Ukucaciswa I-Google yasungula i-Gemini File Search, kunye neengcali zithi ukuba i-homebrew RAG (i-Retrieval Augmented Generation). Umzekelo yinto ukuba ngoku umenzi we-app ayidinga ukutya, ukuqhuba, ukugcina ifayile, i-vector database, i-metadata, ukufikelela kwe-optimization, ukuphathwa kwe-context, kunye nokunye. Kwaye yonke i-document Q&A stack (eyaziwa ngokuba yi-middleware kunye ne-application layer logic) ngoku i-Gemini model kunye neengcango zayo ze-peripheral. Kule nqaku, siza kubandakanya i-Gemini File Search kunye nokuguqulwa kwinkqubo ye-homebrew RAG kumgangatho, ukusebenza, ixabiso, ukuxhaswa, kunye ne-transparency. Uya kukwazi ukwenza ukhetho olufanelekileyo kwimeko yakho yokusebenzisa. Kwaye ukwandisa ukuthuthukiswa yakho, ndiya kuhlanganisa . I-app yam isibonelo kwi-GitHub I-app yam isibonelo kwi-GitHub Ndiyathanda i-original : Ukuhlaziywa kweGoogle Ukuhlaziywa kweGoogle Yenza i-Agentic RAG yakho I-Traditional RAG - I-A Refresher I-Architecture ye-RAG ye-traditional ibonakala ngoko, ebandakanya iinyathelo eziliqela eziliqela. Iingxelo zokuqala zihlanganiswa, zihlanganiswa, kwaye zithunyelwe kwi-vector database. Ngokuvamile, i-metadata ezinxulumene ziquka kwi-database entries. Umxholo wama-username uye lithunyelwe kwi-vector DB search ukufumana iinkcukacha ezifanelekileyo. Kwaye lokugqibela, isibuyekezo yokuqala lomsebenzisi kunye neengxaki ezitholise (njenge-context) zithunyelwa kwiimodeli ze-AI ukuvelisa impendulo lomsebenzisi. I-Agent RAG I-Architecture ye-Agentic RAG system ifumaneka i-reflection & react loop, apho i-agent uyahlola ukuba iziphumo ziquka kunye nokugqibeleleyo, kwaye uqhagamshelane isibuyekezo ukufumana umgangatho sokufunda. Ngoko ke, i-AI model isetyenziselwa kwiindawo ezininzi: ukutshintsha isibuyekezo se-user kwi-vector DB isibuyekezo, ukulawula ukuba isibuyekezo se-satisfactory, kwaye ekugqibeleni ukuvelisa impendulo ye-user. Umzekelo wokusetyenziswa - I-Camera Manual Q&A Kukho abaninzi abaphotographers abatsha abavela ukusetyenziswa kwamakhemikhali ezidlulileyo. Enye yeengxaki ezininzi kubo ukuba amaqemikhali ezininzi ezidlulileyo ziyafumaneka iindlela ezizodwa kwaye ezininzi ezizodwa zokusebenza, ngexesha izinto ezincinane, ezifana ukulayisha ifilimu kunye nokuguqula i-film frame counter. Ukungqongileyo, ungenza ukunciphisa i-camera ukuba ungenza izinto ezithile kwi "ukugqiba olungileyo." Ngoko ke, imiyalelo efanelekileyo kwaye esifanelekileyo kwi-camera manual kubalulekile. I-camera manual archive inikeza iidilesi ze-9000 ze-camera ezidlulileyo, ezininzi ze-PDF ezidlulileyo. Kwi-world epheleleyo, uya kufumana iifomati ezincinane ze-camera yakho, ukufundisa, ukufumana, kwaye uye kwenziwe ngoko. Kodwa thina bonke abantu abadala abaxhasiwe okanye abaphambili. Ngoko ke, kufuneka i-Q&A kunye ne-camera manual PDF ngexesha, njl, kwi-app ye-telephone. Kwaye ndingathanda ukuba iya kuthathwa ngokubanzi kwiintlobo ezininzi (i-instruments yokuzonwabisa, izixhobo ze-Hi-Fi, iinqwelo ze-vintage) ezihlangene ukufumana ulwazi evela kumadokhumenti ze-ancient. I-Homebrew RAG ye-PDF Q&A inkqubo yethu RAG yasungulwa ngexesha elidlulileyo ngexesha elidlulileyo nge customization enkulu: I-LlaMAIndex RAG Workflow I-LlaMAIndex RAG Workflow Ukusebenzisa Qrrant database vector: ixabiso-umgangatho elungileyo, ukunceda metadata. Ukusetyenziswa kwe-Mistral OCR API ukufumana i-PDF: ukusebenza olungcono ekutholeni iifayile ze-PDF ezininzi kunye neentlawulo kunye neentlawulo. Ukugcina iifoto zeephepha ze-PDF ukuze abasebenzisi bangakwazi ukufikelela ngqo iifoto zeengxaki ze-camera kunye neengxaki ze-text. Ukongeza i-agentic loop ye-reflection kunye ne-react ezisekelwe kwi-Google / i-Langchain isibonelo ye-agentic search. I-Google / i-Langchain isibonelo ye-agentic search Yintoni malunga Multi-Modal LLMs? Ukususela ngo-2024, i-LLM ye-multimodal iye yenza kakhulu kakhulu. Umgangatho olungagqibeleleyo yokufaka i-user query kunye ne-PDF epheleleyo kwi-LLM kunye nokufumana isibuyekezo. Kuyinto isisombululo elula kakhulu enokufuneka ukuba akufuneka i-DB ye-vector okanye i-middleware. inkcazelo yethu yokuqala yaba ixabiso, ngoko siye siyenze i-cost calculation and comparison. Yaye ingxelo oluthe ngempumelelo ukuba i-RAG iye ngokukhawuleza, i-efficient, kwaye kunzima kakhulu ngexabiso xa inani lwezilwanyana lwabasebenzisi ngosuku kunokuba ngaphezu kwe-10. Ngoko ke, "ukutya ngokuthe ngqo isibuyekezo lwabasebenzisi kunye nqakraza yonke i-PDF kwi-Multi-modal LLM" iyasebenza kuphela ngokwenene ne-prototyping okanye ukusetyenziswa kwe-volume kakhulu (izibuyekezo ezincinane ngosuku). Kwaye ngexesha, kuqinisekisa ukuba i-homebrew RAG iye yintoni ebalulekileyo xa i-Google ibheka i-Gemini File Search. Ndingathanda ukuba umxholo akuyona elula kakhulu. I-Gemini File Search - Umzekelo Ndava i-app isampuli ye-camera manual Q&A usage case, eyakhelwe yi-Google AI Studio isampuli. ukuze unako ukuyifaka ngokukhawuleza kakhulu. Nazi i-screenshot ye-user interface kunye ne-chat thread. , I-open source kwi-GitHub I-open source kwi-GitHub Umzekelo we-Q&A kunye ne-PDF usebenzisa i-Gemini File Search: https://github.com/zbruceli/pdf_qa https://github.com/zbruceli/pdf_qa Iimpawu eziphambili ezihambelana ne-source code: Yenza i-File Search Store, kwaye uqhagamshelane kwiiyure ezahlukeneyo. Ukukhuphela iifayile ezininzi ngexesha elifanayo, kwaye i-Google backend iya kuthatha zonke i-chunking kunye ne-embedding. Kwakhona ivela iingxaki ze-sampling kubasebenzisi. Ukongezelela, unako ukuguqula i-chunking kunye ne-upload ye-metadata ye-custom. Ukuqhuba i-Standard Generation Query (RAG): Phantsi komhlaba, i-agentic kwaye inokufunda umgangatho weemiphumo ngaphambi kokufumana umxholo wokuqala. Zonke iinkcukacha ze-developer I-GEMINI File Search API i-doc https://ai.google.dev/gemini-api/docs/file-search https://ai.google.dev/gemini-api/docs/file-search Umhlahlandlela nguPhill Schmidt https://www.philschmid.de/gemini-file-search-javascript https://www.philschmid.de/gemini-file-search-javascript Xabiso kwi-Gemini File Search I-Developers ibhaliswe kwi-embeddings ngexesha lokubhalisa ngokutsho kwimali ye-embeddings ezikhoyo (i-$0.15 ngexesha le-1M ye-token). I-Storage iyatholakala khulula. I-Query time embeddings ifumaneka khulula. I-Document Tokens ebonakalayo ifakwe njenge-context tokens ezivamile. Iindleko ze-embeddings I-Context Token Ngoko ke, yintoni engcono? Ngenxa yokuba i-Gemini File Search iyiphi na ingxaki ezintsha, ukubuyekezwa kwimibelelwano yam ngexesha elandelayo ngexesha elidlulileyo. Ukubala umthamo Gemini File Search unayo zonke iimpawu eziphambili ye-homebrew RAG system Chunking (ngaba ukuguqulwa ubungakanani kunye nokufaka) Ukucinga I-Vector DB ekusebenziseni i-custom metadata input Ukucinga Imveliso Generative Kwakhona iimpawu ezininzi ezidlulileyo phantsi kwe-hood: Umgangatho we-agent to assess the quality of retrieval Ukuba kufuneka i-nitpick, i-imaging output ingaphantsi ngoku. Ngoku, i-output ye-Google File Search ibhekwa kuphela kwi-text, kwaye i-RAG eyenziwe ngexabiso ingaba ufumana iifoto kwi-PDF eyenziwe. Ndingathanda ukuba ayikho ingxaki kakhulu ukuba i-Gemini File Search inikeza i-multimodal output ngexesha elandelayo. Ukubala ukusebenza Ukucaciswa: ngexabiso. Akukho ukucaciswa okucacileyo kwi-recovery okanye umgangatho we-generation. I-Gemini File Search ingaba i-slightly faster, njengoko i-vector DB kunye ne-LLM zombini "ngathi" kwi-infrastructure ye-Google Cloud. Iindleko Ukubala Okugqibela, i-Gemini File Search yinkqubo esebenzayo ngokupheleleyo enokwenzeka Ngaphandle kwe-homebrew system. less Ukubunjwa kwedokumenthi kwenziwa nje ngexesha elinye, kwaye iindleko $0.15 ngalinye tokens. Oku kuyinto ixabiso olufanelekileyo ukuba zonke iinkqubo RAG, kwaye kungenziwa i-amortised ngexesha lokuzalwa kwedokument Q & A isicelo. Kwimeko yam ukusetyenziswa kwamakhasimende camera, le ixabiso olufanelekileyo kuyinto ingxenye elincinci yexabiso jikelele. Ngenxa yokuba i-Gemini File Search inikeza isitoreji yefayela kunye ne-database "umhla", oku kubuyekeza kwinkqubo ye-homebrew RAG. Iindleko ye-inference iyona efanayo, njengoko inani le-token ye-input (i-question plus i-vector search results njenge-context) kunye ne-output token zihlanganisa phakathi kwe-Gemini File Search kunye ne-homebrew system. I-Flexibility kunye ne-Transparency ye-Tuning ne-Debugging Ngokwemvelo, i-Gemini File Search ibamba kubandakanya iimodeli ze-Gemini AI ngenxa yokubandakanya kunye nokugqiba. Kukho ngokwemvelo ukufumana ukufumanisa ngexesha lokugqiba kunye nokukhetha. Ngokutsho ukucaciswa kwinkqubo yakho ye-RAG, i-Gemini File Search inikeza inqanaba elifanelekileyo. Umzekelo, unako ukucacisa i-chunkingConfig ngexesha lokukhuthaza ukucacisa iiparametre efana ne-maxTokensPerChunk kunye ne-maxOverlapTokens, kunye ne-customMetadata ukucacisa iipari ze-key-value kwi-document. Nangona kunjalo, kubona engabonakali ukuba unomdla eyinxalenye kwi-Gemini File Search system for debugging and performance tuning. Ngoko ke, unasebenzisa kakhulu okanye ngakumbi njenge-black box. Imibuzo I-Google's Gemini File Search iyona elungileyo kwi-applications ezininzi kunye neempawu ezininzi ngexabiso elihle kakhulu. I-super easy to use and has minimal operational overhead. It is not only good for quick prototyping and mock-ups, but also good enough for a production system with thousands of users. Nangona kunjalo, kukho izicwangciso ezimbalwa ukuba ungenza i-homebrew RAG system: Ngaba awufunayo i-Google yokufakelwa iidokhumenti akho ze-proprietary. Ufuna ukuguqulwa iifoto kubasebenzisi evela kumadokhumenti yokuqala. Ingaba ufuna umlinganiselo epheleleyo kunye ne-transparency ngalinye i-LLM ukuba isetyenziswe kunye ne-inference, indlela yokwenza i-chunking, indlela yokulawula i-agent flow ye-RAG, kunye neendlela yokubhalisa iinkcukacha zekhwalithi zokusetyenziswa. Ngoko ke, ukunika i-Gemini File Search ukuyifuneka kwaye uye uye uyifunayo. Ungasebenzisa i njengoko umdlalo, okanye ungenza . Nceda uqhagamshelane phantsi malunga neziphumo zakho zokusetyenziswa kwimeko zakho zokusetyenziswa. I-Google i-AI Studio I-example code yam kwi-GitHub I-Google i-AI Studio I-example code yam kwi-GitHub
