I-Mistral OCR 4 Iletha Okukhiphayo Okuhlelekile Okucashuniwe ku-RAG, i-Agentic, kanye ne-Enterprise Search Pipelines
Namuhla, iMistral AI ikhululiwe OCR 4imodeli yayo yakamuva yokuqonda idokhumenti. Lokhu kukhishwa okusha kungeza amabhokisi abophayo, ukuhlukaniswa kwamabhulokhi, kanye nezikolo zokuzithemba okusemgqeni eceleni kombhalo okhishiwe. Iyasekela Izilimi ezingu-170 emaqenjini ezilimi ezingu-10 futhi isebenza esitsheni esisodwa ukuze kusetshenziswe okuzibambele ngokwakho ngokugcwele. I-OCR 4 iphinda isebenze njengengxenye yokungenisa ekusesheni kwebhizinisi, i-RAG, namapayipi okubuyiswa aqondene nesizinda esithile.
I-TL;DR
- I-OCR 4 ibuyisela amabhokisi ahlanganisayo, amalebula ebhulokhi elithayiphiwe, kanye nezikolo zokuthenjwa zegama ngalinye, hhayi umbhalo kuphela.
- Isekela izilimi ezingu-170 emaqenjini angu-10, kanye nezinzuzo ezilimini eziyivelakancane nezisetshenziswa kancane.
- Izichasiselo ezizimele zincamela i-OCR 4 kunawo wonke amasistimu ahloliwe, isilinganiso samazinga okuwina angama-72%.
- Intengo ingu-$4 emakhasini angu-1,000, yehlela ku-$2 ngesaphulelo se-Batch-API.
- Isiphetho esisodwa sisebenzisa kokubili ukukhipha okungaphekiwe kanye nokuphumayo kweDokhumenti ye-AI eqhutshwa yi-schema.
I-Mistral OCR 4
I-Mistral OCR 4 ikhipha kanye nokuqukethwe kwezakhiwo kusuka enhlobonhlobo yamadokhumenti. Izizukulwane ezedlule bezigxile ekuguquleni ikhasi libe umbhalo namathebula ahlanzekile. I-OCR 4 esikhundleni salokho ibuyisela ukumelwa okuhlelekile kwayo yonke idokhumenti.
Ibhulokhi ngalinye lenziwa lasendaweni ngebhokisi elibophayo futhi lihlukaniswa ngohlobo. Izinhlobo zokuvimba zihlanganisa izihloko, amathebula, izibalo, amasiginesha, nokuningi. Izikolo zokuzithemba okusemgqeni zikhiqizwa ekhasini ngalinye nangegama ngalinye.
Ngakho-ke amasistimu aphansi afunda okungaphezu kwalokho okushiwo idokhumenti. Baphinde bafunde ukuthi i-elementi ngayinye ihlezi kuphi, idlala yiphi indima, nokuthi imodeli iqiniseka kangakanani. Lowo mongo owengeziwe ubalulekile ekucashunisweni, ekuguqulweni kabusha, nokuqinisekiswa komuntu-in-the-loop.
I-OCR 4 yamukela amafomethi ajwayelekile ebhizinisi, afaka i-PDF, i-DOC, i-PPT, ne-OpenDocument. Imodeli ihlangene ngokwanele ukuthi ingafakwa esitsheni esisodwa. Ukuthunyelwa okuzilawulayo kuyatholakala kumakhasimende ebhizinisi ngokuhlala kwedatha nokuhambisana.
Ibhentshimakhi
I-Mistral iqhathanise i-OCR 4 namamodeli we-AI-native OCR, amamodeli enhloso evamile yomngcele, izinsiza zemibhalo yebhizinisi, kanye ne-Mistral OCR 3.
Inani lezichasiselo ezizimele lincamela i-OCR 4 kunawo wonke amasistimu aholayo ahloliwe. Izinga lokuwina lilinganiselwa ku-72% kuyo yonke isethi yokuqhathanisa. Ukuhlola kusebenzise amadokhumenti angu-600+ ngezilimi ezingu-12+, athathwe kubathengisi bezinkampani zangaphandle. Izichasiselo zilinganisela umphumela wezimbangi ngamunye ngokuqhathaniswa nama-OCR 4, idokhumenti ngedokhumenti.
Kumabhentshimakhi azenzakalelayo, i-OCR 4 ikora 85.20 ku-OlmOCRBench yomphakathi. Lishaye igoli 93.07 ku-OmniDocBench futhi .98 ekuhloleni kwangaphakathi kwe-Mistral Ukukhasa Kwezilimi Eziningi.
Amaphuzu amabili edatha yekhasimende engeza umongo. U-Rogo ubike ukunemba okulinganako cishe ngezindleko eziphansi ezingu-8x kanye ne-latency ephansi engu-17x uma kuqhathaniswa nabahlaluli abaholayo be-ejenti. I-Anaqua ikalwe cishe ngokushesha okungu-4x ekhasini ngalinye kunomhlinzeki wayo okhona.
Ukwehlukanisa, Hhayi Umbhalo Kuphela
Amabhokisi okubopha bekuyikhono leMistral elicelwe kakhulu. Benza umbhalo ukuze bagqanyiswe kokuqukethwe kanye namaphayiphi edatha athembekile.
Izinhlobo zokuvimba kanye nezikolo zokuzethemba zinikeza imisebenzi ehlukene. Bashayela izingcaphuno ezisekelwe emthonjeni, ukuguqulwa kabusha, nokuqinisekiswa komuntu-in-the-loop. Lesi sakhiwo sisekela imisebenzi eminingana eya phansi.
Amabhulokhi ahlanzekile, ahlukanisiwe aba amayunithi angcono okubuyisa e-RAG. Ama-ejenti athola ama-primitives esakhiwo ukuze asebenze kumadokhumenti, hhayi nje ukuwafunda. Izixhumi zithola okukhiphayo okungaguquki, okuthayiphiwe ukuze kungeniswe futhi kufakwe ohlwini.
I-OCR 4 iphinde ibe yingxenye yokungenisa I-Mistral Search Toolkitmanje isibonwa kuqala esidlangalaleni. Ikhithi Yamathuluzi Yokusesha ingumthombo ovulekile, uhlaka lokusesha lwe-Mistral oluhlanganisekayo. Okukhiphayo okuhlelekile kunikeza okokufaka okulungele ukucashunwa ekubuyiseni nasekuhloleni ukuhamba komsebenzi.
Sebenzisa Amacala anezibonelo
I-OCR 4 isekela kokubili amapayipi evolumu ephezulu kanye nokugeleza komsebenzi wedokhumenti okusebenzisanayo.
- Ukuhlaziya nokukhipha idokhumenti: Guqula inkontileka yezilimi eziningi ibe ukumaka okuhlanzekile, okuhlelekile ukuze kufakwe ohlwini.
- I-Retrieval-Augmented Generation (RAG): Phakela amabhlogo ahlukanisiwe ku-Search Toolkit ukuze uthole izimpendulo ezisekelwe emthonjeni ngezingcaphuno.
- Ukugeleza komsebenzi we-Agent: Nikeza umenzeli wokucubungula ama-invoyisi izinkambu ezithayiphiwe namabhokisi okubophezela ukuze agcwalise amafomu ngokuzenzakalelayo.
- Amapayipi anesango lokuzethemba: Hambisa izifunda ezinokuzithemba okuphansi kuziqinisekisi zabantu, futhi ugunyaze ngokuzenzakalelayo okunye.
- Ukusesha kwebhizinisi: Sebenzisa i-OCR 4 njengengxenye yomthombo wedatha yokungenisa nokukhipha ibhizinisi kuyo yonke ingobo yomlando.
Abasebenzisi bangaphambi kwesikhathi basebenzisa i-OCR 4 ukuze baguqule ama-invoyisi abe yizinkambu ezihlelekile futhi benze izingobo zomlando zenkampani zifakwe kwidijithali. Abanye bakhipha umbhalo ohlanzekile emibikweni yobuchwepheshe noma usesho lwebhizinisi lamandla.
Inothi kububanzi obuvela ekukhululweni okusemthethweni kwe-Mistral: I-OCR 4 iyimodeli yokuqonda idokhumenti, hhayi umenzi wesinqumo. Akuhloselwe ukuxilongwa kwezokwelapha, ukwahlulela okusemthethweni, noma izinqumo zezezimali eziphakeme. Futhi ayifaneleki kumasistimu abalulekile okuphepha, ukucubungula ngesikhathi sangempela, noma okokufaka okungewona amadokhumenti njengomsindo ongahluziwe noma ividiyo.
I-OCR 4 ihamba ngemuva kwephoyinti lokuphela le-API. Isicelo ngasinye sisebenzisa imodeli efanayo. Ihlala ibuyisela okuqukethwe okukhishiwe, amabhokisi okubophezela, izinhlobo zamabhulokhi, izikolo zokuzethemba, nokumaka. Okuhlukayo ukuthi ubeka kangakanani phezulu.
| Amandla | Imodi ehlanzekile yokukhipha | I-Document AI Mode (isiphetho esifanayo) |
|---|---|---|
| Okukhiphayo | I-Markdown, ama-bboxes, izinhlobo zamabhulokhi, ukuzethemba | I-JSON ehleliwe ku-schema oyichazayo |
| Isebenza kanjani | Impendulo ye-OCR eluhlaza | Okuphumayo kwe-OCR kudliwe mistral-small-2603 |
| Isichasiselo sesithombe | Akusetshenziswanga | Ikholi yolimi lombono ngesithombe ku-schema |
| Ukwaziswa ngokwezifiso | Cha | Yebo, iqondisa ukuhumusha noma isifinyezo |
| Kuhle kakhulu | Amapayipi, ama-ejenti, ukungeniswa kwenqwaba | Abasebenzisi bebhizinisi, abashayeli bezindiza, akukho logic yokuhlaziya |
| Inani | $4 / 1,000 amakhasi ($2 inqwaba) | $5 / 1,000 amakhasi |
| Ukuzibamba ngokwakho | Itholakalela ibhizinisi | Itholakalela ibhizinisi |
Umthetho wesinqumo ulula. Udinga okuqukethwe okungahluziwe okukhishiwe? Sebenzisa i-OCR 4 njengoba injalo. Udinga okukhiphayo okubunjwe kabusha ku-schema noma kuchazwe ngezinkambu zesizinda? Engeza amapharamitha we-Document AI ocingweni olufanayo.
Ukusebenza Nge-API
Ukukhishwa okuyisisekelo kuthatha i-URL yedokhumenti futhi kubuyisele amakhasi ahlelekile. Setha include_blocks=True ukuze uthole amabhlogo athayiphiwe namabhokisi okubopha.
import os
from mistralai.client import Mistral
client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])
ocr_response = client.ocr.process(
model="mistral-ocr-latest",
document={
"type": "document_url",
"document_url": "
},
include_blocks=True, # typed blocks + bounding boxes
table_format="html", # None (inline), "markdown", or "html"
include_image_base64=True
)Impendulo iyinto ye-JSON ene-a pages uhlu. Ikhasi ngalinye liphethe markdown, images, tables, hyperlinks, dimensionsfuthi confidence_scores. Ukuze uthole ipayipi lokubuyekezwa komuntu, cela ukuzethemba kwegama ngalinye.
ocr_response = client.ocr.process(
model="mistral-ocr-latest",
document={"type": "document_url",
"document_url": "},
confidence_scores_granularity="word" # or "page" for aggregates
)I "word" isilungiselelo sengeza a word_confidence_scores amalungu afanayo ngekhasi ngalinye nokufakiwe kwethebula ngakunye. Ngemisebenzi enevolumu ephezulu, i-Mistral incoma isevisi ye-Batch Inference, enciphisa izindleko zekhasi ngalinye.
Izame: I-Interactive Output Explorer
Okushumekiwe ngezansi kukhombisa okukhiphayo okuhleliwe kwe-OCR 4. Shintsha phakathi kwesampula yamadokhumenti, guqula amabhokisi ahlanganisayo nezinhlobo zamabhulokhi, bese uvula imephu yokushisa yokuqiniseka. Amathebhu we-Markdown kanye ne-JSON akhombisa imimo emibili yokuphuma eceleni. Idatha yesampula ingumfanekiso, hhayi ikholi ye-API ebukhoma.
Hlola Isimemezelo se-Mistral OCR 4, ikhadi lemodeli ye-OCR 4, futhi I-OCR Prosesa amadokhumenti. Futhi, zizwe ukhululekile ukusilandela Twitter futhi ungakhohlwa ukujoyina wethu 150k+ML SubReddit futhi Bhalisela ku Iphephandaba lethu. Linda! ukutelegram? manje ungasijoyina kuthelegramu futhi.
Udinga ukusebenzisana nathi ekuthuthukiseni i-GitHub Repo yakho NOMA Ikhasi Lobuso Lokugona NOMA Ukukhishwa Komkhiqizo NOMA I-Webinar njll.? Xhumana nathi
Imithombo: Isimemezelo se-Mistral OCR 4, ikhadi lemodeli ye-OCR 4, Amadokhumenti e-OCR Processor.
