Digital Nomad
Digital Nomad
Subscribe to Kujira
Subscribe to Kujira
Share Dialog
Share Dialog
<100 subscribers
<100 subscribers
æŽç·Žããã LLM ã®éçºã«ããã RAG ã®åœ¹å²ãæ¢ããŸãã
å ã®ããã°:
https://www.spaceandtime.io/blog/vector-search-to-success
2022 幎以åã¯ãäœåãã®ãã®ã䜿ããã«ããæ°ã«å ¥ãã®æ¬ã®ç¹å®ã®äžç¯ããèŠãã°ããã®æ ç»ã®ã»ãªããããã«æãåºãããå Žåãããããæ€çŽ¢ãšã³ãžã³ãå©çšããããšã«ãªãã§ããããæ éã«äœæãããæ€çŽ¢å ¥åã§ããã³ããã衚瀺ããè¿ãããçµæãè§£æããçããå«ãŸããŠãããšæããã SparkNotes ãŸã㯠IMDB ãªã³ã¯ã«ã¢ã¯ã»ã¹ããæ°å以å ã«ããŒãžäžã§æ¢ããŠããããã¹ããèŠã€ããããšãã§ããŸãã ChatGPT ãéãããæãæåãªã¿ãŒãããŒã¿ãŒã®åèšã¯äœã§ãã?ããŸãã¯ãäºéœç©èªã®åé ã®æ®µèœãæžããŠãã ããããšå ¥åããã ãã§ãæ°ç§ã§éèªçãªçããåŸãããŸãã
å€§èŠæš¡èšèªã¢ãã« (LLM) ã®æãåçŽãªäœ¿çšæ³ã® 1 ã€ã¯ãç¥èããŒã¿ããŒã¹ãšããŠäœ¿çšããããšã§ãã LLM ã¯è±å¯ãªæ å ±ãå«ãå€§èŠæš¡ãªããŒã¿ ã»ããã§ãã¬ãŒãã³ã°ãããŠãããChatGPT ã®ãããªã€ã³ã¿ãŒãã§ã€ã¹ã«ããæ€çŽ¢ã容æã«ãªããŸããããšãã°ãæ ç»ãæžç±ã®ã³ã³ãã³ããè¿ãããã« ChatGPT ã«æç€ºããå Žåããã¬ãŒãã³ã°äžã«å ¬éãããæ å ±ãåŒã³åºãã¢ãã«ã®æ©èœãå©çšããŠããã ãã§ããããããã¿ãŒãããŒã¿ãŒã®èæ¬ã«åºã¥ããŠèšç·ŽãããŠããªãã£ããããã®éã¿ããã£ã±ã³ãºã®äœåã«åœãŠã¯ãŸããªãã£ããã©ããªãã§ãããã?æãåçŽãªãŠãŒã¹ã±ãŒã¹ (åºæ¬çãªæ å ±æ€çŽ¢ãªã©) ã«å¯ŸããŠããæãæ£ç¢ºã§é¢é£æ§ã®é«ãçµæãæäŸããã«ã¯ãLLM ã«ã¯ãå¹ åºãæ å ±ã«æ£ç¢ºã«ã¢ã¯ã»ã¹ã§ããé«åºŠãªã€ã³ããã¯ã¹äœæããã³æ€çŽ¢ã¡ã«ããºã ãå¿ èŠã§ãã
LLM ã³ã³ãã³ãã¯ã次ã®ããŒã¯ã³äºæž¬ãšåŒã°ããããã»ã¹ãéããŠçæãããŸããããã«ãããå¿çãç¶æ³ã«å¿ããŠé©åã§å€æ§æ§ãããã人éã®ãããªçè§£ãããçšåºŠåæ ããããšãä¿èšŒãããŸããæ¬¡ã®ããŒã¯ã³ã®äºæž¬ãã©ã®ããã«æ©èœããããæ®µéçã«èª¬æããŸãã
å ¥ååŠç: ããã³ãããŸãã¯è³ªåãå ¥åãããšããã®å ¥åã¯ããŒã¯ã³ãã€ãŸãåèªãŸãã¯åèªã®æçã«å€æãããŸãã
ã³ã³ããã¹ãã®çè§£: ã¢ãã«ã¯ããŠãŒã¶ãŒãäžããã¿ã°ã調ã¹ããã®ãã¬ãŒãã³ã°ã«åºã¥ããŠãçŸåšã®ãããã¯ãããŠãŒã¶ãŒã䜿çšããŠããããŒã³ãŸã§ã®ãã¹ãŠãå«ãã³ã³ããã¹ããçè§£ããããšããŸãã
次ã®ããŒã¯ã³ã®äºæž¬: ã¢ãã«ã¯ãçè§£ããã³ã³ããã¹ãã䜿çšããŠãæãå¯èœæ§ã®é«ã次ã®ããŒã¯ã³ãäœã§ããããäºæž¬ããŸããåã®åèªã«åºã¥ããŠæšæž¬ããã ãã§ã¯ãããŸããããããŸã§ã®äŒè©±ã®å šäœçãªæèãèæ ®ããŠããŸãã
ããŒã¯ã³ã®éžæ: 次ã«èããããããŒã¯ã³ã®ç¯å²ãäºæž¬ããããšã1 ã€ãéžæãããŸãããã®éžæã¯ç¢ºçãã€ãŸãã¢ãã«ããã¬ãŒãã³ã°ãããããŒã¿ã«åºã¥ããŠçºçããå¯èœæ§ãæãé«ãããŒã«ãŒã«åºã¥ããŠããŸãããã ããããã«ã¯ã©ã³ãã æ§ãããããã倿§ã§ããèªç¶ãªå¿çãçæããã®ã«åœ¹ç«ã€ããšã«æ³šæããŠãã ããã
åºåã®çæ: éžæããããŒã¯ã¢ããã¯äººéãå€èªã§ããããã¹ãã«å€æãããŸããå¿çãäžå®å šãªå Žå (é垞㯠1 ã€ã®ããŒã¯ã³ã®åŸã§ã¯ãããŸãã)ãããã»ã¹ãç¹°ãè¿ãããŸããæ°ããããŒã¯ã³ãã·ãŒã±ã³ã¹ã«è¿œå ãããã¢ãã«ã¯æŽæ°ãããã³ã³ããã¹ãã«åºã¥ããŠæ¬¡ã®ããŒã¯ã³ãäºæž¬ããŸãã
å埩æ¹è¯: 次ã®ããŒã¯ã³ãäºæž¬ããŠã·ãŒã±ã³ã¹ã«è¿œå ããããã»ã¹ã¯ãã¢ãã«ã忢ç¹ã«å°éãããŸã§ç¹°ãè¿ãããŸããããã¯ãå¿çãç¹å®ã®é·ãã«éããã¢ãã«ãæãŸãã¯æ®µèœã®çµããã瀺ãããŒã¯ã³ãäºæž¬ãããšãããŸãã¯ããã³ããã«åã蟌ãŸããæç€ºãæºãããšããªã©ãèããããŸãã
LLM ã¯ããŒã¯ã³ãäºæž¬ãããšãã«ããã®éã¿ã«åã蟌ãŸããå§çž®ãããç¥èãå¹ççã«ååŸããŠæŽ»çšããç¶æ³ã«å¿ããŠé©åãªåºåãçæããŸãããã®ããã«ãLLM ãã¬ãŒãã³ã°ã¯ããŒã¿ããŒã¹å§çž®ãåæ ããŸããããŒã¿ããŒã¹ãé »ç¹ã«ã¢ã¯ã»ã¹ãããããŒã¿ãçŽ æ©ãåŒã³åºããããã«æé©åãããŠããã®ãšåãããã«ãLLM ã¯éã¿ (ç¹å®ã®è£éã¡ã¢ãª) ããæ å ±ãååŸããããã«èšèšãããŠããŸãããã®æ©èœã«ãããé©åã«ã€ã³ããã¯ã¹ä»ããããæ å ±ãããŒã¿ããŒã¹ã«ã¯ãšãªããã®ãšåãããã«ããã¬ãŒãã³ã°äžã«ééããããç¥ãããå 容ã«å¯Ÿããã¯ãšãªã«å¯ŸããŠæ£ç¢ºãªå¿çãçæã§ããŸãããã ããã¢ãã«ãããŸã銎æã¿ã®ãªãã³ã³ãã³ããææ§ãªã³ã³ãã³ãã«ééãããšãå¶éãçããŸãã
ããšãã°ãèæžã®ç¹å®ã®äžç¯ã«ã€ã㊠LLM ã«è³ªåãããšãLLM ã¯ãããããã®ãŸãŸåŒçšããŸããããã¬ãŒãã³ã°äžã«åé·ãªã蚌人ããæããªãã£ãæŠå¿µã«ã€ããŠã¯ããã®æŠå¿µã«é¢é£ããéã¿ã倧ããããããããã®æŠå¿µããã®ãŸãŸåŒçšããããšã¯ã§ããŸããã ããã®æå³ã§ã¯ãLLM ãããŒã¿ããŒã¹ã«äŒŒãŠããŸããããŒã¿ããŒã¹ãæç€ºçã«ä¿åãããŠããããŒã¿ã®ã¿ãè¿ãã®ãšåæ§ã«ãLLM ã¯ãã¬ãŒãã³ã°äžã«ããŸãèŠãŠããªããããã¯ã®ã³ã³ãã³ããçæããã®ãé£ããå ŽåããããŸãã
ãã¡ãããLLM ã¯ãçŽç²ã«æ€çŽ¢ãè¶ ãããã®ããçè§£ãã§ããäžçã®å éšã¢ãã«ãæã£ãŠããããããã®é¡äŒŒæ§ã®ç¯å²ãè¶ ããŠããŸãããã ãããã®é床ã®åçŽåã¯ãLLM ãã³ã³ãã³ããçæããããã«ãã¬ãŒãã³ã°ãããæ¹æ³ã®ããã€ãã®éèŠãªå¶éãçè§£ããã®ã«åœ¹ç«ã¡ãŸãã
ããã«ã次ã®ããŒã¯ã³äºæž¬ã·ã¹ãã ã«ã¯ãåºç€ãšãªãããã¹ãçææ¹æ³ã«èµ·å ããä»ã®åºæã®å¶éããããŸãã
ã³ã³ããã¹ã ãŠã£ã³ã㊠ãµã€ãº: äž»ãªå¶éã® 1 ã€ã¯ãã¢ãã«ã®ã³ã³ããã¹ã ãŠã£ã³ã㊠ãµã€ãºã§ããããã¯ãã¢ãã«ãäºæž¬ãè¡ããšãã«èæ ®ã§ããããã¹ã (ããŒã¯ã³åäœ) ã®æå€§éã§ããå€ãã®ã¢ãã« (GPT ã®ä»¥åã®ããŒãžã§ã³ãå«ã) ã§ã¯ããã®ãŠã£ã³ããŠã¯é·ãäŒè©±ãããã¥ã¡ã³ãã®ã³ã³ããã¹ããç¶æã§ããã»ã©å€§ãããªããããç¶æããå¿ èŠãããé·ãããã¹ããè€éãªè°è«ã®äžè²«æ§ã倱ãããå¯èœæ§ããããŸããåã®ããŒã¯ã³ã
äžè¬åãšç¹ç°æ§: ãããã®ã¢ãã«ã¯å€§èŠæš¡ãªããŒã¿ã»ããã§ãã¬ãŒãã³ã°ãããŠããŸããããã¬ãŒãã³ã°ããäžè¬åããèœåã«ãããäžè¬çãªã³ã³ãã³ããæŒ ç¶ãšé¢é£æ§ã®ããã³ã³ãã³ããçæãããããšããããŸãããã¬ãŒãã³ã° ããŒã¿ä»¥å€ã®è©³çްãªçè§£ãæå 端ã®ââç¥èãå¿ èŠãšãããéåžžã«å ·äœçãŸãã¯åŸ®åŠãªå¿çãçæã§ããªãå ŽåããããŸãã
å€éšç¥èãžã®ã¢ã¯ã»ã¹ã®æ¬ åŠ: 次ã®ããŒã¯ã³äºæž¬ã¢ãã«ã¯ããã¬ãŒãã³ã° ããŒã¿ã»ããã«å«ãŸããæ å ±ã«éå®ãããŸãããã¬ãŒãã³ã°åŸã«æ°ããæ å ±ã«ã¢ã¯ã»ã¹ãããçµ±åãããã§ããªããšããããšã¯ãããã«æ å ±ãå€ããªã£ãããæè¿ã®åºæ¥äºãçºèŠã泚ç®ã®ãããã¯ãªã©ã®çŸåšã®ã³ã³ããã¹ããæ¬ èœãããããå¯èœæ§ãããããšãæå³ããŸãã
åçŸæ§ãšäºæž¬å¯èœæ§: 次ã®ã¿ã°äºæž¬ã®ã¢ã«ãŽãªãºã ã®æ§è³ªã«ãããããã¹ãçæãå埩çãŸãã¯äºæž¬å¯èœã«ãªãå ŽåããããŸããéåžžãã¢ãã«ã¯ã³ã³ããã¹ããèæ ®ãããšãçµ±èšçã«è¿œè·¡ãããå¯èœæ§ãé«ãããŒã¯ã³ãåªå ãããããã«ãŒãã«é¥ã£ãããäžè¬çãªãã¬ãŒãºãåªå ãããããå¯èœæ§ããããããã«ãã£ãŠåºåã®å€åæ§ãäœæžãããŸãã
äžã§è¿°ã¹ãããã«ãLLM ã¯ãã¬ãŒãã³ã°äžã«ããŒã¿ã®ããŸããŸãªåŽé¢ã«å²ãåœãŠãããéã¿ã«åºã¥ããŠå¿çãçæããŸãããããã®éã¿ã¯ãã¢ãã«ãèªèããå ¥åããŒã¿ã®ããŸããŸãªèŠçŽ ã®éèŠæ§ãåæ ããŸãããŠãŒã¶ãŒã®ããã³ããã«ããã¬ãŒãã³ã° ããŒã¿ã§ã¯ããŸã衚ãããŠããªãèŠçŽ ãå«ãŸããŠããå Žåãã¢ãã«ã¯æ£ç¢ºãªå¿çãé©åãªå¿çãçæããªãå¯èœæ§ããããŸãã
äŒè©±ã LLM ã®ã³ã³ããã¹ã ãŠã£ã³ããŠãè¶ ããå ŽåããŸãã¯ããã³ããã LLM èªèº«ã®ãã¬ãŒãã³ã° ããŒã¿ ã»ããã®éèŠãªéã¿å¶éãè¶ ããå Žå (ãŠãŒã¶ãŒãæ¢ããŠããçããæ£ç¢ºã«æãåºããªãããšãæå³ããŸã)ãã¢ãã«ã¯éåžžãå€éšã®ãã¯ãã«æ€çŽ¢ããŒã¿ããŒã¹ã«äŸåããŸããããã«ãããé¢é£ããã³ã³ããã¹ãããŠãŒã¶ãŒ ããã³ããã«è¿œå ã§ããæ°ããããŒã¿ãæ€çŽ¢ã§ããããã«ãªããŸãããã®ããã»ã¹ã¯ãæ€çŽ¢æ¡åŒµçæ (RAG) ãšåŒã°ããŸãã
RAG ããã»ã¹ã¯ããã¯ãã«æ€çŽ¢ããŒã¿ããŒã¹ (ãã¯ãã«åœ¢åŒã§ããŒã¿ãä¿åããã³ç®¡çããé«ã¬ãã«ã®ããŒã¿ããŒã¹) ãéããŠå®è£ ãããŸãããããã®ãã¯ãã«ã¯é«æ¬¡å 空éå ã®ããŒã¿ã衚ãã忬¡å ãããŒã¿ã®æå³ã®äžéšã®åŽé¢ãæããè€éãªé¢ä¿ã屿§ã®è¡šçŸãå¯èœã«ããŸããããã¹ããšèšèªã®ã³ã³ããã¹ãã§ã¯ããã¯ãã«æ€çŽ¢ããŒã¿ããŒã¹ã¯åã蟌ã¿ãªã©ã®æè¡ã䜿çšããŠããã¹ããæ°å€ãã¯ãã«ã«å€æããŸãããã®å€æã«ãããã·ã¹ãã ã¯ã倿¬¡å 空éå ã®ç°ãªãããã¹ãæçã®å¯Ÿå¿ãããã¯ãã«éã®è·é¢ãèšç®ããããšã«ãããç°ãªãããã¹ãæçéã®æå³çé¡äŒŒæ§ã枬å®ã§ããããã«ãªããŸãã
RAG äžã«ãã¯ãšãª (LLM ãžã®ãŠãŒã¶ãŒå ¥åãªã©) ãšä¿åãããããŒã¿ (èšäºãææžãæç« ãªã©) ã®äž¡æ¹ããããã¹ãåã蟌ã¿ã䜿çšããŠãã¯ãã«ã«å€æãããŸãããããã®åã蟌ã¿ã¯ããã¹ã ããŒã¿ãæ°å€ãã¯ãã«ã«å€æããåæ§ã®æå³ããã¯ãã«ç©ºéå ã®è¿ãã®ç¹ã«ãããã³ã°ãããŸããæ¬¡ã«ãããŒã¿ããŒã¹ã¯ã¯ãšãª ãã¯ãã«ãšæ ŒçŽãããããŒã¿ ãã¯ãã«ã®éã®è·é¢ãèšç®ããããã¹ãã®æå³ãã©ã®çšåºŠé¢é£ããŠãããã倿ããŸããããŒã¿ããŒã¹ã¯ããã®ãã¯ãã«ãã¯ãšãª ãã¯ãã«ã«æãè¿ãããŒã¿ ãã€ã³ã (ããã¹ã ã³ã³ãã³ã)ãã€ãŸãå ¥åãšæå³çã«æã䌌ãŠããããŒã¿ ãã€ã³ããååŸããŸãããããã®ããŒã¿ ãã€ã³ãã¯ãã³ã³ããã¹ããšæå³ã®èгç¹ãããæè¿åããšã¿ãªãããŸãã
ãããã®æè¿åã¯ãLLM ãç¬èªã®ãã¬ãŒãã³ã° ããŒã¿å ã§ã¢ã¯ã»ã¹ã§ããªãå¯èœæ§ããããæèäžé¢é£ããè¿œå æ å ±ã®åºç€ãæäŸããLLM ã®åºåã®ç²ŸåºŠãé¢é£æ§ãè±å¯ãã倿§æ§ãå€§å¹ ã«åäžãããããšãã§ããŸãã Sam Altman æ°ãã¯ãã¢ãã«ã®åŸ®èª¿æŽã®ã¿ã§ã¯ãªããRAG ãå©çšããŠãšãŒãžã§ã³ããéçºãããæåã®ããã®ãã¯ãã«æ¢çŽ¢ãã¢ãããŒããæå±ããŠããŸãã
LLM ã®åŸ®èª¿æŽã«ã¯ãç¹å®ã®ããŒã¿ã»ããã§ã®è¿œå ãã¬ãŒãã³ã°ã«åºã¥ããŠã¢ãã«ã®éã¿ã調æŽããŠãç¹å®ã®ã¿ã¹ã¯ã®ããã©ãŒãã³ã¹ãåäžãããããäžéšã®ãã¡ã€ã³ã®çè§£ãæ·±ãããããããšãå«ãŸããŸãããã®ããã»ã¹ã¯ã€ãããŒã·ã§ã³ã®é床ãããé ããã€ãŸãã埮調æŽãããã¢ãã«ã¯æŽæ°ãããã®ãšã»ãŒåãéãã§å€ããªã£ãŠããŸãã ãã§ãªããæ°ããããŒã¿ã«ã察å¿ã§ããŸããã
å¯Ÿç §çã«ãRAG ã䜿çšãããšãã¢ãã«ã¯ãªã¢ã«ã¿ã€ã ã§å€éšããŒã¿ããŒã¹ã«ã¢ã¯ã»ã¹ããçŸåšã®ã¯ãšãªã«é¢é£ããææ°ã®æ å ±ãååŸã§ããŸããåºç€ãšãªãã¢ãã«ãæè¿æŽæ°ãŸãã¯åŸ®èª¿æŽãããŠããªãå Žåã§ããææ°ã®ããŒã¿ãå«ãå¿çãçæã§ããŸããã¢ãã«ã¯ãå€éšã®æ å ±ãœãŒã¹ãååŸããŠã³ã³ããã¹ããçè§£ããããšã§ãæ°ããããŒã¿ã倿Žã«é©å¿ã§ãããããé¢é£æ§ãããé·ãä¿ãããŸãã
RAG ã¯ã埮調æŽã«å ããŠãæšæº LLM ã«é¢é£ããåè¿°ã®èª²é¡ã«ã察åŠããŸãã
ã³ã³ããã¹ãã®çè§£ã®æ¡åŒµ: RAG ã¯ãã¢ãã«ã®å¿çã匷åããææ°æ å ±ãŸãã¯è©³çްæ å ±ãååŸããããšã§ãåŸæ¥ã® LLM ã®ã³ã³ããã¹ã ãŠã£ã³ããŠãæ¡åŒµããŸãã
ç¹ç°æ§ãšç²ŸåºŠã®åäž: RAG ã䜿çšãããšãã¢ãã«ã¯ãã¬ãŒãã³ã°äžã«åŠç¿ãããã¿ãŒã³ã®ã¿ã«äŸåããã®ã§ã¯ãªããååŸããããã¥ã¡ã³ãããã®ç¹å®ã®è©³çްãå¿çã«æ¿å ¥ã§ãããããã¢ãã«ã®ç²ŸåºŠãåäžããã ãã§ãªããçŸåšã®ç¶æ³ã«åãããŠããå ·äœçã«åãåãããããšãã§ããŸãã
éè€ãšäºæž¬å¯èœæ§ã®åæž: RAG ã¯ãã¯ãšãªããšã«ç°ãªãæ å ±ã»ãããåçã«æœåºããããšã§ãã¢ãã«ã®å¿çãå€§å¹ ã«å€æŽã§ããŸãããã®å€åæ§ã¯ãå€éšããŒã¿ã«ãã£ãŠäŒè©±ã«æ°ãã衚çŸã詳现ãå°å ¥ããããããçŽç²ã«çæã¢ãã«ã«ããèŠãããå埩æ§ãäºæž¬å¯èœæ§ã軜æžããã®ã«åœ¹ç«ã¡ãŸãã
ãã ããRAG ã«ã¯ãé å»¶ãšã€ã³ããªãžã§ã³ã¹ã®æ¬ åŠãšããç¬èªã®èª²é¡ããããŸãããŠãŒã¶ãŒãããã³ãããéä¿¡ããLLM ã远å ã®ã³ã³ããã¹ããå¿ èŠã§ããããšã瀺ãããã€ãã®ããŒã¯ã³ãåãåºãããã¯ã¿ãŒãããŒã¿ããŒã¹ãæ€çŽ¢ããŠãŠãŒã¶ãŒã®å ¥åããã³ãããä»ããŠæãè¿ãã³ã³ããã¹ããååŸããæçµçã«äž¡æ¹ãéä¿¡ãããã¿ãŒã³ããŒã¹ã®ãšãŒãžã§ã³ã ãã£ãããããã®äŒè©±ãèããŠã¿ãŸããããæšè«ã®ããã«åã³ LLM ã«éä¿¡ãããŸããæ¬¡ã«ããŠãŒã¶ãŒãè¿ä¿¡ããçªã«ãªããŸãã
ãã®ã·ã¹ãã ã§ã¯ãåãŠãŒã¶ãŒ ããã³ãããè€æ°ã¹ãããã®æäœãéå§ããåã¹ããããåèšåŠçæéã«å ç®ãããŸããããã»ã¹å šäœã®é床ã¯ããã¯ãã«æ€çŽ¢ããŒã¿ããŒã¹ãå¿ èŠãªã³ã³ããã¹ããã©ãã ãæ©ãååŸã§ãããã«ãäŸåããŸããããŒã¿ããŒã¹ ã¯ãšãªãè€éãªå ŽåããŸãã¯ããŒã¿ããŒã¹èªäœã倧ãããŠæé©ãªã€ã³ããã¯ã¹ããªãå Žåããã®ååŸã«ããå€§å¹ ãªé å»¶ãçºçããå¯èœæ§ããããŸããããã«ãç¹ã«ããè€éãªäŒè©±ã§ã¯ãå¿çãå®å šã«æ¹è¯ããããã«ããã®äžé£ã®çæãšååŸãäœåºŠãç¹°ãè¿ãå¿ èŠãããå ŽåããããŸãããã®å埩ã«ãŒãã«ããã¬ã€ãã³ã·ãŒãå¢å ããå éšããŒã¿ã®ã¿ã«äŸåããçŽç²ãªçæã¢ãã«ããã察話ãé ããªããŸãã
ããã«ãRAG ãªãã LLM ã®ã€ã³ããªãžã§ã³ã¹ã¯ããã¯ãã«æ€çŽ¢ããŒã¿ããŒã¹ããååŸãããæ å ±ã®å質ãšé¢é£æ§ã«å€§ããäŸåããŸããããŒã¿ããŒã¹ãå æ¬çã§ãªããææ°ã§ããªããååã«ç¶æãããŠããªãå ŽåãååŸããæ å ±ã®æçšæ§ãå¶éãããå¿çã®å šäœçãªã€ã³ããªãžã§ã³ã¹ã«åœ±é¿ãäžããå¯èœæ§ããããŸãã
é«å質ã®å€éšããŒã¿ãååŸãããå Žåã§ãããã®æ å ±ã LLM ã®æ¢åã®å¿çãã¬ãŒã ã¯ãŒã¯ã«å¹æçã«çµ±åããæ¹æ³ã課é¡ãšããŠæ®ããŸããã¢ãã«ã¯ãã®å€éšããŒã¿ãçµã¿èŸŒãã ãã§ãªããç¶æ³ã«å¿ããŠé©åãã€äžè²«ããæ¹æ³ã§çµã¿èŸŒãå¿ èŠããããŸããã¢ãã«ã®ãã¬ãŒãã³ã°ãšå€éšããŒã¿ã®æ§è³ªã®éã«äžäžèŽããããšãæè¡çã«ã¯æ£ç¢ºã§ããã³ã³ããã¹ãçã«ã¯åãé¢ãããå¿çãçºçããå¯èœæ§ããããŸãã
次äžä»£ã® LLM ã§ã¯ããã¯ãã«æ€çŽ¢ããŒã¹ã® RAG ãšãåŸæ¥ã®ãã¬ãŒãã³ã°/åŸ®èª¿æŽæ¹æ³ããã³æ§é åããŒã¿åŠç (TradFi åžå ŽããŒã¿ãé¢é£ããéèãã¥ãŒã¹ã® SQL ããŒã¿ããŒã¹ãªã©) ãæ··åãããå¯èœæ§ããããŸãã ããããã« 1 ã€ã® LLM ãããã€ããŒããããããããã«å¥åã®ãã¯ãã«æ€çŽ¢ããŒã¿ããŒã¹ããããšããæŠå¿µã¯ãã€ã³ããã¯ã¹äœæäœæ¥ã¡ã¢ãªããã©ãã€ãã®ãã¯ãã«åã³ã³ããã¹ããæã€ããŒã«ã« SSD ã«çŽæçã«æ¡åŒµããæ°ããã¢ãã«ã«ãã£ãŠäœç³»åãããŸãã
Space and Time ã¯ãã§ã«ãSQL ããŒã¿ããŒã¹åŠçã®ç²ŸåºŠãšèæ¹ããæ§ãæ€èšŒãã ZK ãã«ãŒãã§ãã Proof of SQL ãé¡§å®¢ã«æäŸããŠããŸããããã«æè¿ã§ã¯ããã¯ã¿ãŒæ€çŽ¢ã®ååŸãæäŸãã Proof of Vector Search ãåæ§ã®å¹æãæäŸããŠããŸãããããã®æ°ãã蚌æã¯ãLLM ãæ°ããã³ã³ããã¹ããçµ±åããããåºç¯å²ã§ãã詳现ãªããŒã¿ç¯å²ã«ãªã¢ã«ã¿ã€ã ã§ã¢ã¯ã»ã¹ããæ§é åããŒã¿åŠçãçµ±åããŠãããæŽå¯ã«æºã¡ãåæããã¹ãŠè¿œè·¡å¯èœã§æ€èšŒå¯èœãªæ¹æ³ã§è¡ãããšãã§ããæªæ¥ãžã®éãéããŸãããããã®é²æ©ã«ãããæçµçã« LLM ã®é©çšç¯å²ãæ¡å€§ããéèãµãŒãã¹ããã¥ãŒã¹éçŽããªã¹ã¯è©äŸ¡ãªã©ãææ°ã®ããŒã¿ã«å€§ããäŸåããåéã§ã®æçšæ§ãæ¡å€§ããã€ãããŒã·ã§ã³äž»å°å AI ã®æ¬¡ã®æ³¢ãæšé²ããŸãã
æŽç·Žããã LLM ã®éçºã«ããã RAG ã®åœ¹å²ãæ¢ããŸãã
å ã®ããã°:
https://www.spaceandtime.io/blog/vector-search-to-success
2022 幎以åã¯ãäœåãã®ãã®ã䜿ããã«ããæ°ã«å ¥ãã®æ¬ã®ç¹å®ã®äžç¯ããèŠãã°ããã®æ ç»ã®ã»ãªããããã«æãåºãããå Žåãããããæ€çŽ¢ãšã³ãžã³ãå©çšããããšã«ãªãã§ããããæ éã«äœæãããæ€çŽ¢å ¥åã§ããã³ããã衚瀺ããè¿ãããçµæãè§£æããçããå«ãŸããŠãããšæããã SparkNotes ãŸã㯠IMDB ãªã³ã¯ã«ã¢ã¯ã»ã¹ããæ°å以å ã«ããŒãžäžã§æ¢ããŠããããã¹ããèŠã€ããããšãã§ããŸãã ChatGPT ãéãããæãæåãªã¿ãŒãããŒã¿ãŒã®åèšã¯äœã§ãã?ããŸãã¯ãäºéœç©èªã®åé ã®æ®µèœãæžããŠãã ããããšå ¥åããã ãã§ãæ°ç§ã§éèªçãªçããåŸãããŸãã
å€§èŠæš¡èšèªã¢ãã« (LLM) ã®æãåçŽãªäœ¿çšæ³ã® 1 ã€ã¯ãç¥èããŒã¿ããŒã¹ãšããŠäœ¿çšããããšã§ãã LLM ã¯è±å¯ãªæ å ±ãå«ãå€§èŠæš¡ãªããŒã¿ ã»ããã§ãã¬ãŒãã³ã°ãããŠãããChatGPT ã®ãããªã€ã³ã¿ãŒãã§ã€ã¹ã«ããæ€çŽ¢ã容æã«ãªããŸããããšãã°ãæ ç»ãæžç±ã®ã³ã³ãã³ããè¿ãããã« ChatGPT ã«æç€ºããå Žåããã¬ãŒãã³ã°äžã«å ¬éãããæ å ±ãåŒã³åºãã¢ãã«ã®æ©èœãå©çšããŠããã ãã§ããããããã¿ãŒãããŒã¿ãŒã®èæ¬ã«åºã¥ããŠèšç·ŽãããŠããªãã£ããããã®éã¿ããã£ã±ã³ãºã®äœåã«åœãŠã¯ãŸããªãã£ããã©ããªãã§ãããã?æãåçŽãªãŠãŒã¹ã±ãŒã¹ (åºæ¬çãªæ å ±æ€çŽ¢ãªã©) ã«å¯ŸããŠããæãæ£ç¢ºã§é¢é£æ§ã®é«ãçµæãæäŸããã«ã¯ãLLM ã«ã¯ãå¹ åºãæ å ±ã«æ£ç¢ºã«ã¢ã¯ã»ã¹ã§ããé«åºŠãªã€ã³ããã¯ã¹äœæããã³æ€çŽ¢ã¡ã«ããºã ãå¿ èŠã§ãã
LLM ã³ã³ãã³ãã¯ã次ã®ããŒã¯ã³äºæž¬ãšåŒã°ããããã»ã¹ãéããŠçæãããŸããããã«ãããå¿çãç¶æ³ã«å¿ããŠé©åã§å€æ§æ§ãããã人éã®ãããªçè§£ãããçšåºŠåæ ããããšãä¿èšŒãããŸããæ¬¡ã®ããŒã¯ã³ã®äºæž¬ãã©ã®ããã«æ©èœããããæ®µéçã«èª¬æããŸãã
å ¥ååŠç: ããã³ãããŸãã¯è³ªåãå ¥åãããšããã®å ¥åã¯ããŒã¯ã³ãã€ãŸãåèªãŸãã¯åèªã®æçã«å€æãããŸãã
ã³ã³ããã¹ãã®çè§£: ã¢ãã«ã¯ããŠãŒã¶ãŒãäžããã¿ã°ã調ã¹ããã®ãã¬ãŒãã³ã°ã«åºã¥ããŠãçŸåšã®ãããã¯ãããŠãŒã¶ãŒã䜿çšããŠããããŒã³ãŸã§ã®ãã¹ãŠãå«ãã³ã³ããã¹ããçè§£ããããšããŸãã
次ã®ããŒã¯ã³ã®äºæž¬: ã¢ãã«ã¯ãçè§£ããã³ã³ããã¹ãã䜿çšããŠãæãå¯èœæ§ã®é«ã次ã®ããŒã¯ã³ãäœã§ããããäºæž¬ããŸããåã®åèªã«åºã¥ããŠæšæž¬ããã ãã§ã¯ãããŸããããããŸã§ã®äŒè©±ã®å šäœçãªæèãèæ ®ããŠããŸãã
ããŒã¯ã³ã®éžæ: 次ã«èããããããŒã¯ã³ã®ç¯å²ãäºæž¬ããããšã1 ã€ãéžæãããŸãããã®éžæã¯ç¢ºçãã€ãŸãã¢ãã«ããã¬ãŒãã³ã°ãããããŒã¿ã«åºã¥ããŠçºçããå¯èœæ§ãæãé«ãããŒã«ãŒã«åºã¥ããŠããŸãããã ããããã«ã¯ã©ã³ãã æ§ãããããã倿§ã§ããèªç¶ãªå¿çãçæããã®ã«åœ¹ç«ã€ããšã«æ³šæããŠãã ããã
åºåã®çæ: éžæããããŒã¯ã¢ããã¯äººéãå€èªã§ããããã¹ãã«å€æãããŸããå¿çãäžå®å šãªå Žå (é垞㯠1 ã€ã®ããŒã¯ã³ã®åŸã§ã¯ãããŸãã)ãããã»ã¹ãç¹°ãè¿ãããŸããæ°ããããŒã¯ã³ãã·ãŒã±ã³ã¹ã«è¿œå ãããã¢ãã«ã¯æŽæ°ãããã³ã³ããã¹ãã«åºã¥ããŠæ¬¡ã®ããŒã¯ã³ãäºæž¬ããŸãã
å埩æ¹è¯: 次ã®ããŒã¯ã³ãäºæž¬ããŠã·ãŒã±ã³ã¹ã«è¿œå ããããã»ã¹ã¯ãã¢ãã«ã忢ç¹ã«å°éãããŸã§ç¹°ãè¿ãããŸããããã¯ãå¿çãç¹å®ã®é·ãã«éããã¢ãã«ãæãŸãã¯æ®µèœã®çµããã瀺ãããŒã¯ã³ãäºæž¬ãããšãããŸãã¯ããã³ããã«åã蟌ãŸããæç€ºãæºãããšããªã©ãèããããŸãã
LLM ã¯ããŒã¯ã³ãäºæž¬ãããšãã«ããã®éã¿ã«åã蟌ãŸããå§çž®ãããç¥èãå¹ççã«ååŸããŠæŽ»çšããç¶æ³ã«å¿ããŠé©åãªåºåãçæããŸãããã®ããã«ãLLM ãã¬ãŒãã³ã°ã¯ããŒã¿ããŒã¹å§çž®ãåæ ããŸããããŒã¿ããŒã¹ãé »ç¹ã«ã¢ã¯ã»ã¹ãããããŒã¿ãçŽ æ©ãåŒã³åºããããã«æé©åãããŠããã®ãšåãããã«ãLLM ã¯éã¿ (ç¹å®ã®è£éã¡ã¢ãª) ããæ å ±ãååŸããããã«èšèšãããŠããŸãããã®æ©èœã«ãããé©åã«ã€ã³ããã¯ã¹ä»ããããæ å ±ãããŒã¿ããŒã¹ã«ã¯ãšãªããã®ãšåãããã«ããã¬ãŒãã³ã°äžã«ééããããç¥ãããå 容ã«å¯Ÿããã¯ãšãªã«å¯ŸããŠæ£ç¢ºãªå¿çãçæã§ããŸãããã ããã¢ãã«ãããŸã銎æã¿ã®ãªãã³ã³ãã³ããææ§ãªã³ã³ãã³ãã«ééãããšãå¶éãçããŸãã
ããšãã°ãèæžã®ç¹å®ã®äžç¯ã«ã€ã㊠LLM ã«è³ªåãããšãLLM ã¯ãããããã®ãŸãŸåŒçšããŸããããã¬ãŒãã³ã°äžã«åé·ãªã蚌人ããæããªãã£ãæŠå¿µã«ã€ããŠã¯ããã®æŠå¿µã«é¢é£ããéã¿ã倧ããããããããã®æŠå¿µããã®ãŸãŸåŒçšããããšã¯ã§ããŸããã ããã®æå³ã§ã¯ãLLM ãããŒã¿ããŒã¹ã«äŒŒãŠããŸããããŒã¿ããŒã¹ãæç€ºçã«ä¿åãããŠããããŒã¿ã®ã¿ãè¿ãã®ãšåæ§ã«ãLLM ã¯ãã¬ãŒãã³ã°äžã«ããŸãèŠãŠããªããããã¯ã®ã³ã³ãã³ããçæããã®ãé£ããå ŽåããããŸãã
ãã¡ãããLLM ã¯ãçŽç²ã«æ€çŽ¢ãè¶ ãããã®ããçè§£ãã§ããäžçã®å éšã¢ãã«ãæã£ãŠããããããã®é¡äŒŒæ§ã®ç¯å²ãè¶ ããŠããŸãããã ãããã®é床ã®åçŽåã¯ãLLM ãã³ã³ãã³ããçæããããã«ãã¬ãŒãã³ã°ãããæ¹æ³ã®ããã€ãã®éèŠãªå¶éãçè§£ããã®ã«åœ¹ç«ã¡ãŸãã
ããã«ã次ã®ããŒã¯ã³äºæž¬ã·ã¹ãã ã«ã¯ãåºç€ãšãªãããã¹ãçææ¹æ³ã«èµ·å ããä»ã®åºæã®å¶éããããŸãã
ã³ã³ããã¹ã ãŠã£ã³ã㊠ãµã€ãº: äž»ãªå¶éã® 1 ã€ã¯ãã¢ãã«ã®ã³ã³ããã¹ã ãŠã£ã³ã㊠ãµã€ãºã§ããããã¯ãã¢ãã«ãäºæž¬ãè¡ããšãã«èæ ®ã§ããããã¹ã (ããŒã¯ã³åäœ) ã®æå€§éã§ããå€ãã®ã¢ãã« (GPT ã®ä»¥åã®ããŒãžã§ã³ãå«ã) ã§ã¯ããã®ãŠã£ã³ããŠã¯é·ãäŒè©±ãããã¥ã¡ã³ãã®ã³ã³ããã¹ããç¶æã§ããã»ã©å€§ãããªããããç¶æããå¿ èŠãããé·ãããã¹ããè€éãªè°è«ã®äžè²«æ§ã倱ãããå¯èœæ§ããããŸããåã®ããŒã¯ã³ã
äžè¬åãšç¹ç°æ§: ãããã®ã¢ãã«ã¯å€§èŠæš¡ãªããŒã¿ã»ããã§ãã¬ãŒãã³ã°ãããŠããŸããããã¬ãŒãã³ã°ããäžè¬åããèœåã«ãããäžè¬çãªã³ã³ãã³ããæŒ ç¶ãšé¢é£æ§ã®ããã³ã³ãã³ããçæãããããšããããŸãããã¬ãŒãã³ã° ããŒã¿ä»¥å€ã®è©³çްãªçè§£ãæå 端ã®ââç¥èãå¿ èŠãšãããéåžžã«å ·äœçãŸãã¯åŸ®åŠãªå¿çãçæã§ããªãå ŽåããããŸãã
å€éšç¥èãžã®ã¢ã¯ã»ã¹ã®æ¬ åŠ: 次ã®ããŒã¯ã³äºæž¬ã¢ãã«ã¯ããã¬ãŒãã³ã° ããŒã¿ã»ããã«å«ãŸããæ å ±ã«éå®ãããŸãããã¬ãŒãã³ã°åŸã«æ°ããæ å ±ã«ã¢ã¯ã»ã¹ãããçµ±åãããã§ããªããšããããšã¯ãããã«æ å ±ãå€ããªã£ãããæè¿ã®åºæ¥äºãçºèŠã泚ç®ã®ãããã¯ãªã©ã®çŸåšã®ã³ã³ããã¹ããæ¬ èœãããããå¯èœæ§ãããããšãæå³ããŸãã
åçŸæ§ãšäºæž¬å¯èœæ§: 次ã®ã¿ã°äºæž¬ã®ã¢ã«ãŽãªãºã ã®æ§è³ªã«ãããããã¹ãçæãå埩çãŸãã¯äºæž¬å¯èœã«ãªãå ŽåããããŸããéåžžãã¢ãã«ã¯ã³ã³ããã¹ããèæ ®ãããšãçµ±èšçã«è¿œè·¡ãããå¯èœæ§ãé«ãããŒã¯ã³ãåªå ãããããã«ãŒãã«é¥ã£ãããäžè¬çãªãã¬ãŒãºãåªå ãããããå¯èœæ§ããããããã«ãã£ãŠåºåã®å€åæ§ãäœæžãããŸãã
äžã§è¿°ã¹ãããã«ãLLM ã¯ãã¬ãŒãã³ã°äžã«ããŒã¿ã®ããŸããŸãªåŽé¢ã«å²ãåœãŠãããéã¿ã«åºã¥ããŠå¿çãçæããŸãããããã®éã¿ã¯ãã¢ãã«ãèªèããå ¥åããŒã¿ã®ããŸããŸãªèŠçŽ ã®éèŠæ§ãåæ ããŸãããŠãŒã¶ãŒã®ããã³ããã«ããã¬ãŒãã³ã° ããŒã¿ã§ã¯ããŸã衚ãããŠããªãèŠçŽ ãå«ãŸããŠããå Žåãã¢ãã«ã¯æ£ç¢ºãªå¿çãé©åãªå¿çãçæããªãå¯èœæ§ããããŸãã
äŒè©±ã LLM ã®ã³ã³ããã¹ã ãŠã£ã³ããŠãè¶ ããå ŽåããŸãã¯ããã³ããã LLM èªèº«ã®ãã¬ãŒãã³ã° ããŒã¿ ã»ããã®éèŠãªéã¿å¶éãè¶ ããå Žå (ãŠãŒã¶ãŒãæ¢ããŠããçããæ£ç¢ºã«æãåºããªãããšãæå³ããŸã)ãã¢ãã«ã¯éåžžãå€éšã®ãã¯ãã«æ€çŽ¢ããŒã¿ããŒã¹ã«äŸåããŸããããã«ãããé¢é£ããã³ã³ããã¹ãããŠãŒã¶ãŒ ããã³ããã«è¿œå ã§ããæ°ããããŒã¿ãæ€çŽ¢ã§ããããã«ãªããŸãããã®ããã»ã¹ã¯ãæ€çŽ¢æ¡åŒµçæ (RAG) ãšåŒã°ããŸãã
RAG ããã»ã¹ã¯ããã¯ãã«æ€çŽ¢ããŒã¿ããŒã¹ (ãã¯ãã«åœ¢åŒã§ããŒã¿ãä¿åããã³ç®¡çããé«ã¬ãã«ã®ããŒã¿ããŒã¹) ãéããŠå®è£ ãããŸãããããã®ãã¯ãã«ã¯é«æ¬¡å 空éå ã®ããŒã¿ã衚ãã忬¡å ãããŒã¿ã®æå³ã®äžéšã®åŽé¢ãæããè€éãªé¢ä¿ã屿§ã®è¡šçŸãå¯èœã«ããŸããããã¹ããšèšèªã®ã³ã³ããã¹ãã§ã¯ããã¯ãã«æ€çŽ¢ããŒã¿ããŒã¹ã¯åã蟌ã¿ãªã©ã®æè¡ã䜿çšããŠããã¹ããæ°å€ãã¯ãã«ã«å€æããŸãããã®å€æã«ãããã·ã¹ãã ã¯ã倿¬¡å 空éå ã®ç°ãªãããã¹ãæçã®å¯Ÿå¿ãããã¯ãã«éã®è·é¢ãèšç®ããããšã«ãããç°ãªãããã¹ãæçéã®æå³çé¡äŒŒæ§ã枬å®ã§ããããã«ãªããŸãã
RAG äžã«ãã¯ãšãª (LLM ãžã®ãŠãŒã¶ãŒå ¥åãªã©) ãšä¿åãããããŒã¿ (èšäºãææžãæç« ãªã©) ã®äž¡æ¹ããããã¹ãåã蟌ã¿ã䜿çšããŠãã¯ãã«ã«å€æãããŸãããããã®åã蟌ã¿ã¯ããã¹ã ããŒã¿ãæ°å€ãã¯ãã«ã«å€æããåæ§ã®æå³ããã¯ãã«ç©ºéå ã®è¿ãã®ç¹ã«ãããã³ã°ãããŸããæ¬¡ã«ãããŒã¿ããŒã¹ã¯ã¯ãšãª ãã¯ãã«ãšæ ŒçŽãããããŒã¿ ãã¯ãã«ã®éã®è·é¢ãèšç®ããããã¹ãã®æå³ãã©ã®çšåºŠé¢é£ããŠãããã倿ããŸããããŒã¿ããŒã¹ã¯ããã®ãã¯ãã«ãã¯ãšãª ãã¯ãã«ã«æãè¿ãããŒã¿ ãã€ã³ã (ããã¹ã ã³ã³ãã³ã)ãã€ãŸãå ¥åãšæå³çã«æã䌌ãŠããããŒã¿ ãã€ã³ããååŸããŸãããããã®ããŒã¿ ãã€ã³ãã¯ãã³ã³ããã¹ããšæå³ã®èгç¹ãããæè¿åããšã¿ãªãããŸãã
ãããã®æè¿åã¯ãLLM ãç¬èªã®ãã¬ãŒãã³ã° ããŒã¿å ã§ã¢ã¯ã»ã¹ã§ããªãå¯èœæ§ããããæèäžé¢é£ããè¿œå æ å ±ã®åºç€ãæäŸããLLM ã®åºåã®ç²ŸåºŠãé¢é£æ§ãè±å¯ãã倿§æ§ãå€§å¹ ã«åäžãããããšãã§ããŸãã Sam Altman æ°ãã¯ãã¢ãã«ã®åŸ®èª¿æŽã®ã¿ã§ã¯ãªããRAG ãå©çšããŠãšãŒãžã§ã³ããéçºãããæåã®ããã®ãã¯ãã«æ¢çŽ¢ãã¢ãããŒããæå±ããŠããŸãã
LLM ã®åŸ®èª¿æŽã«ã¯ãç¹å®ã®ããŒã¿ã»ããã§ã®è¿œå ãã¬ãŒãã³ã°ã«åºã¥ããŠã¢ãã«ã®éã¿ã調æŽããŠãç¹å®ã®ã¿ã¹ã¯ã®ããã©ãŒãã³ã¹ãåäžãããããäžéšã®ãã¡ã€ã³ã®çè§£ãæ·±ãããããããšãå«ãŸããŸãããã®ããã»ã¹ã¯ã€ãããŒã·ã§ã³ã®é床ãããé ããã€ãŸãã埮調æŽãããã¢ãã«ã¯æŽæ°ãããã®ãšã»ãŒåãéãã§å€ããªã£ãŠããŸãã ãã§ãªããæ°ããããŒã¿ã«ã察å¿ã§ããŸããã
å¯Ÿç §çã«ãRAG ã䜿çšãããšãã¢ãã«ã¯ãªã¢ã«ã¿ã€ã ã§å€éšããŒã¿ããŒã¹ã«ã¢ã¯ã»ã¹ããçŸåšã®ã¯ãšãªã«é¢é£ããææ°ã®æ å ±ãååŸã§ããŸããåºç€ãšãªãã¢ãã«ãæè¿æŽæ°ãŸãã¯åŸ®èª¿æŽãããŠããªãå Žåã§ããææ°ã®ããŒã¿ãå«ãå¿çãçæã§ããŸããã¢ãã«ã¯ãå€éšã®æ å ±ãœãŒã¹ãååŸããŠã³ã³ããã¹ããçè§£ããããšã§ãæ°ããããŒã¿ã倿Žã«é©å¿ã§ãããããé¢é£æ§ãããé·ãä¿ãããŸãã
RAG ã¯ã埮調æŽã«å ããŠãæšæº LLM ã«é¢é£ããåè¿°ã®èª²é¡ã«ã察åŠããŸãã
ã³ã³ããã¹ãã®çè§£ã®æ¡åŒµ: RAG ã¯ãã¢ãã«ã®å¿çã匷åããææ°æ å ±ãŸãã¯è©³çްæ å ±ãååŸããããšã§ãåŸæ¥ã® LLM ã®ã³ã³ããã¹ã ãŠã£ã³ããŠãæ¡åŒµããŸãã
ç¹ç°æ§ãšç²ŸåºŠã®åäž: RAG ã䜿çšãããšãã¢ãã«ã¯ãã¬ãŒãã³ã°äžã«åŠç¿ãããã¿ãŒã³ã®ã¿ã«äŸåããã®ã§ã¯ãªããååŸããããã¥ã¡ã³ãããã®ç¹å®ã®è©³çްãå¿çã«æ¿å ¥ã§ãããããã¢ãã«ã®ç²ŸåºŠãåäžããã ãã§ãªããçŸåšã®ç¶æ³ã«åãããŠããå ·äœçã«åãåãããããšãã§ããŸãã
éè€ãšäºæž¬å¯èœæ§ã®åæž: RAG ã¯ãã¯ãšãªããšã«ç°ãªãæ å ±ã»ãããåçã«æœåºããããšã§ãã¢ãã«ã®å¿çãå€§å¹ ã«å€æŽã§ããŸãããã®å€åæ§ã¯ãå€éšããŒã¿ã«ãã£ãŠäŒè©±ã«æ°ãã衚çŸã詳现ãå°å ¥ããããããçŽç²ã«çæã¢ãã«ã«ããèŠãããå埩æ§ãäºæž¬å¯èœæ§ã軜æžããã®ã«åœ¹ç«ã¡ãŸãã
ãã ããRAG ã«ã¯ãé å»¶ãšã€ã³ããªãžã§ã³ã¹ã®æ¬ åŠãšããç¬èªã®èª²é¡ããããŸãããŠãŒã¶ãŒãããã³ãããéä¿¡ããLLM ã远å ã®ã³ã³ããã¹ããå¿ èŠã§ããããšã瀺ãããã€ãã®ããŒã¯ã³ãåãåºãããã¯ã¿ãŒãããŒã¿ããŒã¹ãæ€çŽ¢ããŠãŠãŒã¶ãŒã®å ¥åããã³ãããä»ããŠæãè¿ãã³ã³ããã¹ããååŸããæçµçã«äž¡æ¹ãéä¿¡ãããã¿ãŒã³ããŒã¹ã®ãšãŒãžã§ã³ã ãã£ãããããã®äŒè©±ãèããŠã¿ãŸããããæšè«ã®ããã«åã³ LLM ã«éä¿¡ãããŸããæ¬¡ã«ããŠãŒã¶ãŒãè¿ä¿¡ããçªã«ãªããŸãã
ãã®ã·ã¹ãã ã§ã¯ãåãŠãŒã¶ãŒ ããã³ãããè€æ°ã¹ãããã®æäœãéå§ããåã¹ããããåèšåŠçæéã«å ç®ãããŸããããã»ã¹å šäœã®é床ã¯ããã¯ãã«æ€çŽ¢ããŒã¿ããŒã¹ãå¿ èŠãªã³ã³ããã¹ããã©ãã ãæ©ãååŸã§ãããã«ãäŸåããŸããããŒã¿ããŒã¹ ã¯ãšãªãè€éãªå ŽåããŸãã¯ããŒã¿ããŒã¹èªäœã倧ãããŠæé©ãªã€ã³ããã¯ã¹ããªãå Žåããã®ååŸã«ããå€§å¹ ãªé å»¶ãçºçããå¯èœæ§ããããŸããããã«ãç¹ã«ããè€éãªäŒè©±ã§ã¯ãå¿çãå®å šã«æ¹è¯ããããã«ããã®äžé£ã®çæãšååŸãäœåºŠãç¹°ãè¿ãå¿ èŠãããå ŽåããããŸãããã®å埩ã«ãŒãã«ããã¬ã€ãã³ã·ãŒãå¢å ããå éšããŒã¿ã®ã¿ã«äŸåããçŽç²ãªçæã¢ãã«ããã察話ãé ããªããŸãã
ããã«ãRAG ãªãã LLM ã®ã€ã³ããªãžã§ã³ã¹ã¯ããã¯ãã«æ€çŽ¢ããŒã¿ããŒã¹ããååŸãããæ å ±ã®å質ãšé¢é£æ§ã«å€§ããäŸåããŸããããŒã¿ããŒã¹ãå æ¬çã§ãªããææ°ã§ããªããååã«ç¶æãããŠããªãå ŽåãååŸããæ å ±ã®æçšæ§ãå¶éãããå¿çã®å šäœçãªã€ã³ããªãžã§ã³ã¹ã«åœ±é¿ãäžããå¯èœæ§ããããŸãã
é«å質ã®å€éšããŒã¿ãååŸãããå Žåã§ãããã®æ å ±ã LLM ã®æ¢åã®å¿çãã¬ãŒã ã¯ãŒã¯ã«å¹æçã«çµ±åããæ¹æ³ã課é¡ãšããŠæ®ããŸããã¢ãã«ã¯ãã®å€éšããŒã¿ãçµã¿èŸŒãã ãã§ãªããç¶æ³ã«å¿ããŠé©åãã€äžè²«ããæ¹æ³ã§çµã¿èŸŒãå¿ èŠããããŸããã¢ãã«ã®ãã¬ãŒãã³ã°ãšå€éšããŒã¿ã®æ§è³ªã®éã«äžäžèŽããããšãæè¡çã«ã¯æ£ç¢ºã§ããã³ã³ããã¹ãçã«ã¯åãé¢ãããå¿çãçºçããå¯èœæ§ããããŸãã
次äžä»£ã® LLM ã§ã¯ããã¯ãã«æ€çŽ¢ããŒã¹ã® RAG ãšãåŸæ¥ã®ãã¬ãŒãã³ã°/åŸ®èª¿æŽæ¹æ³ããã³æ§é åããŒã¿åŠç (TradFi åžå ŽããŒã¿ãé¢é£ããéèãã¥ãŒã¹ã® SQL ããŒã¿ããŒã¹ãªã©) ãæ··åãããå¯èœæ§ããããŸãã ããããã« 1 ã€ã® LLM ãããã€ããŒããããããããã«å¥åã®ãã¯ãã«æ€çŽ¢ããŒã¿ããŒã¹ããããšããæŠå¿µã¯ãã€ã³ããã¯ã¹äœæäœæ¥ã¡ã¢ãªããã©ãã€ãã®ãã¯ãã«åã³ã³ããã¹ããæã€ããŒã«ã« SSD ã«çŽæçã«æ¡åŒµããæ°ããã¢ãã«ã«ãã£ãŠäœç³»åãããŸãã
Space and Time ã¯ãã§ã«ãSQL ããŒã¿ããŒã¹åŠçã®ç²ŸåºŠãšèæ¹ããæ§ãæ€èšŒãã ZK ãã«ãŒãã§ãã Proof of SQL ãé¡§å®¢ã«æäŸããŠããŸããããã«æè¿ã§ã¯ããã¯ã¿ãŒæ€çŽ¢ã®ååŸãæäŸãã Proof of Vector Search ãåæ§ã®å¹æãæäŸããŠããŸãããããã®æ°ãã蚌æã¯ãLLM ãæ°ããã³ã³ããã¹ããçµ±åããããåºç¯å²ã§ãã詳现ãªããŒã¿ç¯å²ã«ãªã¢ã«ã¿ã€ã ã§ã¢ã¯ã»ã¹ããæ§é åããŒã¿åŠçãçµ±åããŠãããæŽå¯ã«æºã¡ãåæããã¹ãŠè¿œè·¡å¯èœã§æ€èšŒå¯èœãªæ¹æ³ã§è¡ãããšãã§ããæªæ¥ãžã®éãéããŸãããããã®é²æ©ã«ãããæçµçã« LLM ã®é©çšç¯å²ãæ¡å€§ããéèãµãŒãã¹ããã¥ãŒã¹éçŽããªã¹ã¯è©äŸ¡ãªã©ãææ°ã®ããŒã¿ã«å€§ããäŸåããåéã§ã®æçšæ§ãæ¡å€§ããã€ãããŒã·ã§ã³äž»å°å AI ã®æ¬¡ã®æ³¢ãæšé²ããŸãã
No activity yet