ã¯ããã«
ååãChatGPTã瀟å ã«é ã£ãŠãããŸã䜿ãããªãæ¬åœã®çç±ãã«ãããŠãçŽ ã®ChatGPTãé ã£ãŠãæã£ããã瀟å ã§äœ¿ãããªãã®ã¯ãChatGPTã¯ç€Ÿå æ å ±ãæã£ãŠããªããããšããäºãæžããŸããã
ã€ãŸããChatGPTãç€Ÿå æ å ±ã«ã粟éãããšãã¹ããŒããšããŠæ¥åã§ã䜿ã£ãŠããã«ã¯ãå ¬éããŒã¿ã§ã¯ãªãç€Ÿå æ å ±ãããã«ChatGPTã«èªã¿èŸŒãŸããããšããäºããã€ã³ãã«ãªããŸãã
Fine-Tuningãšããéžæè¢ããã¡ãããããŸãããåŠç¿æ
å ±ã®ã¢ããããŒããå¶åŸ¡ãšè©äŸ¡ãæ
å ±å
¬éå
ã®æš©éèšå®çã®éçšé¢ã®é£ãããããããŸãã以äžã®MSã®è«æã§ããè€æ°ã¿ã¹ã¯ã§ã®æ¯èŒã§ãFine-TuningããRAGã®ã»ããããã©ãŒãã³ã¹ãè¯ããšããå®éšçµæã玹ä»ãããŠããŸãã
åèïŒFine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs
äžé·äžçã§ã¯ããã®ã§ãäžæŠã«ã©ã¡ããè¯ããšããããã§ã¯ãããŸããããçŸæç¹ã§ã¯ãããããã¡ãŒã¹ããªãã·ã§ã³ãšãªãRAGã«ã€ããŠãRAGã®ç²ŸåºŠåäžã«äœ¿ããäž»èŠãªæ¡åŒµææ³ããŸãšãã€ã€ãRAGã®å®è£ æŠç¥ã«ã€ããŠèããŠãããããšæããŸãã
RAGã®æ¡åŒµææ³
RAGã®æ¡åŒµææ³ã«ã€ããŠã¯ãã¢ã€ãã£ã¢ããŒã¹ã®ç©ãå«ãäžãã ããšããããªãã®ã§ãããä»åã¯OSSã§ãå®è£ ãããŠããäž»èŠãªæ¡åŒµææ³ãäžå¿ã«ãŸãšããŠãããŸãã
| # | ææ³ | æŠèŠ |
|---|---|---|
| 1 | ãã€ããªãããµãŒã | è€æ°ã®æ€çŽ¢æ¹åŒã®çµã¿åããã«ããæ€çŽ¢æ§èœã®åäž |
| 2 | ãªã©ã³ãã³ã° | æ€çŽ¢ã§åŸãææžã«å¯ŸããŠãå床ã©ã³ã¯ä»ããè¡ããé¢é£æ§ãé«ãææžã®ã¿ãå©çš |
| 3 | ãµãã¯ãšãª | è€åçãªã¯ãšãªããµãã¯ãšãªã«åè§£ããŠãããããã®ã¬ã¹ãã³ã¹çµæãçµ±å |
| 4 | HyDE | ã¯ãšãªã«å¯Ÿããä»®åçãçæãããä»®åçã§æ€çŽ¢ |
| 5 | ã¹ãããããã¯ããã³ãã | ã¯ãšãªãäžæ®µæœè±¡åãããæœè±¡åããã¯ãšãªãžã®æ€çŽ¢çµæãåæããŠåç |
| 6 | RAG Fusion | é¡äŒŒã¯ãšãªãçæãããåã¯ãšãªã§ã®æ€çŽ¢çµæãçµ±åããŠåçãçæ |
| 7 | ãã«ãã¹ãããã¯ãšãª | ã¯ãšãªã®åè§£ãšæ®µéçãªæšé²ã«ãããåçãæŽç·Žãããææ³ |
| 8 | ãã£ã³ã¯æ¡åŒµ | æ€çŽ¢ã§ãããããååŸã®ãã£ã³ã¯ãã³ã³ããã¹ãã«å«ãã |
| 9 | Pandas Dataframe | CSVã®ãããªããŒã¿ããŒãã«ã®æ§é ãä¿ã£ããŸãŸã¯ãšãªãå¯èœ |
| 10 | TextToSQL | èªç¶èšèªãSQLã«å€æããŠãããŒã¿ããŒã¹ããå¿ èŠãªæ å ±ãæœåº |
ããã§ã¯ä»¥äžãç°¡åãªãã¢ãå«ããªããåå¥ã«èª¬æããŠãããŸãã
1. ãã€ããªãããµãŒã
ãã¡ãã¯ãæ§è³ªã®ç°ãªãè€æ°ã®æ€çŽ¢æ¹åŒïŒäŸãã°ãã¯ãã«æ€çŽ¢ãšããŒã¯ãŒãæ€çŽ¢ïŒãçµã¿åãããŠæ€çŽ¢ç²ŸåºŠãåäžãããææ³ã«ãªããŸãã
åæ€çŽ¢æ¹åŒåäœã®å Žåã«æ¯ã¹ãæ§è³ªã®ç°ãªãæ€çŽ¢æ¹åŒãçµã¿åãããããçš®ãããšãåããããäºã§ãæ€çŽ¢æ§èœã®åäžãæåŸ ã§ããŸãã
ä»åã¯BM25ã§ã®ããŒã¯ãŒãããŒã¹ã®é¡äŒŒåºŠæ€çŽ¢ãšéåžžã®ãã¯ãã«æ€çŽ¢ãçµã¿åãããŠãããŸãã
BM25ã«ã€ããŠç°¡åã«èª¬æããŠãããšãæèãæç« æ§é ã¯å®å šã«ç¡èŠããäžã§ãææžå ã®åèªãå šãŠãã©ãã©ã«åå²ããææžå ã®ååèªã®åºçŸé »åºŠãšææžéã«ãããã¬ã¢åºŠãå å³ããç¹åŸŽéãç®åºããŸãã
ã€ãŸããç¹å®ã®ææžå ã®ååèªã®æ°ãã«ãŠã³ãããŠãã¹ãã°ã©ã ãäœãã°ã䌌ããããªææžã«ã¯åããããªåèªãããåºãã¯ããªã®ã§(åããããªãã¹ãã°ã©ã ã®åœ¢ã«ãªãã®ã§)ãé¡äŒŒåºŠãé«ããªãæ§è³ªãæã€ãšããç¹åŸŽéã«ãªããŸãã
ãã ãã©ã®ææžã«ãããåºãåèªã¯ç¹åŸŽéãšããŠããŸãæå³ããªããªãã®ã§ãææžéã®ã¬ã¢åºŠãå å³ããŠãããç¹å®ã®ææžã«ããåºãªãåèªã«ã€ããŠã¯ããã®éã¿ã倧ããããäºã§ãç¹åŸŽéãšããŠã®è¡šçŸåãäžããŠããŸãã
æç³»åçãªè©±ããããšãTFIDFãBM25ãšãã£ãææ³ãã¡ãžã£ãŒã§ãã£ãæã«BERTãåºãŠããŠãåèªãæç« ã®æå³ãèžãŸããŠãã¯ãã«åãããšããæµããåºãŠããŸããã
ã€ãŸãããããã¯ã©ã¡ãããšãããšå€å žçãªæ©æ¢°åŠç¿ã®ææ³ã«åé¡ãããŸãããAzure AI Searchãæ§ã ãªæ€çŽ¢ãšã³ãžã³ã§ãçŸåœ¹ã§äœ¿ãããŠããæçšãªææ³ã«ãªããŸãã
ãããšèããã ãã ãšãåèªåäœã§ãã©ãã©ã«ããŠããŸã£ãŠå€§äžå€«ãªã®ïŒããšããæ°ããããããããŸãããããããŸã§æå³çè§£ã§ã¯ãªãææžã®é¡äŒŒåºŠæ€çŽ¢ã®ç¯å²ã§ã¯è²ã ãšã¡ãªãããåºãŠããŸãã
ç¹ã«éèŠã«ãªãã®ãåºæåè©ã§ãããšããã®ããç€Ÿå æ å ±ã«ã¯äžéã§äœ¿ãããäºã®ãªã倿°ã®åºæåè©ããããæ€çŽ¢ã®éã¯ãã®åºæåè©ãããŒã«ãªãããšãå°ãªããããŸããã
äžæ¹ã§ãã¯ãã«æ€çŽ¢ã¯ãæç« å šäœãæå³ãã¯ãã«ã«å€æãããããå šäœçãªæå³ã¯äžæãæããäºãã§ããŸãããåºæåè©ãå«ãååèªã®åœ±é¿èªäœã¯èããªã£ãŠããŸããŸãã
BERTãåºãé ã«ãã£ãäºããã人ã¯èæãããããšæããŸããããã¥ãŒã¹ã®ã«ããŽãªåé¡ããã¬ããžåæãšãã£ããã¯ãçãªåŠçã¯äžæãäžæ¹ã§ãåºæåè©ã倿°å ¥ãç€Ÿå ææžã®é¡äŒŒåºŠæ€çŽ¢ã¯ãTFIDFãBM25ã®ããã«ãåºæåè©ãå«ããåèªããã®ãŸãŸäœ¿ããææ³ã®ã»ãã粟床ãåºããããŸãã
ã€ãŸããæå³ã«åŒ·ããã¯ãã«æ€çŽ¢ãšããŒã¯ãŒãã«åŒ·ãBM25ã®æ€çŽ¢ã®ãã€ããªããã«ããäºã§ãæ€çŽ¢ç²ŸåºŠãåäžããããšããç®çã«ãªããŸãã
å°ãå眮ããé·ããªããŸããããç°¡åãªãã¢ã§èŠãŠãããŸãããã
ãµã³ãã«ã¯äœã§ãè¯ãã®ã§ãããä»åã¯ãDXã¬ããŒã2.2ãã䜿ã£ãŠãããŸãã
å°ã極端ãªäŸã§ã¯ãããŸãããåããããããéèŠããããã«ãå š14ããŒãžã®PDFã®äžã§1ç®æã ãåºãŠããã以äžã®ããã©ãããã©ãŒã ããšããåèªãçããããšæããŸãã
ããã©ãããã©ãŒã ã«ã€ããŠã®èšåã¯ãããŸããïŒããšããã¯ãšãªã«å¯Ÿããããã¯ãã«æ€çŽ¢ãšBM25ã®æ€çŽ¢çµæã®éããããããããã£ã³ã¯ã®äžäœ5åãŸã§ãæœåºããŠèŠãŠã¿ãŸãã
ãŸãã¯ãã¯ãã«æ€çŽ¢ããã§ãã
â ãã¯ãã«æ€çŽ¢ã§ã®é¡äŒŒãã£ã³ã¯ååŸçµæïŒäžäœ5åïŒ
ãã©ãããã©ãŒã ã«ã€ããŠèšåãããŠããç®æ(p13)ã¯å«ãŸããŠããŸãããé äœãšããŠã¯4çªç®ã§ãããé¡äŒŒåºŠã0.8以äžãšãªã£ãŠããŸãã
ããã©ã«ãã®ååŸãã£ã³ã¯æ°èšå®(äžäœ2å)ããé¡äŒŒåºŠã®éŸå€ã®ã«ãããªãã䜿ãå Žåã¯(0.8以äžã®ã¿ååŸç)ãç¡é¢ä¿ãªãã£ã³ã¯ãšããŠèœãšããŠããŸãå¯èœæ§ããããŸãã
ç¶ããŠBM25ã§ã®æ€çŽ¢çµæã§ãã
â BM25ã§ã®é¡äŒŒãã£ã³ã¯ååŸçµæïŒäžäœ5åïŒ
ãã¡ãã¯ãã¯ãã«æ€çŽ¢ãšéãããã©ãããã©ãŒã ããšããåèªãå«ãŸããp13ãæäžäœã«æ¥ãŠããŸãã
ãã®ããã«ãæå³ãã¯ãã«ã«å€æããã«åèªããã®ãŸãŸäœ¿ãããããç¹å®ã®åç§°ã«ããããããããšããå©ç¹ãèŠãŠãšããŸãã
åãªãå šææ€çŽ¢ã§ã¯ãªããã¯ãšãªã«å«ãŸããååèªã®åºçŸãã¿ãŒã³ãšé¡äŒŒãããã£ã³ã¯ãååŸããŠãããããããã¯ãã«æ€çŽ¢ã®èŠæãªæãäžæãã«ããŒããŠãããææ³ãšèšããŸãã
以äžã®MSã®ãã³ãããŒã¯ã§ããããŒã¯ãŒãããŒã¹ã®æ€çŽ¢ãšãã¯ãã«æ€çŽ¢ãçµã¿åããããã€ããªããæ€çŽ¢ã®ç²ŸåºŠãé«ãåºããšããçµæãå ±åãããŠããŸãã
åŒçšïŒAzure AI Search: Outperforming vector search with hybrid retrieval and ranking capabilities
ãã€ããªããæ€çŽ¢ãããããã±ãŒã¹ã§è¯ããšããããã§ã¯ãªãã§ããããã¯ãã«æ€çŽ¢ã§äžæããããããªãå Žåã¯ããã€ããªããæ€çŽ¢ã詊ããŠã¿ããšããã®ã¯æå¹ãªéžæè¢ã«ãªãããšæããŸãã
2. ãªã©ã³ãã³ã°
ç¶ããŠãªã©ã³ãã³ã°ãèŠãŠãããŸãã
ãªã©ã³ãã³ã°ã¯å ã»ã©ã®ããã«æ€çŽ¢ã§ååŸãããã£ã³ã¯ããæŽã«ã¯ãšãªãšã®é¡äŒŒåºŠãé«ãé ã«äžŠã³å€ããææ³ã«ãªããŸãã
ãã¯ãã«æ€çŽ¢ã¯æå³çãªæ€çŽ¢ãã§ãããšèšããŸããããæç« 衚çŸãäžã€ã®æå³ãã¯ãã«ã«æŒã蟌ãã§ãããããäžå®ã®æ å ±éã¯åœç¶èœã¡ãŠããŸãã
ããã§ããã衚çŸåã®é«ãèšèªã¢ãã«ã䜿ã£ãŠãã¯ãšãªãšæœåºãããã£ã³ã¯ã®é¡äŒŒåºŠãåç®åºããé¡äŒŒåºŠã®é«ãé ã«å床䞊ã³å€ããäžäœãã£ã³ã¯ã®ã¿ã䜿ããšããææ³ã«ãªããŸãã
ãåæ€çŽ¢ããããããªãããªãæåãã䜿ããªãã®ïŒããšæããããããŸãããããã¯ãã«æ€çŽ¢ãåçŽãªãã¯ãã«ã®å ç©èšç®ã§ããã®ã«æ¯ã¹ãèšèªã¢ãã«ã¯èšç®ã³ã¹ããéåžžã«é«ããããã¥ã¡ã³ãå šäœã«é©çšããã®ã¯çŸå®çã§ã¯ãªãããã§ãã
ãã®ããããã¯ãã«æ€çŽ¢ãBM25ã§å¹ åºãé¢é£ããããªãã£ã³ã¯ãååŸãããã®äžã§ããªã©ã³ãã³ã°ã«ãã£ãŠäžäœãã£ã³ã¯ã ããæœåºãããšãã圢ã«ããäºã§ãLLMã«ã³ã³ããã¹ããšããŠæž¡ãåç §æ å ±ã®ç²ŸåºŠãäžããäºãã§ããŸãã
Azure AI Searchã®ã»ãã³ãã£ãã¯ã©ã³ãã³ã°ã§ããBM25çã§äžäœ50ãã£ã³ã¯ãå ã«æœåºããŠããããªã©ã³ã¯ãããšãããã€ãã©ã€ã³ã«ãªã£ãŠããŸãã
ãªã©ã³ãã³ã°ã«ã¯TransformerãLLMã䜿ããã®ããããŸãããä»åã¯SentenceTransformerã䜿ã£ããªã©ã³ãã³ã°ã詊ããŠãããŸãã
ãªã©ã³ãã³ã°èªäœã¯åçš®OSSã«å®è£ ãããŠãããããç°¡åã«è©Šãäºãã§ããŸãã
以äžããå ã»ã©ã®ãã¯ãã«æ€çŽ¢çµæããªã©ã³ãã³ã°ãããçµæã§ãã
â ãã¯ãã«æ€çŽ¢ã®ãªã©ã³ãã³ã°çµæ
å ã»ã©ã¯4çªç®ã ã£ãp13ã®è©²åœç®æããããã®é äœã«ãªã£ãŠããŸãã
ãã®ããã«ããã¯ãã«æ€çŽ¢ããã€ããªããæ€çŽ¢çã§å¹ åºãæ€çŽ¢ãããã£ã³ã¯ããªã©ã³ãã³ã°ããäžäœãã£ã³ã¯ã ãã«çµã£ãŠLLMã«æž¡ãäºãã§ããŸãã
ããããTransformerã䜿ãã®ã¯èšç®è³æºçã«ããŒãã«ãé«ãé¢ãããããããã®ãªã©ã³ãã³ã°ãLLMã«ãããããšããæ¹æ³ããããŸãã
以äžãLlamaIndexã®LLMRerankã®ããã³ãããµã³ãã«ã§ããããã£ãŠããäºã¯ã·ã³ãã«ã§ãã¯ãšãªãšæç« ãæž¡ããŠã¯ãšãªãšé¢é£åºŠã®é«ãé ã«äžŠã³ãããããŠããŸãã
æåã®ãã¯ãã«æ€çŽ¢ã®çµæãLLMRerankã«æž¡ããå Žåããå ã»ã©ãšåãããã«p13ãæäžäœã«äžŠã³å€ããããŸãã
ãã ãRerankã¯2段éã®ããã»ã¹ã«ãªã以äžãå°ãªããšãã¬ã¹ãã³ã¹æéã¯äŒžã³ãããããã¡ããéåžžã®ãã¯ãã«æ€çŽ¢ã§äžååãªã±ãŒã¹ã§ã®å©çšãæ€èšããã®ãè¯ãã§ãããã
ç¹ã«ãèŠæš¡ã®å€§ããªããã¥ã¡ã³ããããå€å°ã¬ã¹ãã³ã¹ãç ç²ã«ããŠã粟床è¯ãåããããšããã±ãŒã¹ã§ã¯éèŠã«ãªãããšæããŸãã
ã¡ãªã¿ã«å€§éã®ã³ã³ããã¹ãããã®ãŸãŸæž¡ããã«ãRerankã§äžäœãã£ã³ã¯ã®ã¿ãæž¡ãäºããªãéèŠãªã®ãã«ã€ããŠã¯ãåŸåã®ã³ã³ããã¹ãé·ã®é·ãã®åœ±é¿ã®ç®æã§æ¹ããŠè§ŠããããšæããŸãã
3. ãµãã¯ãšãª
ãã¡ãã¯ãå ¥åã¯ãšãªããã®ãŸãŸãã¯ãã«åããŠæ€çŽ¢ã«å©çšããã®ã§ã¯ãªããå ¥åã¯ãšãªããµãã¯ãšãªã«åè§£ããããããã®ãµãã¯ãšãªã§ã®ã¬ã¹ãã³ã¹ãæåŸã«åæããŠåçãããšããææ³ã«ãªããŸãã
ç¹ã«å ¥åã¯ãšãªãè€éãªå Žåããè€åçãªæå³ãæã£ãŠããå Žåã«æå¹ãªææ³ãšèšããŸãã
ãã¡ããå®éã®äŸãèŠãŠãããŸãããããµã³ãã«ãšããŠãDXã¬ããŒãã«é¢ãã以äžã®ã¯ãšãªãèããŠã¿ãŸãã
â å
¥åã¯ãšãª
ãäœäœå®å®ãšããžã¿ã«ç£æ¥å®£èšã®æå³ãããããæããŠã
ãã®ã¯ãšãªã¯1æã§ãããæå³çã«ã¯2ã€ã®äºãèããŠããè€åçãªã¯ãšãªã«ãªã£ãŠããŸãã
ãã®ãŸãŸãã¯ãã«åãããšãã©ã¡ãã®æå³ãå Œãããããªäžéçãªãã¯ãã«ãçæãããææžå ã®é¡äŒŒåºŠãæ€çŽ¢ããäºã«ãªããŸããããã®2ã€ã¯ç¬ç«ã®é ç®ã§ãããããäžéçãªãã¯ãã«ã§ã®æ€çŽ¢ã«ãã£ãŠãã©ã¡ãã«ãããããã«ãããªããšããå¯èœæ§ãåºãŠããŸãã
äžæ¹ã§ãå ¥åã¯ãšãªãæå³çã«åè§£ãããµãã¯ãšãªã«ããäºã§ãããããã®æå³ãä¿æããäžã§ãé¡äŒŒåºŠã®æ€çŽ¢ãã§ããŸãã
å®éã«çæããããµãã¯ãšãªãšåçã¯ä»¥äžã«ãªããŸãïŒå°ãé·ãã§ããããã®ãŸãŸèŒããŸãïŒã
â å
¥åã¯ãšãª
ãäœäœå®å®ãšããžã¿ã«ç£æ¥å®£èšã®æå³ãããããæããŠã
â LLMã«ãã£ãŠçæããããµãã¯ãšãª
ãµãã¯ãšãª1ïŒäœäœå®å®ã®è©³çްãªå®çŸ©ãšãã®æå³ã¯äœã§ããïŒ
ãµãã¯ãšãª2ïŒããžã¿ã«ç£æ¥å®£èšã®è©³çްãªå®çŸ©ãšãã®æå³ã¯äœã§ããïŒ
â ãµãã¯ãšãªãžã®åç
ããµãã¯ãšãª1ãžã®åçã
"äœäœå®å®"ã¯ããŠãŒã¶ãŒäŒæ¥ãšãã³ããŒäŒæ¥ãåºå®ãããé¢ä¿ãæããŸãããã®é¢ä¿ã§ã¯ããŠãŒã¶ãŒäŒæ¥ã¯ITãæŽ»çšããŠæ¥åã®å¹çåãç®æãã仿§ãäœããITãã³ããŒãç«¶äºãããŠITã³ã¹ããåæžããŸããäžæ¹ããã³ããŒäŒæ¥ã¯åŽåéã«å¯Ÿãã察䟡ãšããŠå€ä»ãããäœãªã¹ã¯ã®ããžãã¹ã享åããŸãããããããã®é¢ä¿ã¯ãã³ããŒä»»ãã«ãªãããšã§IT察å¿èœåãè²ãããITã·ã¹ãã ããã©ãã¯ããã¯ã¹åãããã³ããŒããã¯ã€ã³ã«ããçµå¶ã®ã¢ãžãªãã£ãäœäžããŸãããŸãã顧客ãžã®è¿
éãªäŸ¡å€ææ¡ãã§ããªããªããäœãå©çæ°Žæºã«ããå€éäžè«ãæ§é ãçãã売ãäžãç·éã®ç¢ºä¿ãå¿
èŠãšãªããŸããããã«ãåŽåéãäžããããçç£æ§åäžã®ã€ã³ã»ã³ãã£ããåãããäœå©ççã®ããæè¡éçºæè³ãå°é£ã«ãªããŸãããããã®çµæã"ããžã¿ã«"ã®ææ¡ãã§ããªããªããŸãã
ããµãã¯ãšãª2ãžã®åçã
ããžã¿ã«ç£æ¥å®£èšã¯ãããžã¿ã«ã掻çšããŠåçåäžãéæããç¹åŸŽãå
šç€Ÿã«æµžéãããããã®ãè¡åæéãã§ããã瀟äŒéåè«çã¢ãããŒãã®å®è·µã«ããããçµå¶è
èªãã®ã䟡å€èгããå€éšãžçºä¿¡ããããã«çå®ãããŸãããã®å®£èšã¯ãDXæšé²ã®èŠç¯çäŒæ¥ããåŸãããé
ç®ãåºã«ãçµå¶è
ã®çè§£ãé«ããããã«5é
ç®ã«éçŽãããŠããŸããããã«ãçµå¶è
ãèªãã®å®£èšãšããŠç·Žãäžããããšãæå³ããŠããŸãã
â æçµã¬ã¹ãã³ã¹
äœäœå®å®ã¯ããŠãŒã¶ãŒäŒæ¥ãšãã³ããŒäŒæ¥ãåºå®ãããé¢ä¿ã瀺ãçšèªã§ãããã®é¢ä¿ã§ã¯ããŠãŒã¶ãŒäŒæ¥ã¯ITãæŽ»çšããŠæ¥åã®å¹çåãç®æããITãã³ããŒãç«¶äºãããŠITã³ã¹ããåæžããŸããäžæ¹ããã³ããŒäŒæ¥ã¯åŽåéã«å¯Ÿãã察䟡ãšããŠå€ä»ãããäœãªã¹ã¯ã®ããžãã¹ã享åããŸãããããããã®é¢ä¿ã¯äžé£ã®åé¡ãåŒãèµ·ããå¯èœæ§ããããŸãã
äžæ¹ãããžã¿ã«ç£æ¥å®£èšã¯ãããžã¿ã«ã掻çšããŠåçåäžãéæããç¹åŸŽãå
šç€Ÿã«æµžéãããããã®è¡åæéã§ããããã¯ãçµå¶è
èªãã®äŸ¡å€èгãå€éšãžçºä¿¡ããããã«çå®ãããDXæšé²ã®èŠç¯çäŒæ¥ããåŸãããé
ç®ãåºã«ãçµå¶è
ã®çè§£ãé«ããããã«5é
ç®ã«éçŽãããŠããŸãã
ãŒãŒãŒãŒãŒãŒãŒãŒãŒãŒãŒãŒãŒãŒ
è€åçãªã¯ãšãªãåå¥ã®ãµãã¯ãšãªã«åè§£ãããããçš®ã·ã£ãŒããªæ€çŽ¢ãšåçãå®çŸã§ããŠããŸãã
ã¯ãšãªã®åè§£ãšåçã®åæã¯åæãããå¿ èŠããããŸããããµãã¯ãšãªåäœã®åŠçã¯äžŠååã§ããããããµãã¯ãšãªãå¢ããåãç·åœ¢ã§ã¬ã¹ãã³ã¹ãé ããªããšããããã§ã¯ãããŸããã
ãŠãŒã¹ã±ãŒã¹ã«ããããŸãããç¹ã«å ¥åã¯ãšãªãè€éã§ãã£ãããé·ãåŸåã«ããå Žåããã®ãŸãŸãã¯ãã«åããããããµãã¯ãšãªåäœã§åŠçããæ¹ã广çãªå ŽåããããããããŸããã
4. HyDE
HyDEã¯Hypothetical Document Embeddingsã®ç¥ã§ãå ¥åã¯ãšãªãçŽæ¥ãã¯ãã«åãããã®ã§ã¯ãªããLLMã«ä»®ã®åçãäœãããŠããã®ä»®ã®åçããã¯ãã«åããŠæ€çŽ¢ã«äœ¿ããšããæ¹æ³ã§ãã
ãã®ã¢ãããŒã·ã§ã³ãã©ãã«ããã®ããšãããšããã¯ãã«æ€çŽ¢ã¯ãå ¥åã¯ãšãªãšããã¥ã¡ã³ãã®é¡äŒŒåºŠãé«ãç®æãæ€çŽ¢ããŸãããããã¯ããã°ãQ&Aã§èšããšãQãšAãåœãŠã«ããé¢ä¿ãªãã®ã§ãå¿ ããããããè¿ããšã¯éããªãããã§ãã
äŸãã°ãããã¥ã¡ã³ãã«ã¬ã¹ãã©ã³ã®æ å ±ã®äžèЧããã£ããšããŠããäœãçŸå³ããç©ãé£ã¹ããããšããã¯ãšãªãšããã¥ã¡ã³ãã®ã¬ã¹ãã©ã³ã®æ å ±ã®é¡äŒŒåºŠãé«ãããšèšããšãå¿ ããããããšã¯éããŸããã
ãããããããäœãçŸå³ããç©ãé£ã¹ããããšããã¯ãšãªã«å¯ŸããŠãLLMã«ä»®ã®åçãçæããïŒäŸãã°ãè¿ãã«ããæçã®ãåºããããŸãããšãããããªä»®åç)ããã®ä»®åçã§æ€çŽ¢ãããã»ãããQ&Aã§èšããšãAå士ã§ããããããäºã«ãªãã®ã§ãç¹ã«ã¯ãšãªãšåçã®å 容ã倧ããç°ãªããããªã±ãŒã¹ã§æå¹ã«ãªãã®ã§ã¯ãªãããšããäºã§ãã
è«ççã«ã¯ç¢ºãã«ããã ãªãšæãã€ã€ããLlamaIndexã®ããã¥ã¡ã³ãã«ãäžæããããªãã±ãŒã¹ã«ã€ããŠã®èšåããããããä»åã¯ãããŠå€±æããäºäŸãèŒããŠãããŸãã
DXã¬ããŒãã«å¯ŸããŠããäœäœå®å®ãšã¯ïŒããšããã¯ãšãªã§è©ŠããŠã¿ãŸãããã
äœäœå®å®èªäœã¯ã以äžã®æå³ã«ãªããŸãã
ãäœäœå®å®ãšã¯ããŠãŒã¶ãŒäŒæ¥ãšãã³ããŒäŒæ¥ãåºå®ãããé¢ä¿ã«ããç¶æ
ãæããŸãããã®ç¶æ
ã§ã¯ããŠãŒã¶ãŒäŒæ¥ã¯ITã®æŽ»çšç®çãšããŠæ¥åã®å¹çåã远æ±ããITãã³ããŒãç«¶äºãããŠITã³ã¹ããåæžããŸããäžæ¹ããã³ããŒäŒæ¥ã¯åŽåéã«å¯Ÿãã察䟡ãšããŠå€ä»ãããäœãªã¹ã¯ã®ããžãã¹ã享åããŸãããããããã®é¢ä¿ã¯IT察å¿èœåã®è²æãè¿
éãªäŸ¡å€ææ¡ãæè¡éçºæè³ãªã©ã®é¢ã§åé¡ãåŒãèµ·ããå¯èœæ§ããããŸããã
ããã§ã¯ãHyDEã䜿ã£ãŠãäœäœå®å®ãšã¯ïŒãã«å¯Ÿããä»®ã®åçãçæãããŠã¿ãŸããããçµæã¯ä»¥äžã§ãã
â ã¯ãšãª
äœäœå®å®ãšã¯ïŒ
â HyDEã«ããä»®åç
äœäœå®å®ãšã¯ãæ°è±¡åŠã®çšèªã§ã倧æ°ã®ç¶æ
ã衚ããã®ã§ããããã¯ãå°è¡šè¿ãã®ç©ºæ°ãå·ãããé«åºŠãäžããã«ã€ããŠç©ºæ°ãæãããªãç¶æ
ãæããŸãããã®ç¶æ
ã§ã¯ãå°è¡šè¿ãã®å·ãã空æ°ã¯éããäžæããããšãé£ãããªããŸãããã®çµæã倧æ°ã¯å®å®ããä¹±æµãç©ä¹±é²ã®çºçãæå¶ãããŸãããã®ãããªç¶æ
ã¯ãæŽããæ¥ã«ããèŠãããŸãããŸããäœäœå®å®ã¯å€§æ°æ±æç©è³ªãå°è¡šè¿ãã«æ»çããããç¶æ
ãäœãåºãããã倧æ°è³ªã«ã圱é¿ãäžããŸãã
ä»®åçèªäœã¯LLMã«çæãããããããã®ããã«å¿ ãããæåŸ ãããããªç¯å²ã®åçã§ã¯ãªãå¯èœæ§ããããŸãã
ããã§ãã¯ãã«æ€çŽ¢ãããŠããæå³ãå šãç°ãªãããããããæ€çŽ¢ç²ŸåºŠãèœã¡ãŠããŸãã§ãããã
ãã¡ããHyDEãæå¹ãªå Žé¢ãããããšæããŸãããããããRAGã«ãããŠã¯LLMã®èªç±åºŠã®é«ããã©ãå¶åŸ¡ããŠè¡ãããšããäºãéèŠãªãã€ã³ãã§ãããHyDEã®å Žåã¯ãå éšã§LLMã®çæåŠçãæŽã«å ¥ãã®ã§ãå¶åŸ¡ã®è€éããäžæ®µäžãã£ãŠããŸããšãèšããŸãã
å ¥åã¯ãšãªãšåç §ãã¹ãããã¥ã¡ã³ãã®é¡äŒŒåºŠãäœãã±ãŒã¹ã§ã¯æå¹ãªææ³ãšãªãå¯èœæ§ãããäžæ¹ãä»®åçã®çæãšå¶åŸ¡ãããããé£ãããšããç¹ã¯èªèããŠããå¿ èŠãããããšæããŸãã
5. ã¹ãããããã¯ããã³ãã
ãã¡ãã¯å ¥åã¯ãšãªããã®ãŸãŸå©çšããã®ã§ã¯ãªããäžæ®µæœè±¡åãã質åå 容(ã¹ãããããã¯è³ªå)ãçæããå ¥åã¯ãšãªãšã¹ãããããã¯è³ªåããããã§æ€çŽ¢ãè¡ãææ³ã«ãªããŸãã
ã¹ãããããã¯è³ªåã®ãµã³ãã«ã¯ä»¥äžã§ãã
å
¥åã¯ãšãªïŒçæ³æ°äœã®å§åPã«ã€ããŠã枩床ã2åã«å¢å ããäœç©ã8åã«å¢å ããå Žåãå§åPã¯ã©ãå€åããŸããïŒ
ã¹ãããããã¯è³ªåïŒãã®è³ªåã®èåŸã«ããç©çåŠã®åçã¯äœã§ããïŒ
å
¥åã¯ãšãªïŒãšã¹ãã©ã»ã¬ãªãã«ãã¯1954幎8æãã11æã®éã«ã©ã®åŠæ ¡ã«éã£ãŠãããïŒ
ã¹ãããããã¯è³ªåïŒãšã¹ãã©ã»ã¬ãªãã«ãã®æè²æŽã¯ã©ããªã£ãŠãããïŒ
ã€ãŸããå ¥åã¯ãšãªãäžæ®µæœè±¡åãããäºã«ãã£ãŠãå ¥åã¯ãšãªã«å¯ŸããŠãã³ãã€ã³ãã§è©²åœç®æããªãã±ãŒã¹ã§ããã¹ãããããã¯è³ªåã®æ€çŽ¢çµæãšåæããäºã§ãæšè«ãå¯èœã«ãªãã±ãŒã¹ããããŸãã
以äžãåè«æããã®ã€ã¡ãŒãžã§ãã
åŒçšïŒTake a Step Back: Evoking Reasoning via Abstraction in Large Language Models
ã¹ãããããã¯ããã³ããã¯RAGã«é¢ä¿ãªã䜿ããææ³ã§ã¯ãããŸãããRAGã§äœ¿ãå Žåã¯ã以äžã®ããã«å ¥åã¯ãšãªã®ã³ã³ããã¹ããšã¹ãããããã¯è³ªåã®ã³ã³ããã¹ããåãããŠåçãçæãããŸãã
response_prompt_template = """You are an expert of world knowledge. I am going to ask you a question. Your response should be comprehensive and not contradicted with the following context if they are relevant. Otherwise, ignore them if they are not relevant.
{normal_context}
{step_back_context}
Original Question: {question}
Answer:"""
ãã¡ããäžèŠè¯ããããªææ³ã§ãããHyDEãšåæ§ã«ãLLMã«ããçæãããäžã€å ¥ãäºã«ãªãã®ã§ãå¶åŸ¡ãé£ãããšããåŽé¢ããããŸãã
äŸãã°ä»¥äžã®ãããªã±ãŒã¹ã§ãã
å
¥åã¯ãšãªïŒå€§è°·ç¿å¹³ã¯2010幎ã«äœãããŸãããïŒ
ã¹ãããããã¯è³ªåïŒ2010幎ã«ç¹å®ã®äººç©ãäœããããïŒ
ãã®å Žé¢ã§ã®æåŸ å€ãšããŠã¯ãã倧谷ç¿å¹³ã®çµæŽã¯ïŒããšãããããªè³ªåã®çæã§ãããã倧谷ç¿å¹³èªäœãæœè±¡åããã2010幎ã«ã©ã®ãããªåºæ¥äºãèµ·ãã£ããïŒã«ãã©ãŒã«ã¹ãåœãã£ãŠããŸã£ãŠããŸãã
ãã ããæœè±¡åãšããæå³ã§ã¯ééãã§ã¯ãªãããã®åé¡èšå®ã§é£ããã®ã¯ãæœè±¡åããŠæ¬²ããç®æãšæœè±¡åããŠæ¬²ãããªãç®æãæé»çã«ååšãããšããäºã§ãã
å®éã®ã¹ãããããã¯è³ªåã®äŸã§ã¯ãã±ãŒã¹æ¯ã«Few-shotããã³ãããçšæããŠããã®èŸºãã®æåŸ å€ãå¶åŸ¡ããããšããŠããŸãã
ãŠãŒã¶ãŒã®ã¯ãšãªãããªããã³ãã€ã³ãã§ã察åã®ããã¥ã¡ã³ããšã®æå³çãªç²åºŠãæã£ãŠããªãå Žåã«æ€çŽ¢ç²ŸåºŠãåäžããå¯èœæ§ããããŸããããã¡ããLLMã®çæåŠçãå éšã§ããäžã€å ¥ããšããå¶åŸ¡ã®é£ããã¯èªèããŠããå¿ èŠãããããšæããŸãã
6. RAG Fusion
ãã¡ãã¯å ¥åã¯ãšãªã«é¡äŒŒããã¯ãšãªãããã€ãçæããããããã®ã¯ãšãªã®æ€çŽ¢çµæãçµ±åããäžã§ãäžäœã«æ¥ããã£ã³ã¯ãåçã«å©çšãããšããæ¹æ³ã§ãã
å ¥åã¯ãšãªã ãã§ãªããé¡äŒŒããã¯ãšãªãLLMã«è€æ°çæãããŠæ€çŽ¢ããäºã§ãé¢é£æ§ã®é«ãç®æãå¹ åºãååŸããäºãæåŸ ã§ããŸãã
é¡äŒŒã¯ãšãªã§ãäžäœã«æ¥ããã£ã³ã¯ã¯éèŠãªãã£ã³ã¯ã ãšããä»®å®ã§ãç°¡åã«èšãã°æç¥šå¶ã§ãéèŠãªãã£ã³ã¯ãååŸããæ¹æ³ã«ãªããŸãã
RAG Fusionã®åºæ¬çãªæµãã¯ä»¥äžã§ãã
â RAG Fusionã®æµã
- é¡äŒŒã¯ãšãªã®çæïŒLLMãçšããŠãå ¥åã¯ãšãªã«é¡äŒŒããã¯ãšãªãè€æ°çæ
- æ€çŽ¢ïŒå ¥åã¯ãšãªãšçæãããé¡äŒŒã¯ãšãªããããã§æ€çŽ¢ãå®è¡
- ãªã©ã³ãã³ã°ïŒåã¯ãšãªã®æ€çŽ¢çµæã®ã©ã³ãã³ã°ãçµ±åããäžäœãã£ã³ã¯ãååŸ
- ã¬ã¹ãã³ã¹ïŒãªã©ã³ãã³ã°ã§æœåºãããäžäœã®ã³ã³ããã¹ãæ å ±ãå ã«ãåçãçæ
ãã£ãŠããäºã¯éåžžã«ã·ã³ãã«ã§ã以äžã®LangChainã®èšäºã®å³ãããããããã§ãã
åŒçšïŒQuery Transformations/LangChain
ãã®å³ã§ã¯ãæ€çŽ¢ããã¯ãã«æ€çŽ¢ãšãªã£ãŠããŸãããReciprocal Rank Fusionã¯åçŽã«é äœã§çµ±åãããããç°ãªãæ€çŽ¢ææ³ã®çµæãçµ±åããäºãã§ããŸãã
ã€ãŸãããã¯ãã«æ€çŽ¢ãšBM25ã®ãããªãã€ããªããæ€çŽ¢ã§ãå©çšå¯èœã§ãçæããé¡äŒŒã¯ãšãªããã¯ãã«æ€çŽ¢ãšBM25ã«ããããæããäžã§æçµçã«çµæãçµ±åãããããšãã§ããŸãã
以äžãRAG Fusionã®ãµã³ãã«ã§ãã
â å
¥åã¯ãšãª
äœäœå®å®ãšã¯ïŒ
â çæãããé¡äŒŒã¯ãšãª
- äœäœå®å®ã®å®çŸ©ãšã¯ïŒ
- äœäœå®å®ã®å ·äœçãªäŸ
- äœäœå®å®ã®çè«ãšãã®å¿çš
- äœäœå®å®ã«é¢ããç ç©¶æç®
â æçµçãªåç
äœäœå®å®ãšã¯ããŠãŒã¶ãŒäŒæ¥ãšãã³ããŒäŒæ¥ãåºå®ãããé¢ä¿ã«ããç¶æ
ãæããŸãããã®ç¶æ
ã§ã¯ãäŒæ¥ã¯ITã·ã¹ãã ã®å¹çåãã³ã¹ãåæžãç®æããITãã³ããŒãšç«¶äºãããäžæ¹ã§ããã³ããŒäŒæ¥ã¯åŽåéã«å¯Ÿãã察䟡ãšããŠäœãªã¹ã¯ã®ããžãã¹ã享åããŸãããããããã®é¢ä¿ã¯IT察å¿èœåã®è²æãé»å®³ããITã·ã¹ãã ããã©ãã¯ããã¯ã¹åããå¯èœæ§ããããŸãããŸãããã³ããŒããã¯ã€ã³ã«ããçµå¶ã®ã¢ãžãªãã£ãäœäžãã顧客ãžã®è¿
éãªäŸ¡å€ææ¡ãå°é£ã«ãªãå¯èœæ§ããããŸãã
ãŒãŒãŒãŒãŒãŒãŒãŒãŒãŒãŒãŒãŒãŒãŒ
ä»åã¯åé¡èšå®ãç°¡åãªã®ã§ããŸãæçšåºŠããããã«ãããããããŸããããçæãããé¡äŒŒã¯ãšãªãèŠãŠã¿ããšãå ¥åã¯ãšãªåäœã«æ¯ã¹ãŠãé¢é£ç®æãå¹ åºãããããããããã€ã¡ãŒãžãæ¹§ãã®ã§ã¯ãªãã§ããããã
HyDEãã¹ãããããã¯ããã³ããã«æ¯ã¹ããšãåäžã®çæã§ã¯ãªã倿°æ±ºã®ãããªç¶æ ã«ãªãã®ã§ãé 匷æ§ã®é«ãã®é¢ã§ã䜿ããããææ³ãªã®ã§ã¯ãªãããšæããŸãã
å¹ åºãæ€çŽ¢ããããã®ã®ããã¡ãããµãã¯ãšãªãšåæ§ã«äžŠååã§ããç¹ãã倧ããªå©ç¹ã®äžã€ã«ãªããŸãã
7. ãã«ãã¹ãããã¯ãšãª
ãã¡ãã¯ãµãã¯ãšãªãRAG Fusionã®ãããªäžŠåçãªææ³ãšã¯ç°ãªãã倿®µéã®æšè«ãé çªã«å®æœããŠããæ¹åŒã«ãªããŸãã
ãµãã¯ãšãªãšåæ§ã«ã¯ãšãªãåè§£ããŸããã䞊åã«åŠçããã®ã§ã¯ãªããåã¯ãšãªã®åçãèžãŸããŠã次ã®ã¯ãšãªãçæãããŸãã
ã€ãŸããåã¯ãšãªãšãã®åçã®çµæãèžãŸããŠæ¬¡ã®ã¯ãšãªãçæãããšããäºãç¹°ãè¿ããŠãåçãæ®µéçã«æŽç·Žãããªãããæçµçãªã¢ãŠãããããçæãããšããææ³ã«ãªããŸãã
以äžãæŠç¥å³ã§ããã1ã€ç®ã®è³ªåã«å¯Ÿããåçã®ãY Combinatorããæ¬¡ã®è³ªåã§äœ¿ãããŠããäºãããããŸãã
ããã°ReActããšãŒãžã§ã³ãã®ãããªåããªã®ã§ã粟床ã¯äžããã ãããªãšæãã€ã€ãçŽåã§ããåŠçã§ããªããããã¬ã¹ãã³ã¹ã®åŸ ã¡æéãšããŠèš±å®¹ã§ããããšããäºãè«ç¹ã«ãªãããšæããŸãã
8. ãã£ã³ã¯æ¡åŒµ
ãã¡ãã¯ãç¹ã«å ·äœçãªååãã€ããŠããããã§ã¯ãªãéåžžã«åçŽãªææ³ã§ãããæ€çŽ¢ã§ããããããã£ã³ã¯ã®ååŸã®ãã£ã³ã¯ãã³ã³ããã¹ãã«å«ããŠLLMã«æž¡ããšããæ¹æ³ã«ãªããŸãã
ãšããã®ããæ€çŽ¢ã«ããããããã£ã³ã¯ã®ååŸã«ã¯ããã®å 容ã®åæãšãªãæ å ±ãªã©ãå«ãé¢é£æ§ã®é«ãæ å ±ãããå¯èœæ§ãé«ããããæ€çŽ¢ã«ã¯çŽæ¥ãããããªããã®ã®ãã³ã³ããã¹ããšããŠã¯æž¡ããã»ããåç粟床ãé«ããªãäºãæåŸ ã§ããããã§ãã
ãã£ã³ã¯ãµã€ãºèªäœã倧ãããããšããã®ãäžã€ã§ããããã£ã³ã¯ã倧ããããã°ããã»ã©ããã¯ãã«ãšããŠã¯æ å ±ãèããªãããããã£ã³ã¯ãµã€ãºãããŸã倧ãããããšæ€çŽ¢ç²ŸåºŠãäžãããšããå ±åãåºãŠããŸãã
å®éã«ãã¯ãã«æ€çŽ¢ããã£ãŠã¿ããšãæèãšããŠååŸã®æ å ±ãæž¡ããæ¹ãè¯ãããã ãªãšæãå Žé¢ã¯å°ãªããªãã®ã§ãã·ã³ãã«ãªææ³ã§ã¯ãããã®ã®ããŠãŒã¹ã±ãŒã¹ã«ãã£ãŠã¯ã詊ããŠã¿ããšå¹æçã«æ©èœãããããããŸããã
9. Pandas Query
ããããã¯å°ãæ¯è²ãéããŸãããCSVãªã©ã®ããŒãã«ããŒã¿ã®åŠçã«é¢ããŠã§ãã
å°ããCSVã§ããã°ããã¹ãã®ãŸãŸåŠçããŠãäžæããããŸãããäžå®å€§ããªCSVã®å Žåã¯ãéåžžã®ããã¹ãåŠçãšãã£ã³ã¯åå²ã§ã¯ããŒã¿ããŒãã«ã®æ§é ãèœã¡ãŠããŸãã®ã§ãããŒãã«æ§é ãä¿æãããŸãŸåŠçããäºãéèŠã«ãªããŸãã
ä»åã¯ã以äžã®5åãããªãåŸæ¥å¡ãµã³ãã«ããŒã¿1000è¡ã䜿ã£ãŠãããŸãã
â ãµã³ãã«ããŒã¿ïŒ1000è¡ã®åŸæ¥å¡ããŒã¿ïŒ
ãã¡ãã«å¯ŸããŠãéåžžã®ãã¯ãã«æ€çŽ¢ãšãPandasã§åŠçã§ããPandasQueryã䜿ã£ãŠãããŸãã
ãŸãã¯éåžžã®ãã¯ãã«æ€çŽ¢ã§ãã
| # | ã¯ãšãª | ãã¯ãã«æ€çŽ¢ã§ã®åç |
|---|---|---|
| 1 | åŸæ¥å¡ã¯å šéšã§äœåããŸããïŒ | åŸæ¥å¡ã¯åèšã§57åããŸãã |
| 2 | å è€éœäžããã®åŸæ¥å¡IDãæããŠäžãã | åŸæ¥å¡IDã«é¢ããæ å ±ã¯æäŸãããŠããŸããã®ã§ãå è€éœäžããã®åŸæ¥å¡IDãæäŸããããšã¯ã§ããŸããã |
| 3 | 山壿·³ããã®æå±éšçœ²ã¯ïŒ | 山壿·³ããã®æå±éšçœ²ã«ã€ããŠã®æ å ±ã¯æäŸãããŠããŸããã |
| 4 | ããŒã±ãã£ã³ã°éšãšéçºéšã«ã¯ããããäœåããŸããïŒ | ããŒã±ãã£ã³ã°éšã«ã¯7åãéçºéšã«ã¯13åããŸãã |
| 5 | 2020幎以éã«å ¥ç€Ÿãã人ã¯äœåããŸããïŒ | 2020幎以éã«å ¥ç€Ÿãã人ã¯15åããŸãã |
ã¡ãªã¿ã«ããã¯å šãŠééã£ãŠããŸãã
ããããåèšã®ãããªæŒç®ãå¿ èŠãªåŠçã®å Žåã¯ããã£ã³ã¯åå²ã§åŠçããäºã¯ã§ããªãäžã«ããããæ å ±ã¯æåã®ãã£ã³ã¯ã«ãããªããããæå±éšçœ²ãåŸæ¥å¡IDã®ãããªæ å ±ãäžæãèªèã§ããŠããªãäºãããããŸãã
ç¶ããŠPandasQueryãèŠãŠã¿ãŸããããããããããããã«ãå éšã§çæãããŠããPandasã®åŠçãèšèŒããŸãã
| # | ã¯ãšãª | å éšã§èªåçæãããPandasã®åŠç | PandasQueryã®åç |
|---|---|---|---|
| 1 | åŸæ¥å¡ã¯å šéšã§äœåããŸããïŒ | df['åŸæ¥å¡ID'].count() | 1000 |
| 2 | å è€éœäžããã®åŸæ¥å¡IDãæããŠäžãã | df[df['åå'] == 'å è€éœäž']['åŸæ¥å¡ID'].values[0] | 9 |
| 3 | 山壿·³ããã®æå±éšçœ²ã¯ïŒ | df[df['åå'] == '山壿·³']['æå±éšçœ²'].values[0] | 人äºéš |
| 4 | ããŒã±ãã£ã³ã°éšãšéçºéšã«ã¯ããããäœåããŸããïŒ | df[df['æå±éšçœ²'].isin(['ããŒã±ãã£ã³ã°éš', 'éçºéš'])]['æå±éšçœ²'].value_counts() | éçºéš214, ããŒã±ãã£ã³ã°éš188 |
| 5 | 2020幎以éã«å ¥ç€Ÿãã人ã¯äœåããŸããïŒ | df[df['å ¥ç€Ÿæ¥'].apply(lambda x: int(x.split('/')[0]) >= 2020)].shape[0] | 429 |
ãã¡ãã¯å šãŠæ£è§£ã§ãããããããã³ãŒãã€ã³ã¿ãŒããªã¿ãŒã®ãããªåœ¢ã§ãèªç¶èšèªã®å ¥åã¯ãšãªããå¿ èŠãªå éšã¯ãšãªãèªåã§çæããããŒã¿ãã¬ãŒã ãžã®ã¯ãšãªçµæãã¬ã¹ãã³ã¹ãšããŠè¿ããŠããŸãã
ããŒãã«ã®æ§é ãç¶æãããŸãŸåŠçã§ããããã«ãCSVã®ãµã€ãºã倧ãããªã£ãŠããå®äœã¯ããã°ã©ã åŠçãªã®ã§ãLLMã®æ¶è²»ããŒã¯ã³èªäœã¯éåžžã«å°ãããšããç¹åŸŽããããŸãã
10. TextToSQL
ãã¡ãã¯ã以åå¥ã®èšäºã§ãèšèŒããŸããããèªç¶èšèªã§ã®ããŒã¿ããŒã¹ããã®æ€çŽ¢ã«ãªããŸãã
ããŒã¿ããŒã¹ã®æ¥ç¶æ å ±ã ããæž¡ããŠãããšãèªç¶èšèªã®å ¥åã¯ãšãªãã€ã³ãããã«ãããŒãã«åãã«ã©ã åããSQLãèªåã§çæã»å®è¡ããŠãæçµçãªåçãè¿ããŠãããŸãã
ã€ã³ããããCSVã®å Žåã¯ãPandas Queryã䜿ãã°è¯ãã§ããã瀟å ã®ããŒã¿ããŒã¹ãã€ã³ãããã«ãããå Žåã¯ãã¡ããå©çšããã®ãè¯ãã§ãããã
ããŒãã«ãã«ã©ã ã®åå¥èª¬æãªãã§ããããŒãã«åãã«ã©ã åããèªåã§é¡æšããŠãããã®ã§ãSQLãããããªããŠãŒã¶ãŒåãã®ç¿»èš³ããŒã«ãšããŠæ©èœããŸãã
åºæ¬çã«ã¯æšæºçãªBIããªãã¹ãçšæããã¹ãããšã¯æããŸããããã£ããããŒã¹ã§ãã³ãã€ã³ãã«ããŒã¿ããŒã¹ã®æ å ±ãååŸã§ããããã«ãªããšãããŒã¿ã®å©æŽ»çšãšæ°äž»åãé²ãã®ã§ã¯ãªãã§ããããã
SQLãäžåã§äžæãçæã§ããªãã±ãŒã¹ã¯åœç¶ããã®ã§ãAgentã§åãã®ãè¯ãããšæããŸãã
RAGã®æ¡åŒµææ³ã«é¢ããææ
ãããŸã§è²ã ãšç޹ä»ããŠããŸããããå人çã«ã¯ãã€ããªãããµãŒããšRAG Fusionã®çµã¿åããã®ã³ã¹ããè¯ãããã ãªãšæããŸããã
ãŠãŒã¶ãŒããã®ã€ã³ããããè€åçãªã¯ãšãªã«ãªããã¡ãªã±ãŒã¹ã®å Žåã¯ããµãã¯ãšãªã广çãããããŸããããã©ã¡ãããšãããšãŠãŒã¶ãŒããã®ã¯ãšãªã¯ã·ã³ãã«ãããŠå°ãã±ãŒã¹ã®ã»ããå€ãã®ã§ãRAG Fusionã§æšªã«æ¡åŒµããŠããããšæ©èœããã±ãŒã¹ãå€ãã®ããªãšæããŸãã
Rerankerã§Transformerç³»ã䜿ãã®ã¯ãGPUçã®ç°å¢é¢ã§ã®ããŒãã«ãé«ããããLLMããŒã¹ããŸãã¯æè»œããªãšæããŸãããããŒã¯ã³æ°ãå¢ããã®ãšãã¬ã¹ãã³ã¹ã¯åœç¶é ããªãã®ã§ããã®èŸºããèžãŸããäžã§ã®æ¡çšæ€èšã«ãªãããšã¯æããŸãã
æ°è¡çšåºŠã®ããŒãã«ã§ããã°ãããã¹ããšããŠãã®ãŸãŸèªã¿èŸŒãã§ã倧äžå€«ã§ãããäžå®ã®ãµã€ãºã®ããããŒãã«ããŒã¿ã®å Žåã¯ããŒã¯ã³æ°çã«ãéçãããããŸãããŒãã«ã®æ§é çãªæå³ãèœãšããªãããã«Pandas QueryãTextToSQLã䜿ãã®ããã¹ãã«ãªããšæããŸãã
æ§ã ãªæ¡åŒµææ³ããããã®ã®ãLangChainã®èšäºã«ããThere is no "one-size-fits-all" solution because different problems require different retrieval techniques.ããšã®èšèŒãããããã«ãRAGã®æ¡åŒµææ³ã«ã€ããŠã¯ããšããããããããã£ãŠããã°è¯ããšãããã¹ããã©ã¯ãã£ã¹ã¯ãªãã®ã§ããããŸã§ãã¿ãŒã³ãšããŠææ¡ããŠããã察象ã®ãŠãŒã¹ã±ãŒã¹ã«åãããŠèŠåŠãæ€èšããŠããäºãéèŠã«ãªããŸãã
å®éã®RAGã®é²ãæ¹
ãããŸã§RAGã®æ¡åŒµææ³ã«ã€ããŠãŸãšããŠããŸããããå®éã«RAGãé²ããéã¯ããšã«ããå®è£ ãé 匵ãã°è¯ããšããèš³ã§ã¯ãªãã以äžã®3ã€ãåããŠèããå¿ èŠããããŸãã
â ãŠãŒã¶ãŒã¯ãšãªã«åé¡ããªãã
â¡åç
§ããã¥ã¡ã³ãã«åé¡ããªãã
â¢RAGã®å®è£
ã«åé¡ããªãã
ããããè«ãšããŠãAIããã£ãŠããã®ã¯ã€ã³ããããå ã«ããã¢ãŠãããããžã®å€æåŠçãªã®ã§ãã€ã³ããããæªããã°è¯ãã¢ãŠããããã¯åºããŸããã
ã€ãŸãããŠãŒã¶ãŒããæž¡ãããã¯ãšãªãšåç §ããã¥ã¡ã³ãèªäœã®è³ªãæªããã°ãã©ãã ãAIãè³¢ããŠãè¯ãã¢ãŠããããã¯åºããªããšããäºã§ãã
ããªããã©ãã ãè³¢ãã£ããšããŠããããããããªãæç€ºãšã誰ãã¡ã³ãããŠããªãããã¥ã¢ã«ãæž¡ãããå Žåã«äžæãä»äºãé²ããããªãã®ãšåã話ã§ãã
ã€ãŸãããããã¿ã€ãã§äžæãè¡ããªãã£ãå Žåã¯ãâ ã¯ãšãªã®åé¡ãªã®ããâ¡ããã¥ã¡ã³ãã®åé¡ãªã®ããâ¢RAGã®å®è£ ã®åé¡ãªã®ããèŠæ¥µããªãã察åŠããŠããå¿ èŠããããŸãã
ç°¡åãªãã§ãã¯ãã€ã³ããšããŠã¯ããèªåããã®ã¯ãšãªãšããã¥ã¡ã³ããæž¡ãããŠåçã§ãããïŒããšããäºãèããŠã¿ãã®ãè¯ãã®ã§ã¯ãªãããšæããŸãã
èªåã§ãåçã§ããªãã®ã§ããã°ããããã®ã€ã³ããããæªããšãã話ã«ãªããŸãããèªåãªãåçã§ãããšããäºã§ããã°ãæ€çŽ¢ãšå¿ççæã®åé¡ã®ãããRAGã®å®è£ ãèŠçŽããæ¹ãè¯ãã§ãããã
ã€ã³ãããèªäœãæªãç¶æ ã§RAGã®å®è£ ãã²ãããé 匵ãã®ã¯ã倧éã®ãã€ãºãæ¬ æããŒã¿ãå«ãããã°ããŒã¿ãæž¡ããããšã³ãžãã¢ãããã£ãŒãã©ãŒãã³ã°ã®ãã¥ãŒãã³ã°ã ããã²ãããé 匵ãã®ãšåã話ã§(çµå±ç²ŸåºŠãåºãªãã®ã§)ããŸãèŠçŽãã¹ãã¯äžæµã®ã€ã³ãããã«ãªããŸãã
ãŠãŒã¶ãŒã¯ãšãªã®æ¹åã«ã€ããŠã¯ãèªç±å ¥åããããŠã¿ããšããã®ã1ã€ãã€ã³ãã«ãªãããšæããŸãã
ãšããã®ããèªç±å ¥åã¯å¿ ããããŠãŒã¶ããªãã£ãé«ããªãããã§ãã
äŸãã°ãAIã¢ã·ã¹ã¿ã³ãã«ã仿¥ã¯äœãé£ã¹ããã§ããïŒäœã§ãèšã£ãŠäžããããšèšããããããã仿¥ã¯äœãé£ã¹ããã§ããïŒã€ã¿ãªã¢ã³ã§ããã°ãè¿ãã«æåãªããã®ãåºãããããã«äºçŽãåããŸããäžè¯æçã§ããã°ïœïœãåé£ã§ããã°ïœïœããšããããã«ãäœã§ãèããŠãšãã圢ãããææ¡åã§ãããçš®èããç¯å²ãçµã£ãŠãããã»ããæ¥œãšããå ŽåããããšæããŸãã
ã€ãŸããèªç±å ¥åã§å šãŠããŠãŒã¶ãŒåŽã«èããããã®ã§ã¯ãªããé©åã«èªå°ãããªãã察象ç¯å²ãçµã£ãŠãããšãã圢ã«ããäºã§ããŠãŒã¶ãŒãããã®ã€ã³ããããåãåãAIãæ¥œã«ãªãå¯èœæ§ããããŸãã
èªç±å ¥åã®ã€ã³ãããæ ã ãçšæããŠããæå³ããããããããã«è©³çްã«å ¥åããŠäžããããšãŠãŒã¶ãŒã«ãé¡ãããã®ã¯ç¡çãããã®ã§ããã«ããŠã³ãèãè¿ãçã®ã¢ã¯ã·ã§ã³ãå©çšããªãããAIã«å ¥ãåã®å ¥åã®è³ªãäžããèšèšãéèŠãªèæ ®ãã€ã³ãã«ãªããŸãã
ããã¥ã¡ã³ãã®æ¹åã«ã€ããŠã¯ãã¯ã¬ã³ãžã³ã°ã®èгç¹ã§ãAIã«æ ¡æ£ããããšããäºãäžã€ã®æå¹ãªæã«ãªããŸãã
誀åè±åãéè€ç®æãèšèŒå 容ãäžæçãªç®æããè€æ°ã®ããã¥ã¡ã³ãã§èšèŒå 容ãççŸããŠããç®æãªã©ã®ãã§ãã¯ãå šãŠäººæã§ããã®ã¯é£ããã®ã§ãæ ¡æ£ã®è£å©ããã»ã¹ã§AIã䜿ããã¯ã¬ã³ãžã³ã°ããŠããã€ã³ããã¯ã¹åãããšããã®ãããŸãã¯ããã»ã©æéããããã«å®æœã§ããæœçããšæããŸãã
ããã°ã©ãã³ã°ã«ãããã³ãŒãã£ã³ã°ãã§ãã¯ããŒã«ã®ãããªäœçœ®ä»ãã§ãããã¥ã¡ã³ãã®èªåãã§ãã¯ããã€ãã©ã€ã³åããŠããŸããšå°è¯ããããããŸããã
ãªããªãäžççžã«ã¯ãããªãã§ããã端çã«èšã£ãŠããŸããšãããã¥ã¡ã³ãå šäœãMECEã«ãªã£ãŠããã°ããã¯ãã«æ€çŽ¢ã§åé¡ãªãååŸã§ããã¯ãã§ãåªç§ãªäººãäœãã誰ã«ãšã£ãŠãèªã¿ãããããã¥ã¡ã³ã(æ§æã®ç¶ºéºãªããã¥ã¡ã³ã)ã¯ãAIã«ãšã£ãŠãèªã¿ããããšèšããŸãã
ããã«AIã«ã³ã³ããã¹ãæ å ±ãæž¡ãããšããäºãè«ç¹ã«ãªã£ãŠããäžçæã«ãããŠã¯ããã®ããŒã¹ãšãªããã¬ããžãããžã¡ã³ããããçš®æãéèŠã«ãªããããããŸããã
ãããã«ããŠãããšã³ãžãã¢ã«å®è£ ã ãäžžæãããŠãäžæããããªãã®ã§ããã©ããããã«ããã¯ãªã®ãïŒããšããäºãèŠæ¥µããŠããããžã§ã¯ããšããŠå¯ŸåŠããŠããå¿ èŠããããŸãã
å ¥åå¯èœãªããŒã¯ã³æ°ãé·ããªããšRAGã¯ãããªããŠãè¯ãïŒ
æè¿ãªãªãŒã¹ãããLLMã¢ãã«ã®ç¹åŸŽãšããŠãå ¥åå¯èœãªããŒã¯ã³æ°ãå€§å¹ ã«å¢ãããšããç¹ããããŸãã
gpt4ã¯æå€§çŽ32,000ããŒã¯ã³ã§ãããgpt4-Turboã§ã¯128,000ããŒã¯ã³ãGemini Pro1.5ã§ã¯æå€§çŽ100äžããŒã¯ã³ãŸã§åŠçå¯èœãšãããããªæ å ±ãåºãŠããŸãã
ãã®æ å ±ãèŠããšããå ¥åå¯èœãªããŒã¯ã³ãå¢ãããã ãããããRAGã§ãã¯ãã«æ€çŽ¢ãšããããã«å šéšã³ã³ããã¹ãã«çªã£èŸŒãã°ããããããªãã®ïŒããšããçåãæ¹§ããŠããŸãã
ããã«é¢ããŠã¯ã仿ç¹ã§äœãšãèšãé£ãã®ã§ãããå°ãªããšããã³ã³ããã¹ãé·ãé·ããªãã»ã©ãLLMã®ç²ŸåºŠãäžãããšããäºè±¡ãå ±åãããŠããäºã¯é ã«çœ®ããŠãããã»ããè¯ããšæããŸãã
以äžãè«æããã®æç²ã«ãªããŸããã10äžã100äžããŒã¯ã³ãšããã¬ãã«ããã¯ããã«å°ããã3000ããŒã¯ã³ä»¥äžã§ããã€ã³ãããã®ããŒã¯ã³æ°ã®å¢å ã«äŒŽã粟床ã®äœäžãå ±åãããŠããŸãã
(ããã¹ãã®èµ€ãå¡ãã€ã¶ããé¢é£ããç®æããã以å€ãç¡é¢ä¿ãªç®æã§ãå³ã«è¡ãã»ã©ãç¡é¢ä¿ãªã€ã³ããããæž¡ãããŠããç¶æ )
ãŸãããã¡ãã®MSã®è«æã§ãåæ§ã«ãã€ã³ããããšããŠæž¡ãããã¥ã¡ã³ãæ°ãå€ããªãã»ã©ç²ŸåºŠãäžãããšããäºè±¡ãå ±åãããŠãããå ããŠãéèŠãªããã¥ã¡ã³ããããã³ããã®äžç€ã«äœçœ®ããŠãããšç²ŸåºŠãèœã¡ã(å é ãæåŸãéèŠããåŸåã«ãã)ãšããè峿·±ãäºè±¡ãå ±åãããŠããŸãã
åŒçšïŒLongLLMLingua: Accelerating and Enhancing LLMs in Long Context Scenarios via Prompt Compression
äžèšãèžãŸãããšãã»ãŒç¡é¢ä¿ãªãã®ãå«ãããã¥ã¡ã³ãã倧éã«æž¡ããŠããã®äžããé¢é£åæãèŠã€ããŠåçããŠãšããã®ã¯ãææŸãã§æ¡çšã§ããæ¹åŒã§ã¯ãªãã®ã§ã¯ãªãããšæããŸãã
ãã¯ãã«æ€çŽ¢ã§é¢é£ç®æãæœåºããæŽã«ãªã©ã³ã«ãŒã§å é ããäžŠã³æ¿ãããã®ãã€ã³ãããã«ãããšç²ŸåºŠãé«ããªããšããã®ã¯ãäžèšã®çµæãããçãéã£ãŠããã®ã§ã¯ãªãã§ããããã
ãããããããŒã¯ã³æ°ãå€§å¹ ã«å¢ãããšããã®ã¯ãã¿ã¹ã¯ãšããŠã¯åœç¶ããé£ãããªãã®ã§ã粟床ãšããŠã¯ãã¬ãŒããªãã«ãªããããåŸãŸããã
極端ãªèšãæ¹ããããšã粟床ã床å€èŠããŠè¯ããã°ããããã§ãå ¥åããŒã¯ã³æ°èªäœã¯å¢ãããã¯ããªã®ã§ãå ¥åããŒã¯ã³æ°ãå¢å ããã¢ãã«ãçºè¡šãããå Žåã¯ãåçŽã«ä»¥åããè¯ããªã£ããšããèš³ã§ã¯ãªããç²ŸåºŠã¯æ ä¿ã§ããŠããã®ãïŒãšããèŠç¹ãéèŠã«ãªãããšæããŸãã
æ®æ®µãããã䜿ã£ãŠãã人ã¯ãååŸãããã£ã³ã¯æ°(simirality-top-kã®ãµã€ãº)ãåçŽã«å€§ãããããšããã³ããã®æç€ºãããŸãå®ããªããªã£ããã倧éæç« ã®èŠçŽã§äžç€ããã軜èŠãããŠãããããªäºè±¡ãäœæãããŠããã®ã§ã¯ãªãããšæããŸãã
ããããç¡é¢ä¿ãªãã®ãå šãŠæž¡ããšããã®ã¯ãæ¶è²»ããŒã¯ã³æ°ãåžžã«å€§ãããªããŸããããããã管çã®åéãã€ããªããªããããã©ã¡ãããšãããšãã¡ã€ã³åäœã§ã€ã³ããã¯ã¹ãåãã察象ããã¥ã¡ã³ãã«å¯ŸããRAGã®æé©åãå«ããŠã¢ãžã¥ãŒã«åããäžæãã«ãŒãã£ã³ã°ãããæ¹æ³ãèããæ¹ãéçšé¢ãå¹çé¢ãèžãŸããŠãè¯ãæ°ãããŸãã
ãã¡ããã³ã³ããã¹ãé·ãé·ããªããšç²ŸåºŠãäœäžããäºè±¡ã«ã€ããŠã¯ãLLMã®éçºè éãåœç¶èªèããŠããã¯ããªã®ã§ãåŸã ã«æ¹åããŠããã®ã§ã¯ãªãããšæããŸãã
ãã ãä»ã®ãšããããšã«ããæå€§é·ã«åãŸããªãäœã§ãçªã£èŸŒãã§è¯ããšãã話ã§ã¯ãªãäºã¯æèããŠããå¿ èŠãããããšæããŸãã
ãŸãšã
ãããã ã£ãã§ããããã
ä»åã¯RAGã®æ¡åŒµæ¹åŒã®ç޹ä»ãäžå¿ã«ãããã«ããã¯ãšããŠæ€èšãã¹ããã€ã³ããã³ã³ããã¹ãé·ã®å¢å ã«äŒŽãRAGã®èãæ¹ã«ã€ããŠç°¡åã«èšèŒããŸããã
æ¡åŒµææ³ã¯è²ã ãšãããŸãããããããéã®åŒŸäžžã¯ãªããŠãŒã¹ã±ãŒã¹ã«å€§ããäŸåãããããä»å玹ä»ããææ³ããã以å€ã®ææ³ãå«ããŠåŒãåºããšããŠæã£ãŠãããããã«PDCAãéãåããããéèŠã«ãªãã®ã§ã¯ãªãããšæããŸãã
詊è¡é¯èª€ãéããªããæ¢ã«åã«é²ãã§ããäŒæ¥ãšãæœè±¡åºŠã®é«ãæºäžããŒã¹ã®è°è«ã«çãŸã£ãŠããäŒæ¥ã«å°ããã€å·®ãåºãŠããŠããå°è±¡ã®ãããRAGå®è£ ãå«ããçæAIã®å©æŽ»çšãé²ããäžå©ã«ãªãã°å¹žãã§ãã



