Kacaanka Ugaadhsiga Bahalnimada: AI-la xoojiyey D&D kulanada

Ciidan ka tirsan AI-ku shaqeeya ee mugdiga

Waa habeen cayaareed, saaxiibadaa waxay taagan yihiin hareeraha miiska ciyaaraha, iyagoo sugaya inay arkaan Dungeons & Dragons (D&D) dabeecadda ay noqon doonaan oo ay raadin doonaan inay bilaabaan. Caawa, waxaad tahay Dungeon Master (sheeko-sheeko iyo hage), farsamayaqaanka kulammada xiisaha leh si aad ula tartanto oo aad u xiisayso ciyaartoygaaga. Buug-gacmeedkaaga D&D Monster ee aaminka ah waxa ka kooban kumanaan makhluuqaad ah. Helitaanka bahal ku habboon xaalad kasta oo ka mid ah xulashooyinka tirada badan waxay noqon kartaa mid culus. Cadawga ku habbooni wuxuu u baahan yahay inuu la jaanqaado goobta, dhibka, iyo sheekada xilligan.

Maxaa dhacaya haddii aan samayn karno qalab isla markiiba helaya bahalnimada ugu habboon xaalad kasta? Qalab tixgalinaya arrimo badan isku mar , hubinta in kulan kastaa uu yahay mid xiiso badan oo suurtogal ah?

Aan guda galno raadinta anaga: dhisno habka ugu dambeeya ee raadinta bahalnimada, anagoo adeegsanayna awooda raadinta sifo badan!

Abuuritaanka makhluuq leh raadinta vector, maxay tahay sababta?

Raadinta Vector waxay u taagan tahay kacaanka soo celinta macluumaadka. Ku-xidhka Vector-adiga oo tixgelinaya macnaha guud iyo macnaha semantic- waxay awood u siinaysaa raadinta vector si ay u soo celiso natiijooyin sax ah oo dheeraad ah, ma aha oo kaliya xogta habaysan laakiin sidoo kale xogta aan habaysanayn iyo luqado badan, iyo miisaan. Laakiin si loo dhaliyo jawaabaha tayada sare leh ee codsiyada adduunka dhabta ah, waxaan inta badan u baahanahay inaan ku meeleyno miisaanyo kala duwan sifooyin gaar ah oo walxaha xogtayada ah.

Waxaa jira laba hab oo caadi ah oo loo adeegsado raadinta sifo-badan. Labaduba waxay ku bilaabaan iyagoo si gaar ah u guntanaya sifo kasta oo shay xog ah. Farqiga ugu weyn ee u dhexeeya labadan hab ayaa ah sida loo kaydiyo loona baadho waxyaalaha la isku dhejiyo .

1. habka naive - ku kaydi sifo kasta oo ka mid ah dukaamada vector gaar ah (mid ka mid ah sifo), samee raadinta u gaar ah sifa kasta, isku dar natiijooyinka raadinta, iyo post-process (tusaale, miisaanka) sida loo baahan yahay.
2. Habka Superlinked - isku xidh oo ku kaydi dhammaan vectors sifooyin isku mid ah dukaanka vector (adoo isticmaalaya Superlinked's functionality), kaas oo noo ogolaanaya inaan hal mar kaliya baadho , iyadoo la helayo faa'iidada waxtarka shaqaalaha. spaces Superlinked sidoo kale waxay noo ogolaanayaan inaan miisaanno sifo kasta wakhtiga su'aasha si aan u soo saarno natiijooyin badan oo khuseeya, iyada oo aan lahayn habraac ka dib.

Hoosta, waxaan u isticmaali doonaa labadan hab si aan u hirgelino qalab raadinta vector-ka sifooyin badan leh - Dungeons and Dragons monster finder! Hirgelintayada fudud, gaar ahaan kan labaad, ayaa muujin doona sida loo abuuro habab goobid oo xoog badan oo dabacsan, kuwaas oo si fudud u maarayn kara su'aalo adag oo dhinacyo badan leh, xaalad kasta oo aad isticmaasho.

Haddii aad ku cusub tahay raadinta isku midka ah vector, ha welwelin! Waanu kuu daboolnay - eeg maqaalladayada dhismayaasha .

Hagaag, aan tagno ugaarsiga bahalnimada!

Xogta

Marka hore, waxaanu soo saari doonaa xog-ururin yar oo ah daanyeerka, anagoo kicinayna Qaabka Luuqadda Weyn (LLM):

 Generate two JSON lists: 'monsters' and 'queries'. 1. 'monsters' list: Create 20 unique monsters with the following properties: - name: A distinctive name - look: Brief description of appearance (2-3 sentences) - habitat: Where the monster lives (2-3 sentences) - behavior: How the monster acts (2-3 sentences) Ensure some monsters share similar features while remaining distinct. 2. 'queries' list: Create 5 queries to search for monsters: - Each query should be in the format: {look: "...", habitat: "...", behavior: "..."} - Use simple, brief descriptions (1-3 words per field) - Make queries somewhat general to match multiple monsters Output format: { "monsters": [ {"name": "...", "look": "...", "habitat": "...", "behavior": "..."}, ... ], "queries": [ {"look": "...", "habitat": "...", "behavior": "..."}, ... ] }

Aynu eegno muunad ka mid ah xog-ururinta LLM ee uu soo saaray. Fiiro gaar ah: Jiilka LLM ma aha mid go'aamin kara, markaa natiijooyinkaagu way kala duwanaan karaan.

Waa kuwan shantayada xayawaan ee ugu horreeya:

... iyo su'aalahayaga soo baxay:

Fiiri xogta asalka ah iyo tusaalooyinka weydiimaha halkan .

Soo celinta

Aynu dejinno cabbirrada aan u adeegsan doono labadayada hab - naive iyo Superlinked - hoos.

Waxa aanu ku soo saarnay qalabyada vector-ka ah:

 sentence-transformers/all-mpnet-base-v2.

Fududeynta awgeed, waxaan ku xaddidi doonaa wax soo saarkayaga 3da kulan ee ugu sarreeya. (Si aad u heshid kood dhammaystiran, oo ay ku jiraan soo dejinta lagama maarmaanka ah iyo hawlaha caawiyaha, eeg buug-yaraha .)

 LIMIT = 3 MODEL_NAME = "sentence-transformers/all-mpnet-base-v2"

Hadda, aynu helno raadintayada daanyeerka sifada badan leh! Marka hore, waxaan isku dayi doonaa habka nafiska ah .

Habka nacasnimada ah

Habkayaga caqli-galnimada ah, waxaanu si madax-bannaan u dhejinnay sifooyin waxaanu ku kaydinnaa tilmaamo kala duwan. Wakhtiga waydiinta, waxaanu wadnaa baadhitaano badan oo kNN ah dhammaan tusmooyinka, ka dibna dhammaan natiijooyinkayada qayb ahaan aan isku darno mid.

Waxaan ku bilaabaynaa qeexida fasalka

 NaiveRetriever

si loo sameeyo raadinta isku midka ah ee ku saleysan xog-ururinta, iyadoo la adeegsanayo all-mpnet-base-v2 -ku-xidhka-soo-saarka.

 class NaiveRetriever: def __init__(self, data: pd.DataFrame): self.model = SentenceTransformer(MODEL_NAME) self.data = data.copy() self.ids = self.data.index.to_list() self.knns = {} for key in self.data: embeddings = self.model.encode(self.data[key].values) knn = NearestNeighbors(metric="cosine").fit(embeddings) self.knns[key] = knn def search_key(self, key: str, value: str, limit: int = LIMIT) -> pd.DataFrame: embedding = self.model.encode(value) knn = self.knns[key] distances, indices = knn.kneighbors( [embedding], n_neighbors=limit, return_distance=True ) ids = [self.ids[i] for i in indices[0]] similarities = (1 - distances).flatten() # by definition: # cosine distance = 1 - cosine similarity result = pd.DataFrame( {"id": ids, f"score_{key}": similarities, key: self.data[key][ids]} ) result.set_index("id", inplace=True) return result def search(self, query: dict, limit: int = LIMIT) -> pd.DataFrame: results = [] for key, value in query.items(): if key not in self.knns: continue result_key = self.search_key(key, value, limit=limit) result_key.drop(columns=[key], inplace=True) results.append(result_key) merged_results = pd.concat(results, axis=1) merged_results["score"] = merged_results.mean(axis=1, skipna=False) merged_results.sort_values("score", ascending=False, inplace=True) return merged_results naive_retriever = NaiveRetriever(df.set_index("name"))

Aan isticmaalno weydiinta koowaad ee liiskayaga sare ee la soo saaray, oo aan raadinno wuxuush anagoo adeegsanayna naive_retriever :

 query = { 'look': 'glowing', 'habitat': 'dark places', 'behavior': 'light manipulation' } naive_retriever.search(query)

Our

 naive_retriever

waxay soo celisaa natiijooyinka soo socda ee raadinta sifo kasta:

Cajiib! Natiijooyinka bahalnimadayada ee la soo celiyay ayaa khuseeya - kulligood waxay leeyihiin sifo "iftiin leh".

Aan aragno waxa hab-dhaqanka caqli-xumada ahi soo noqdo marka aynu baadhno labada sifo ee kale.

Dhammaan daanyeerka la soo saaray waxay leeyihiin sifooyinka la doonayo. Jaleecada hore, natiijooyinka raadinta caajiska ah waxay u muuqan karaan rajo. Laakiin waxaan u baahannahay inaan helno daanyeerka leh dhammaan saddexda sifo isku mar . Aan isku darno natiijooyinkeena si aan u aragno sida wanaagsan ee uushayaasheenu u sameeyaan gaaritaanka yoolkan:

Oo halkan, xadka habka caqli-galnimada ayaa noqda mid cad. Aynu qiimayno:

1. Ku habboon sifada:
   • look : Saddex bahal ayaa la soo saaray (Shade Shade, Sandstorm Djinn, iyo Luminoth).
   • habitat : Kaliya hal bahal oo ka mid ah natiijooyinka look ayaa khuseeyey (Hadhka qaylo-dhaanta).
   • behavior : Kaliya hal bahal oo ka mid ah natiijooyinka look ayaa khuseeya (Luminoth), laakiin way ka duwan tahay kan ku habboon habitat .
2. khusaynta guud:
   • Hal bahal midna lama soo saarin dhammaan saddexda sifo isku mar.
   • Natiijooyinku waa kala qaybsan yihiin: xayawaanno kala duwan ayaa ku habboon sifooyin kala duwan.

Marka la soo koobo, habka raadinta naive wuxuu ku guuldareystaa inuu helo xayawaanno qanciya dhammaan shuruudaha hal mar. Waxaa laga yaabaa in aan xallin karno arrintan annagoo si firfircoon u soo saarnay wuxuush badan sifo kasta? Aynu ku tijaabino 6 bahal sifo kasta, halkii aad ka ahaan lahayd 3. Aan eegno waxa habkani dhalinayo:

Waxaan hadda soo saarnay 13 bahal (in ka badan kala badh xogtayada yar ee xogta ah!), Oo haddana waxaan leenahay isla arrin isku mid ah: mid ka mid ah dabaylahan lama soo saarin dhammaan saddexda sifo.

Kordhinta tirada daanyeerka la soo saaray (ka badan 6) ayaa laga yaabaa inay xalliso mushkiladayada, laakiin waxay abuurtaa arrimo dheeraad ah:

1. Soo saarista, soo celinta natiijooyin badan (baaritaanno badan oo kNN ah) waxay kordhisaa wakhtiga raadinta si muuqata.
2. Sifo kasta oo cusub oo aan soo bandhigno, fursadahayaga ah inaan helno bahal "dhameystiran" - oo leh dhammaan sifooyinka ku jira weydiinteena - ayaa hoos u dhacaysa si aad ah. Si taas looga hortago, waa inaan soo saarno derisyo badan oo kale oo noo dhow (bahal), taasoo ka dhigaysa tirada guud ee daanyeerrada la soo saaray inay si aad ah u koraan.
3. Wali ma haysano dammaanad qaadka inaan soo ceshan doonno bahalnimada leh dhammaan sifooyinka aan doonayno.
4. Haddii aan awoodno inaan soo saarno daanyeerka qanciya dhammaan shuruudaha hal mar, waa inaan bixinaa natiijooyin dib-u-heshiisiin oo dheeraad ah.

Isku soo wada duuboo, habka caqli-galku waa mid aan la hubin oo aan waxtar u lahayn raadinta sifooyin badan oo macquul ah, gaar ahaan wax soo saarka.

Habka isku xidhka sare

Aan hirgelinno habkayaga labaad si aan u aragno inuu ka fiican yahay kan aan caqli-gal ahayn.

Marka hore, waxaanu qeexaynaa qaabka, meelaha bannaan, tusmada, iyo weydiinta:

 @schema class Monster: id: IdField look: String habitat: String behavior: String monster = Monster() look_space = TextSimilaritySpace(text=monster.look, model=MODEL_NAME) habitat_space = TextSimilaritySpace(text=monster.habitat, model=MODEL_NAME) behavior_space = TextSimilaritySpace(text=monster.behavior, model=MODEL_NAME) monster_index = Index([look_space, habitat_space, behavior_space]) monster_query = ( Query( monster_index, weights={ look_space: Param("look_weight"), habitat_space: Param("habitat_weight"), behavior_space: Param("behavior_weight"), }, ) .find(monster) .similar(look_space.text, Param("look")) .similar(habitat_space.text, Param("habitat")) .similar(behavior_space.text, Param("behavior")) .limit(LIMIT) ) default_weights = { "look_weight": 1.0, "habitat_weight": 1.0, "behavior_weight": 1.0 }

Hadda, waxaan bilownay fuliyaha oo soo rarnaa xogta:

 monster_parser = DataFrameParser(monster, mapping={monster.id: "name"}) source: InMemorySource = InMemorySource(monster, parser=monster_parser) executor = InMemoryExecutor(sources=[source], indices=[monster_index]) app = executor.run() source.put([df])

Aynu wadno isla waydiintii aanu ku wadnay fulinta hab-dhaqankeena nafiska ah ee sare:

 query = { 'look': 'glowing', 'habitat': 'dark places', 'behavior': 'light manipulation' } app.query( monster_query, limit=LIMIT, **query, **default_weights )

Iyo voila! Markan, mid kasta oo ka mid ah daanyeerkayaga ugu sarreeya ee soo laabtay ayaa si heer sare ah ugu jira buundooyin ka dhigan nooc ka mid ah "macnaha" dhammaan saddexda astaamood ee aan rabno inuu xayawaankeenna yeesho. Aynu si faahfaahsan u jebinno dhibcaha bahal kasta:

Natiijooyinkayada labaad iyo saddexaad, Luminoth iyo Living Graffiti, labaduba waxay leeyihiin dhammaan saddexda astaamood ee la rabo. Natiijada ugu sareysa, Shade Shade, in kasta oo ay ku yar tahay marka la eego khalkhalgelinta iftiinka - sida ka muuqata dhibcaha behavior (0.006), waxay leedahay sifooyin "iftiin" iyo jawi mugdi ah oo ka dhigaya look (0.168) iyo habitat (0.203) dhibco aad u badan. sare, taasoo siinaya wadarta guud ee dhibcaha ugu sarreeya (0.377), taasoo ka dhigaysa daanyeerka ugu habboon guud ahaan. Maxaa horumar ah!

Ma ku celin karnaa natiijooyinkeena? Aan isku dayno weydiin kale oo aan ogaano.

 query = { 'look': 'shapeshifting', 'habitat': 'varied landscapes', 'behavior': 'illusion creation' }

Wayn! Natiijooyinkayagu waa kuwo aad u wanaagsan mar kale.

Ka warran haddii aan rabno inaan helno xayawaanno la mid ah xayawaan gaar ah oo ka mid ah xogtayada? Aynu ku tijaabino bahal aynaan weli arag - Harmonic Coral. Waanu soo saari karnaa sifada bahalkan oo aanu gacanta ku samayn karno cabbiraadaha weydiinta. Laakiin Superlinked waxay leedahay hab with_vector oo aan u isticmaali karno shayga weydiinta. Sababtoo ah aqoonsiga bahal kastaa waa magaciisa, waxaan u muujin karnaa codsigeena sida fudud:

 app.query( monster_query.with_vector(monster, "Harmonic Coral"), **default_weights, limit=LIMIT )

Natiijada ugu sareysa waa tan ugu habboon, Harmonic Coral lafteeda, sida la filayo. Labada bahal ee kale ee raadintayadu waa Dreamweaver Octopus iyo Aqua Wraith. Labaduba waxay wadaagaan mawduucyo ( sifo ) muhim ah oo leh Harmonic Coral:

1. Degaanada biyaha ( habitat )
2. Awood u leh inay saameyn ku yeelato ama wax ka qabato deegaankooda ( behavior )
3. Sifooyin muuqaal ah oo firfircoon ama dareere ah ( look )

Miisaanka sifo

Ka soo qaad, hadda, in aan rabno in aan siino muhiimad dheeraad ah sifada look . Qaab dhismeedka Superlinked wuxuu noo ogolaanayaa inaan si fudud u hagaajino miisaanka wakhtiga su'aasha. Si sahlan loo barbardhigo, waxaan raadin doonaa daanyeerka la mid ah Harmonic Coral, laakiin miisaankeena la hagaajiyay si loo eego look .

 weights = { "look_weight": 1.0, "habitat_weight": 0, "behavior_weight": 0 } app.query( monster_query.with_vector(monster, "Harmonic Coral"), limit=LIMIT, **weights )

Natiijooyinkayagu dhammaantood (si habboon) waxay leeyihiin muuqaal isku mid ah - "Laamood oo leh jilicsanaan gariiraya", "Makhluuqa dhirta u eg oo leh jir geed canab ah oo qalloocan iyo qodxan", "Snake-like".

Haddaba, aynu samayno raadin kale, inagoo iska indha tirayna muuqalka, oo aynu halkii ka raadino bahal la mid ah habitat ahaan iyo behavior isku mar:

 weights = { "look_weight": 0, "habitat_weight": 1.0, "behavior_weight": 1.0 }

Mar labaad, habka Superlinked wuxuu soo saaraa natiijooyin fiican. Dhammaan saddexda bahal waxay ku nool yihiin deegaanno biyo leh waxayna leeyihiin awood maskax-xakameyn.

Ugu dambeyntii, aan isku dayno baaritaan kale, oo aan miisaanno dhammaan saddexda sifo si kala duwan - si aan u helno daanyeerka marka la barbardhigo Harmonic Coral waxay u egyihiin kuwo isku mid ah, ku nool degaanno kala duwan, oo leh dabeecad aad isku mid ah:

 weights = { "look_weight": 0.5, "habitat_weight": -1.0, "behavior_weight": 1.0 }

Natiijooyin fiican mar labaad! Labadayada xayawaan ee kale ee dib loo soo celiyay - Luminoth iyo Zephyr Dancer - waxay leeyihiin dabeecad la mid ah Harmonic Coral waxayna ku nool yihiin degaanno ka duwan Harmonic Coral's. Waxay sidoo kale aad uga duwan yihiin Harmonic Coral. (In kasta oo jilbaha Harmonic Coral iyo anteenada Luminoth ay xoogaa isku mid yihiin, haddana waxaan kaliya hoos u dhignay look_weight 0.5, isu ekaanshaha labada daanyeerka ayaa halkaas ku dhammaanaya.)

Aynu aragno sida buundooyinka guud ee xayawaanku u kala baxaan marka loo eego sifada shaqsiga ah:

Miisaanka taban ee habitat_weight (-1.0), waxaanu si badheedh ah "u riixnay" daanyeerka leh deegaan la mid ah, taa beddelkeedana daanyeerka dusha sare ee deegaankoodu ka duwan yahay Harmonic Coral's - sida lagu arkay dhibcaha habitat xun ee Luminoth's iyo Zephyr Dancer. Luminoth's iyo Zephyr Dancer's buundooyinkooda behavior aad bay u sarreeyaan, taas oo muujinaysa in ay la mid yihiin dhaqankooda Harmonic Coral. Natiijooyinkooda look waa kuwo togan laakiin ka hooseeya, oo ka tarjumaya qaar laakiin aan ahayn muuqaal muuqaal ah oo la mid ah Harmonic Coral.

Marka la soo koobo, istaraatijiyadeena hoos u dhigista habitat_weight ilaa -1.0 iyo look_weight ilaa 0.5 laakiin ku haysashada behavior_weight 1.0 waxay cadeynaysaa wax ku ool ah soo saarista daanyeerka la wadaaga sifooyinka habdhaqanka muhiimka ah ee Harmonic Coral laakiin leh deegaan aad u kala duwan una eeg ugu yaraan xoogaa ka duwan.

Gabagabo

Raadinta sifooyin badan oo kala duwan ayaa ah horumar la taaban karo oo laga gaaray soo celinta macluumaadka, iyadoo bixinaysa saxsanaan ka badan, faham guud, iyo dabacsanaan marka loo eego raadinta aasaasiga ah ee isku ekaanshaha semantic. Weli, habkayaga nafiska ah (korka) - kaydinta iyo raadinta sifooyinka sifo gaar ah, ka dib isku darka natiijooyinka - waxay ku xaddidan tahay kartida, khiyaanada, iyo waxtarka marka aan u baahanahay inaan soo saarno walxaha leh sifooyin badan oo isku mar ah. (Intaa waxaa dheer, baaritaanno badan oo kNN ah ayaa qaata waqti ka badan hal baaritaan oo leh vectors isku xiran.)

Si loola tacaalo xaaladaha sidan oo kale ah, waxa fiican in aad ku kaydiso sifadaada oo dhan isla dukaanka vector oo aad samayso hal goob , adiga oo miisaamaya sifooyinkaaga wakhtiga waydiinta. Habka Superlinked waa ka saxsan yahay, ka hufan yahay, oo la miisaami karaa marka loo eego habka fudud ee codsi kasta oo u baahan degdeg, la isku halayn karo, nuanceed, soo celinta sifooyin badan - haddii kiiskaaga adeegsi uu la tacaalayo caqabadaha xogta dhabta ah ee adduunka e-ganacsigaaga ama nidaamka talada. ... ama shay gabi ahaanba ka duwan, sida la dagaalanka wuxuush.

Wax ku darsaday

• Andrey Pikunov, Qore
• Mor Kapronczay, Tafatire