okuyisisekelo ukuya ekunembayo okuqondiwe. Uhlinzeka ngedathasethi kanye nokwaziswa kokuqala. I-FAPO ibe isihlola, ihlukanise ukwehluleka, iphakamise okuhlukile, iqinisekise, futhi iphindaphinde. Yonke iluphu ihlelwe ngabasebenzeli be-Claude Code. Iphrojekthi ithumela umthombo ovulekile ngaphansi kwe-Apache 2.0, futhi isekela i-Codex njenge-ejenti yokwenza kahle.

Ekuhlolweni okubikiwe kwe-Cisco, i-FAPO yehlula i-GEPA, isikhuthazi sesimanjemanje, ekuqhathaniseni imodeli-benchmark engu-15 kwezingu-18. Kumabhentshimakhi amabili lapho i-FAPO inyukele khona ekushintsheni kwamapayipi, inzuzo emaphakathi ngaphezu kwe-GEPA ifinyelele ku-+33.8pp.

I-TL;DR

I-FAPO iwuhlelo oluqhutshwa yi-Claude Code oluthuthukisa ngokuzenzakalelayo amaphayiphi e-LLM anezinyathelo eziningi ukusuka ekwazisweni okuyisisekelo kuya ekunembayo okuqondiwe, umthombo ovulekile ngaphansi kwe-Apache 2.0.
Inyuka ngamaleveli amathathu – ngokushesha, ipharamitha, bese kwakheka iketango – isebenzisa isichasiso sokwehluleka kwezinga lesinyathelo ukunquma ukuthi yini okufanele ushintshe ngokulandelayo.
Ekuhlaziyeni kwe-Cisco, i-FAPO yehlule i-GEPA ngokuqhathanisa okungu-15 kokungu-18 kwemodeli-benchmark, ngenzuzo yesilinganiso engu-+14.1pp.
Ku-HoVer ne-IFBench, lapho ikhuphukele ekushintsheni kwamapayipi, i-FAPO iwine wonke amapheya ayisithupha ngenzuzo esho ukuthi +33.8pp; I-AIME kube ukuphela kokuwina kwe-GEPA, ngaphakathi komsindo wamasampuli.
Izivivinyo ezimelene nokufakwa ngokweqile zibandakanya ukuhlolwa kokuhlukaniswa kokuqeqeshwa kuphela, amafayela ahlukile angaguquleki, kanye nombuyekezi ozimele kuso sonke isiphakamiso.

Yini i-FAPO

I-FAPO wuhlaka lokuhlola nokuthuthukisa abaqashi abaningi. Isiqashi iphrojekthi yokuthuthukisa eziqukethwe yona. Uhla lwemibhalo ngalunye lomqashi luphethe imiyalo yomsebenzi owodwa, isethi yedatha, incazelo yeketanga, umshayimagoli, kanye ne-config. Abaqashi bahlala bodwa, ngakho-ke imisebenzi engahlobene yenza kube ngcono ngapha nangapha ngaphandle kokuphazamiseka.

Injini ewumgogodla ibizwa nge-hephaestus futhi isizinda-agnostic. Iphatha ukuhlola, ukukhishwa kweketango, nokushaya amaphuzu. Amaketanga angamagrafu esifunda se-LangGraph acubungula ikesi ngalinye lokuhlola. Ngaphandle kwebhokisi, I-FAPO isekela abahlinzeki abathathu: I-OpenAI, i-Baseten, ne-SageMaker.

Okokufaka okukodwa okufanele ukulethe isethi yedatha. Kungukufakwayo okubhanqiwe kanye nemiphumela elindelekile echaza impumelelo. I-FAPO iyihlukanisa ibe yisethi yokuqinisekisa kanye nesethi yokuhlola ebanjiwe. Isethi yokuqinisekisa ishayela ukuphindaphinda; isethi yokuhlola isetshenziswa kuphela ekuhloleni kwesibhamu esisodwa sokugcina. Kusukela encazelweni yomsebenzi, u-Claude angakwazi ukuhlanganisa okusele: ukwaziswa kokuqala, iketango, kanye nomshayi amaphuzu.

Isebenza kanjani i-Optimization Loop

Uma izingcezu sezikhona, i-FAPO isebenzisa iluphu evaliwe kuze kufinyelelwe ukunemba kwethagethi. Umjikelezo ngamunye uqhuba izigaba eziyisithupha:

Linganisa — sebenzisa uchungechunge kudathasethi, qoqa amaphuzu ekesi ngalinye kanye nemiphumela yezinga lesinyathelo.
Isibaluli – hlukanisa ukwehluleka ngembangela usebenzisa i-heuristics esekelwe kumthetho kanye nokuhlaziywa kwe-LLM.
Phakamisa – khiqiza okuhlukile okuqondise iqoqo lokuhluleka elivelele.
Buyekeza – i-ejenti ezimele iqinisekisa isiphakamiso sokuthotshelwa kobubanzi kanye nokuvuza kwedatha.
Qhathanisa — yamukela okuhlukile kuphela uma kuba ngcono kokudlule, uma kungenjalo yenqaba.
Phinda futhi – qhubeka kuze kufinyelelwe ukunemba kwethagethi noma isabelomali sokuthuthukisa sesiphelile.

Uhlelo lusebenza kumazinga amathathu akhuphukayo. Ukuhlela ngokushesha kuyizindleko eziphansi futhi kuzanywe kuqala. Ipharamitha iyashintsha lungisa amanani okumisa njenge retrieval_k noma temperature. Izinguquko zesakhiwo zishintsha i-topology yeketango, njengokwengeza i-node yokuziphendulela noma ukushintshela kuphethini ye-ReAct. I-FAPO iqeda ileveli eyodwa ngaphambi kokuthi yenyukele kwelandelayo.

Isinyathelo sokulinganisa sihlunga ukuhluleka kube izigaba ezine. Ukuhluleka kokubuyisa kubuyisela okuqukethwe okungenalutho noma okungabalulekile. Ukuhluleka kwe-Cascading kuqala lapho isinyathelo sokuqala sikhiqiza okukhiphayo okungenalutho. Ukwehluleka kwefomethi kufihla impendulo efanele ngaphakathi kombhalo umenzi wegoli akakwazi ukuncozulula. Ukwehluleka kokucabanga kwenzeka lapho imibono emihle isaveza isiphetho esingalungile. Izinkinga zefomethi nokucabanga ziyalungiseka ngokushesha. Izinkinga zokubuyisa kanye ne-cascade ziyalungiseka.

Ama-Guardrails agcina i-optimizer ingachithi kakhulu. Ihlola kuphela izimo ezihlukene zokuqeqesha, kuyilapho ukuqinisekiswa nokuhlola kuveza amaphuzu ahlanganisiwe kuphela. Konke okuhlukile kuyifayela elisha elingaguquleki, elingakaze lihlelwe endaweni yalo. Umbuyekezi ozimele uhlola isiphakamiso ngasinye ngaphambi kokuthi siqalise.

I-Benchmark Case: I-FAPO iqhudelana ne-GEPA

Ithimba le-Cisco lihlole i-FAPO iqhathanisa ne-GEPA (Generalized Evolutionary Prompt Architecture), indlela yesimanjemanje yokuthuthukisa. I-GEPA isebenzisa ukusesha kokuziphendukela kwemvelo ngama-opharetha ofuzo ukuze ithuthukise ukwaziswa kwamapayipi anezinyathelo eziningi. Womabili amasistimu aqale ngamapayipi ayisisekelo afanayo kanye nemiyalelo. I-FAPO ingase inyukele ezinguqukweni zesakhiwo lapho ukuchasisa kutholwa izingqinamba. I-GEPA ibikhawulelwe ekuthuthukisweni kweleveli esheshayo.

Ukuqhathanisa kuthathe amabhentshimakhi ayisithupha namamodeli emisebenzi amathathu: GPT-4.1-mini, GPT-5.4-mini, ne-Gemma 3-12B. U-Claude Opus 4.6 usebenze njenge-orchestrator ye-FAPO kanye nesibonisi se-GEPA. Izikolo ezingezansi zilinganiselwa kuwo wonke amamodeli emisebenzi amathathu.

Ibhentshimakhi	Isisekelo	I-GEPA	I-FAPO	Inzuzo vs. GEPA
HoVer	35.9	48.5	83.8	+35.3pp
IFBench	35.7	48.5	80.7	+32.2pp
I-LiveBench-Math	51.0	52.6	62.0	+9.4pp
I-HotpotQA	50.9	61.8	68.3	+6.5pp
I-Papillon	73.6	90.7	94.9	+4.2pp
AIME	16.7	16.0	12.9	-3.1pp

I-FAPO iwine okungu-15 kokungu-18 kokuqhathanisa okuyimodeli, ngenzuzo emaphakathi engu-+14.1pp kune-GEPA. Ku-HoVer ne-IFBench, lapho i-FAPO inyukele ekushintsheni kwamapayipi, iwine wonke amapheya ebhentshimakhi ayisithupha. Isilinganiso senzuzo lapho sasingu-+33.8pp. Kumabhentshimakhi amane ngaphandle kwezinguquko zesakhiwo, i-FAPO isawine okungu-9 kwayi-12 ngokulungiselela ngokushesha kuphela. I-AIME ukuphela kwebhentshimakhi lapho i-GEPA ihole khona, ngo-3.1pp. Igebe lincane kunokuchezuka okujwayelekile kuzo zonke izivivinyo zestochastic.

Ukuqhathanisa amandla kubonisa umehluko wedizayini obikwe yiCisco. Yonke imigqa engezansi ibonisa incazelo yomthombo wamasistimu amabili.

Amandla	I-GEPA	I-FAPO
Amazinga okuthuthukisa	Umbhalo osheshayo kuphela	Yazisa → ipharamitha → isakhiwo
Ingashintsha ukwakheka kweketango	Cha	Yebo, lapho ukubonakaliswa kuthola izingqinamba
Iqhutshwa kanjani	Ukusesha kwemvelo ngama-genetic operators	Claude Code noma i-Codex agent loop
Umphumela kuwo wonke amapheya ebhentshimakhi angu-18	Ireferensi	Uwina 15 kokungu-18; +14.1pp kusho

Lapho Ilingana: Sebenzisa Amacala

I-FAPO iqondise amapayipi e-LLM anezinyathelo eziningi, hhayi ukwaziswa okukodwa. Izibonelo ezimbalwa ezibambekayo:

Ukuphendula imibuzo ye-Multi-hop: Uchungechunge lubuyisa amadokhumenti, lukhiphe amaqiniso, izizathu phezu kobufakazi, futhi lufomethe impendulo. Ku-walkthrough ye-Cisco ebhaliwe, uchungechunge lwe-QA lwe-multi-hop lukhuphuke lusuka ku-39.3% lwaya ku-70.3% lokuqinisekisa ukufana ncamashi ekuphindaphindweni okubili. Isichasiso sibe sesihlaba umkhosi ukwehluleka okusele njengokukhawulelwe kokubuyisa, okubonisa ukulungiswa kwesakhiwo. Ngokwehlukana, kubhentshimakhi ye-HotpotQA, i-FAPO ifinyelele ukunemba kokuhlolwa okungu-68.3% uma iqhathaniswa nama-61.8% we-GEPA.
Umyalelo olandelayo: Ku-IFBench, ukuhluleka kwefomethi yokukhawulela kuphushe i-FAPO ukuthi inyukele ngale kokwaziswa, yafinyelela ukunemba kokuhlolwa okungu-80.7%.
Ukwahlukanisa: Umsebenzi we-software-igama-to-isigaba ungacushwa nge-Claude Code, bese uthuthukiswa ukuze ufane ncamashi okuhlosiwe.
Ama-ejenti we-React: Isandiso sokugeleza komsebenzi kwe-MCP silungiselela ithuluzi lokubiza i-ejenti ye-ReAct kusetshenziswa i-trajectory scoring kanye ne-LLM-as-Judge scoring.

Ukuqalisa

Indlela eshesha kakhulu ukuvumela i-Claude Code ukuthi idale amafayela abaqashi. Kusukela ku-repo, chaza umsebenzi wakho ngesiNgisi esilula, bese wengeza idathasethi ye-JSONL. Umugqa ngamunye unekesi elilodwa lokuhlola case_id, task_type, context, expectedfuthi metadata:

{"case_id": "1", "task_type": "qa", "context": {"question": "What is the capital of France?"}, "expected": {"answer": "Paris"}, "metadata": {}}
{"case_id": "2", "task_type": "qa", "context": {"question": "What is 2 + 2?"}, "expected": {"answer": "4"}, "metadata": {}}

Oshaya amagoli uqhathanisa umphumela weketango nempendulo elindelekile. Iyasebenza validate_case ukubamba idatha embi kusenesikhathi futhi score_case ukubuyisela isikolo esiyinhlanganisela:

from hephaestus.scoring.scorer import Scorer as BaseScorer

class Scorer(BaseScorer):
    def validate_case(self, case, scoring_profile):
        assert "answer" in case.expected, "Missing 'answer' in expected"

    def score_case(self, case, output_text, scoring_profile):
        expected = case.expected["answer"].strip().lower()
        predicted = output_text.strip().lower()
        em = 100.0 if predicted == expected else 0.0
        return {"composite_score": em, "score_breakdown": {"exact_match": em}}

Qinisekisa ukusetha ngokuhlola okuyisisekelo:

export OPENAI_API_KEY="sk-..."
python -m hephaestus.cli eval --config tenants/my_project/configs/eval.json

Bese ucela i-ejenti yokuthuthukisa ngesiqashi, ukulungiselelwa, kanye nemibandela yokuphumelela njenge composite_score >= 90. I-Claude Code ikhiqiza inkontileka ye-scope, bese iphindaphinda ngokuzenzakalelayo. Konke okuhlukile kokwaziswa, ukulungiselelwa, nokuhlaziya okuhlukile kubhalwe kudiski, ngakho ukugijima ngakunye kuhlala kufundeka. I-UI yasendaweni yokufunda kuphela ebizwa nge-FAPO Explorer iphequlula ama-artifact kamuva.

Amandla Nobuthakathaka

Amandla

I-Pipeline-aware yafaka ukwehluleka kokuthola amaphuzu esinyathelweni esibadalile, hhayi nje umphumela wokugcina.
Ukukhuphuka kwamaleveli amathathu kusingatha ukwehluleka okungalungiswa kukodwa nje.
Izivikelo ezimelene nokufakwa ngokweqile: ukuhlolwa kokuhlukaniswa kokuqeqeshwa kuphela, okuhlukile okungaguquleki, kanye nombuyekezi ozimele.
Umthombo ovulekile ngaphansi kwe-Apache 2.0, kukho kokubili i-Claude Code ne-Codex esekelwe.

Ubuthakathaka

Ikhwalithi yokuthuthukisa ibophezelwe ikhwalithi yedathasethi kanye nokufakwa kwayo, okufanele ukunikeze.
Iphrojekthi yakamuva, ngakho-ke amarekhodi okukhiqiza azimele asalinganiselwe.
Iluphu ezenzakalelayo incike kumathuluzi okubhala amakhodi e-ejenti (Ikhodi ye-Claude noma i-Codex) kunesilungiseleli esizimele.

Isichazi Esisebenzisanayo

“+d.name+': I-FAPO iqede kokuthi '+d.fapo.toFixed(1)+', '+Math.abs(beat).toFixed(1)+'pp ngaphansi kwe-GEPA. Lona ukuphela kokulinganisa lapho i-GEPA ihola khona, futhi igebe lihlala phakathi komsindo wamasampuli.'; } okunye {txt=”<b>“+d.name+':</b> I-FAPO ifinyelele ku-'+d.fapo.toFixed(1)+', '+beat.toFixed(1)+'pp phezu kwe-GEPA'+ (d.struct?' ngemva kokukhuphukela ezinguqukweni zamapayipi esakhiwo.':' ngokuhlelwa ngokushesha nangepharamitha kuphela.'); } v.innerHTML=txt; v.classList.add(“show”); } okunye {v.classList.remove(“show”); } } umsebenzi wokupenda(osebenzayo){ renderLoop(active===undefined?-1:active); amamitha(); attr(); logRender(); Ibhokisi lesigwebo(); } ukusetha kabusha umsebenzi(){ cycle=0; acc=0; ukugijima=amanga; plan=buildPlan(DATA[cur]); document.getElementById(“stepBtn”).disabled=false; document.getElementById(“runBtn”).disabled=false; ama-chips (); upende(-1); kufanelekile (); } umsebenzi applyCycle(cb){ var s=plan[cycle]; // phila ezigabeni eziyisi-6 ngokushesha var i=0; var iv=setInterval(umsebenzi(){renderLoop(i); uma(i===4){// Qhathanisa -> buyekeza ukunemba uma kwamukelwe uma(s.accepted) acc=s.acc; metres(); attr(); } i++; if(i>=6){ clearInterval(iv); cycle(x); buyisela emuva(x); buyisela emuva(x);(render) fit(); uma(umjikelezo>=plan.length){document.getElementById(“stepBtn”).disabled=true(“runBtn”).disabled=true; } document.getElementById(“stepBtn”).onclick=function(){ if(running||cycle>=plan.length) return; ukugijima=iqiniso; setBtns(iqiniso); applyCycle(function(){ running=false; setBtns(false); }); }; document.getElementById(“runBtn”).onclick=function(){ if(running||cycle>=plan.length) return; ukugijima=iqiniso; setBtns(iqiniso); (umsebenzi olandelayo(){ if(cycle>=plan.length){ running=false; setBtns(false); buyisela; } applyCycle(function(){ setTimeout(next,120); }); })(); }; document.getElementById(“resetBtn”).onclick=function(){ if(running) return; setha kabusha(); }; umsebenzi setBtns(dis){ document.getElementById(“runBtn”).disabled=dis||cycle>=plan.length; document.getElementById(“stepBtn”).disabled=dis||cycle>=plan.length; } // —–shintsha usayizi ngokuzenzakalela wokushumeka kwe-WordPress —- function fit(){ try{ var el=document.getElementById(“fapo-wrap”); var h=el.offsetHeight+40; uma(window.parent){ window.parent.postMessage({fapoHeight:h},”*”); } }bamba(e){} } window.addEventListener(“layisha”,faka); window.addEventListener(“resize”,fit); setha kabusha(); })();</p> <p>

Source link

nimda 3 hours ago

0 0 5 minutes read