Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Vespaを利用したテクいベクトル検索

sz_dr
October 10, 2024

 Vespaを利用したテクいベクトル検索

Search Engineering Tech Talk 2024 Summer
https://search-tech.connpass.com/event/331560/

sz_dr

October 10, 2024
Tweet

More Decks by sz_dr

Other Decks in Technology

Transcript

  1. ˜-:$PSQPSBUJPO  ࠓ೔͓࿩͢͠Δ͜ͱ େن໛ݴޠϞσϧɺը૾ݴޠϞσϧɺʜͷൃలʹ൐͍ɺʮϕΫτϧݕࡧʯ͸೤͍🔥τϐοΫ l ࣭໰Ԡ౴γεςϜɿϢʔβʔͷ࣭໰จͱؔ࿈ੑͷߴ͍ճ౴จΛݕࡧ l 3"(ʹ͓͚Δ3FUSJFWBM෦෼Ͱͷ׆༻ l ϚϧνϞʔμϧͳ঎඼ݕࡧγεςϜɿΩʔϫʔυͱؔ࿈ੑͷߴ͍঎඼ը૾Λݕࡧ

    ϕΫτϧݕࡧͷύϫʔΛ͞ΒʹҾ͖ग़ͨ͢Ίʹɺ l ͲΜͳ՝୊ɾղܾࡦ͕͋Δͷ͔ʁ l ݕࡧΤϯδϯ7FTQBͰ͸ͲͷΑ͏ʹ࣮૷Ͱ͖Δͷ͔ʁ Λʢ-5࿮ͳͷͰߴ଎ʹʣ͓࿩͍͖ͯ͠͠·͢ʂ ˞7FTQBɿΦʔϓϯιʔεͷCJHEBUBTFSWJOHFOHJOFɻੑೳ͕ߴ͘ɺػೳ΋๛෋ͳݕࡧΤϯδϯɻ l4FBSDI&OHJOFFSJOH5FDI5BML8JOUFSzͷୋऔ͞Μͷ͝ൃද͕ඇৗʹࢀߟʹͳΓ·͢ɻ IUUQTTFBSDIUFDIDPOOQBTTDPNFWFOU
  2. ˜-:$PSQPSBUJPO  จࣈྻҰகݕࡧͱϕΫτϧݕࡧͷ૊Έ߹Θͤ ϕΫτϧݕࡧ͕೤͍🔥ͱ͸͍͑ɺґવͱͯ͠ैདྷͷจࣈྻҰகݕࡧ͸ॏཁ ೖྗΩʔϫʔυ จࣈྻҰகݕࡧ݁Ռ ϕΫτϧݕࡧ݁Ռ ࿨ू߹ ݕࡧ݁Ռ 7FTQBͰ࣮૷͢ΔͳΒʜ

    IUUQTCMPHWFTQBBJSFEFpOJOHIZCSJETFBSDIQPTTJCJMJUJFTXJUIWFTQB ͞Βʹɺ"/%Ͱ݁߹ͨ͠Γɺෳ਺ͷϕΫτϧݕࡧ݁ՌΛ݁߹ͨ͠Γɺ݁߹݁ՌͷϦϥϯΫ΋Մೳʂ IUUQTEPDTWFTQBBJFOOFBSFTUOFJHICPSTFBSDIHVJEFIUNMNVMUJQMFOFBSFTUOFJHICPSTFBSDIPQFSBUPSTJOUIFTBNFRVFSZ จࣈྻҰகݕࡧͱϕΫτϧݕࡧͷݕࡧ݁ՌΛ૊Έ߹Θͤͯɺ݁ՌΛฦ͍ͨ͠ʂ WFTQBRVFSZ bZRMTFMFDU GSPNXIFSF\UBSHFU)JUT BQQSPYJNBUFUSVF^OFBSFTU/FJHICPS  PS VTFS2VFSZ ` lRVFSZจࣈྻҰகݕࡧΩʔϫʔυz lJOQVURVFSZ ΩʔϫʔυϕΫτϧ໊ <ϕΫτϧ> ϕΫτϧݕࡧΫΤϦ 03Ͱ݁߹ RVFSZ
  3. ˜-:$PSQPSBUJPO  είΞ͕௿͍ίϯςϯπΛݕࡧ݁Ռ͔Βআ֎ ϕΫτϧݕࡧͰ͸ɺΑ͘ྨࣅ౓্Ґ݅Λऔಘͯ͠ܝग़ʙͱ͢Δ͕ɺ ద߹ίϯςϯπ͕ຊ࣭తʹগͳ͍৔߹ɺϊΠζͩΒ͚ͷϕΫτϧݕࡧ݁Ռʹͳͬͯ͠·͏ɻɻ ྫɿϑΝογϣϯ&$αΠτͰͷݕࡧʮϛοΩʔ͕ඳ͔Εͨࠇ஍Ͱγϯϓϧͳ5γϟπʯ ˠ ΫΤϦҙਤʹԊͬͨ঎඼͸͍͍ͤͥ਺͔݅͠ͳ͍ɻɻ ҙਤͱେ͖͘ҟͳΔ঎඼͸ݕࡧ݁Ռʹग़ͨ͘͠ͳ͍ʂ ྨࣅ౓είΞͷ௿͍ίϯςϯπ͸ݕࡧ݁Ռ͔Βআ֎͍ͨ͠ʂ

    7FTQBͰ࣮૷͢ΔͳΒʜ ϥϯΩϯάఆٛ FYQSFTTJPOʢϥϯΫείΞʣͷ஋͕ΛԼճͬͨ৔߹ɺ ݕࡧ݁Ռ͔Βআ֎͢Δ SBOLQSPGJMFSBOLJOH@XJUI@TDPSF@ESPQ\ GJSTUQIBTF\ SBOLTDPSFESPQMJNJU FYQSFTTJPO\ JG ྨࣅ౓είΞ <ᮢ஋>  ྨࣅ౓είΞ ^ ^ ^ ᮢ஋ະຬͷ৔߹ɺϥϯΫείΞͱ͢Δɻ ᮢ஋Ҏ্ͷ৔߹ɺϥϯΫείΞྨࣅ౓ͱ͢Δɻ ྨࣅ౓είΞ͕ ϥϯΫείΞͷͱ͖ɺSBOLTDPSFESPQMJNJUʹҾ͔͔ͬͬͯɺ ݕࡧ݁Ռ͔Βআ֎͞ΕΔ👋 ✅ ✅ ❌ ❌ ❌ ❌ ❌ ❌ ❌ ❌ ❌ ❌ ϊΠζͩΒ͚😇 IUUQTYDPNKPCFSHVNTUBUVT IUUQTEPDTWFTQBBJFOGBRIUNMIPXUPTFUBEZOBNJDRVFSZUJNFSBOLJOHESPQUISFTIPME
  4. ˜-:$PSQPSBUJPO  ͭͷίϯςϯπʹର͠ෳ਺ͷϕΫτϧ͕ඥͮ͘έʔε ྫɿ8JLJQFEJBͷهࣄݕࡧ ϕΫτϧݕࡧਫ਼౓ΛߴΊΔͨΊʹɺͭͷʮจষʯʹର͠ʮஈམʯ͝ͱʹϕΫτϧԽ͢Δ Ҿ͖౰͍ͯͨͷ͸ʮจষʯ୯Ґ͕ͩɺ؅ཧ͢Δͷ͸ʮஈམʯ୯Ґ l ߋ৽ϑϩʔ͕൥ࡶɿʮจॻʯͷ௥Ճ΍ʮஈམʯͷมߋɾ࡟আ l ݕࡧޙॲཧ͕൥ࡶɿʮஈམʯͰҾ͖౰ͯͯɺʮจॻʯ୯ҐͰϚʔδͯ͠ɺʜ

    ָʹϕΫτϧݕࡧΛ࣮ݱ͍ͨ͠ʂ λΠτϧ ஈམ ஈམ <  > <  > ϕΫτϧݕࡧ Τϯδϯ υΩϡϝϯτ@ஈམ υΩϡϝϯτ@ஈམ υΩϡϝϯτ@ஈམ υΩϡϝϯτ@ஈམ  7FTQBͰ࣮૷͢ΔͳΒʜ ϑΟʔϧυఆٛ GJFMEQBSBHSBQI@FNCFEEJOHTUZQFUFOTPSGMPBU Q\^ FNC<> \ ϕΫτϧݕࡧ༻ઃఆ  ^ จॻ͝ͱʹஈམ਺͸ՄมͰ͋Δ͜ͱʹ஫ҙɻ ࣍ݩ໨͸NBQQFEEJNFOTJPOͰѻ͍ɺ ࣍ݩ໨͸ݻఆαΠζͷϕΫτϧͰѻ͏ɻ NFUBEBUBͱͯ͠ ֤ϊʔυ͝ͱʹ υΩϡϝϯτ*%Λ͍࣋ͬͯΔ ෳ਺ϕΫτϧ͕ඥͮ͘έʔεͷ )/48಺෦ߏ଄ IUUQTCMPHWFTQBBJTFNBOUJDTFBSDIXJUINVMUJWFDUPSJOEFYJOH
  5. ˜-:$PSQPSBUJPO  ϕΫτϧͷ֨ೲ༰ྔΛઅ໿ ྫɿ ࣍ݩͷϕΫτϧΛѻ͏৔߹ ֤࣍ݩΛGMPBU CZUFT Ͱද͢ͱɺίϯςϯπ͋ͨΓ  CZUFTඞཁ

    #JOBSJ[BUJPOʹΑͬͯ֨ೲ༰ྔΛઅ໿Ͱ͖Δ͕ɺڑ཭ܭࢉʹίαΠϯྨࣅ౓಺ੵϢʔΫϦουڑ཭͸ෆద ༰ྔઅ໿ͷςΫχοΫͷҰͭɿ#JOBSJ[BUJPO <       > <       > ͳΒ ͳΒͰɺม׵ ࣍ݩ͝ͱʹJOU CZUF Ͱද͢ͱ  ʸ࣍ݩʹѹॖ ˠ CZUFTʂ ϋϛϯάڑ཭ͰϕΫτϧݕࡧ͍ͨ͠ʂ 7FTQBͰ࣮૷͢ΔͳΒʜ ϑΟʔϧυఆٛ GJFMEFNCFEEJOHTUZQFUFOTPSJOU FNC<> \ BUUSJCVUF\ EJTUBODFNFUSJDIBNNJOH ^  ^ 7FTQBͰ͸ϕΫτϧݕࡧͷڑ཭ܭࢉʹ FVDMJEFBOBOHVMBSEPUQSPEVDUIBNNJOH ͱɺ͞·͟·ͳࣜΛར༻Մೳʂ IUUQTEPDTWFTQBBJFOSFGFSFODFTDIFNBSFGFSFODFIUNMEJTUBODFNFUSJD IUUQTCMPHWFTQBBJCJMMJPOTDBMFLOO