Indlela yokugcina i-$ 70K yokwakha iGrafu yoLwazi lwe-RAG kwi-6M yamaphepha e-Wikipedia

Ukusebenzisa iigrafu zolwazi ukuphucula iziphumo ze-retrieval-augmented generation (RAG) izicelo ziye zaba ngumxholo oshushu. Imizekelo emininzi ibonisa indlela yokwakha igrafu yolwazi kusetyenziswa inani elincinane lamaxwebhu. Oku kungenxa yokuba indlela eqhelekileyo - ukukhupha ulwazi olucokisekileyo, olugxile kwiziko - alunasikali. Ukuqhuba uxwebhu ngalunye ngomzekelo ukukhupha amaziko (iindawo) kunye nobudlelwane (imiphetho) kuthatha ixesha elide (kwaye kubiza kakhulu) ukuqhuba kwiidatha ezinkulu.

Siye saxabana oko Iigrafu zolwazi olugxile kumxholo - iVector-store evumela amakhonkco phakathi kwee-chunks - kulula ukuyisebenzisa kunye neyona ndlela isebenzayo. Apha, sibeka oko kuvavanyo. Silayisha iseti esezantsi yamanqaku eWikipedia ukusuka kwi 2wikimultihop isethi yedatha isebenzisa iindlela zombini kwaye uxoxe ukuba oku kuthetha ukuthini ukulayisha isethi yedatha iyonke. Sibonisa iziphumo zeminye imibuzo kwidatha elayishiweyo. Siza kulayisha kwakhona yonke idathasethi - phantse amaxwebhu azizigidi ezi-6 - kumxholo-ophakathi GraphVectorStore .

Iqumrhu eliphakathi: LLMGraphTransformer

Ukulayisha amaxwebhu kwivenkile yegrafu yequmrhu efana ne-Neo4j yenziwe kusetyenziswa iLLMGraphTransformer yeLangChain. Ikhowudi isekelwe kwiLangChain's "Indlela yokwenza iigrafu zolwazi."

 from langchain_core.documents import Document from langchain_experimental.graph_transformers import LLMGraphTransformer from langchain_openai import ChatOpenAI llm = ChatOpenAI(temperature=0, model_name="gpt-4-turbo") llm_transformer = LLMGraphTransformer(llm=llm) from time import perf_counter start = perf_counter() documents_to_load = [Document(page_content=line) for line in lines_to_load] graph_documents = llm_transformer.convert_to_graph_documents(documents_to_load) end = perf_counter() print(f"Loaded (but NOT written) {NUM_LINES_TO_LOAD} in {end - start:0.2f}s")

Umxholo-ophakathi: GraphVectorStore

Ukulayisha idatha kwi-GraphVectorStore kufana nje nokulayisha kwivenkile ye-vector. Ekuphela kwento eyongeziweyo kukuba sidibanisa imetadata ebonisa ukuba iphepha ngalinye lidibana njani namanye amaphepha.

 import json from langchain_core.graph_vectorstores.links import METADATA_LINKS_KEY, Link def parse_document(line: str) -> Document:    para = json.loads(line)    id = para["id"]    links = {        Link.outgoing(kind="href", tag=id)        for m in para["mentions"]        if m["ref_ids"] is not None        for id in m["ref_ids"]    }    links.add(Link.incoming(kind="href", tag=id))    return Document(        id = id,        page_content = " ".join(para["sentences"]),        metadata = {            "content_id": para["id"],            METADATA_LINKS_KEY: list(links)        },    )

Lo ngumzekelo omhle wendlela onokuthi wongeze ngayo amakhonkco akho phakathi kweenodi.

 from langchain_openai import OpenAIEmbeddings from langchain_community.graph_vectorstores.cassandra import CassandraGraphVectorStore import cassio cassio.init(auto=True) TABLE_NAME = "wiki_load" store = CassandraGraphVectorStore( embedding = OpenAIEmbeddings(), node_table=TABLE_NAME, insert_timeout = 1000.0, ) from time import perf_counter start = perf_counter() from datasets.wikimultihop.load import parse_document kg_documents = [parse_document(line) for line in lines_to_load] store.add_documents(kg_documents) end = perf_counter() print(f"Loaded (and written) {NUM_LINES_TO_LOAD} in {end - start:0.2f}s")

Ilayisha imilinganiselo

Ukuqhuba kwimigqa ye-100, indlela ye-entric-centric esebenzisa i-GPT-4o ithathe i-405.93s ukukhupha i-GraphDocuments kunye ne-10.99s ukuyibhalela i-Neo4j, ngelixa i-content-centric approach ithatha i-1.43s. Ukongezelela, kuya kuthatha iiveki ezingama-41 ukulayisha onke amaphepha angama-5,989,847 kusetyenziswa indlela yequmrhu eliphakathi kunye neeyure ezingama-24 kusetyenziswa indlela yokujonga umxholo. Kodwa ngenxa yeparallelism, indlela yokujonga umxholo isebenza kwiiyure ezi-2.5 kuphela! Ukuthatha izibonelelo ezifanayo zokunxulunyaniswa, kusenokuthatha ngaphezulu kweeveki ezine ukulayisha yonke into kusetyenziswa indlela yequmrhu. Khange ndiyizame kuba iindleko eziqikelelweyo ziya kuba yi-58,700 yeedola - ndicinga ukuba yonke into yasebenza okokuqala!

Umgca ongezantsi: indlela yequmrhu-centric yokukhupha iigrafu zolwazi kumxholo usebenzisa i-LLM yayikubini ixesha kunye neendleko ezingavumelekanga kwisikali. Ngakolunye uhlangothi, ukusebenzisa i-GraphVectorStore yayikhawuleza kwaye ingabizi.

Umzekelo iimpendulo

Kweli candelo, imibuzo embalwa, ethathwe kwi-subset yamaxwebhu alayishiwe, iyacelwa ukuba ijongane nomgangatho weempendulo.

I-Entity-centric isetyenziswe i-7324 i-tokens ekhawulezayo kwaye ixabisa i-$ 0.03 ukuvelisa iimpendulo ezingenamsebenzi ngokusisiseko, ngelixa umxholo-centric usebenzisa i-450 iithokheni ezikhawulezayo kunye nexabiso le-$ 0.002 ukuvelisa iimpendulo ezimfutshane ukuphendula imibuzo ngokuthe ngqo.

Kungamangalisa ukuba igrafu ye-Neo4j ecolekileyo ibuyisela iimpendulo ezingenamsebenzi. Ukujonga ukugawulwa kwemithi kwikhonkco, sibona ezinye zezizathu zokuba oku kusenzeka:

 > Entering new GraphCypherQAChain chain... Generated Cypher: cypher MATCH (a:Album {id: 'The Circle'})-[:RELEASED_BY]->(r:Record_label) RETURN a.id, r.id Full Context: [{'a.id': 'The Circle', 'r.id': 'Restless'}] > Finished chain. {'query': "When was 'The Circle' released?", 'result': "I don't know the answer."}

Ke, i-schema-grained schema ibuyise kuphela ulwazi malunga neleyibhile yerekhodi. Iyavakala into yokuba iLLM ayikwazanga ukuphendula umbuzo ngokusekelwe kulwazi olufunyenweyo.

Ukuqukumbela

Ukukhupha iigrafu zolwazi ezichanekileyo, ezingqamene nequmrhu lixesha- kunye neendleko-zingavumelekanga kwisikali. Xa ebuzwa imibuzo malunga ne-subset yedatha elayishiwe, i-granularity eyongezelelweyo (kunye neendleko ezongezelelweyo zokulayisha igrafu ecocekileyo) ibuyisele amathokheni amaninzi ukuba abandakanye i-prompt kodwa ivelise iimpendulo ezingenamsebenzi!

I-GraphVectorStore ithatha i-coarse-grained, indlela yomxholo-centric eyenza kube lula kwaye kulula ukwakha igrafu yolwazi. Unokuqala ngekhowudi yakho ekhoyo yokuzalisa iVectorStore usebenzisa iLangChain kwaye ungeze amakhonkco (imiphetho) phakathi kweenqununu ukuphucula inkqubo yokubuyisela.

I-RAG yeGrafu sisixhobo esiluncedo sokwenza usetyenziso lwe-AI RAG oluvelisayo ukuba lufumane kwakhona iimeko ezifanelekileyo. Kodwa ukusebenzisa i-grained-grained, indlela yequmrhu-centric ayilinganisi kwiimfuno zemveliso. Ukuba ujonge ukongeza ulwazi lwegrafu kwisicelo sakho seRAG, zama GraphVectorStore .

NguBen Chambers , DataStax