[{"data":1,"prerenderedAt":524},["ShallowReactive",2],{"project-41048194":3},{"id":4,"name":5,"line":6,"line_source":7,"title":8,"url":9,"domain":10,"maker":11,"posted_at":12,"year":13,"points":14,"comments":15,"rank_in_year":16,"shown_in_year":17,"top":18,"tier":19,"category":20,"subcategory":21,"format":22,"keywords":23,"reception":28,"sentiment":29,"stance_mix":30,"scored_comments":35,"aspects":36,"strip":46,"post_text":49,"what":50,"said":77,"maker_in_thread":199,"shape":208,"grid":212,"attempts":487,"same_week":488,"maker_shows":39,"maker_hidden":516,"reader_version":517,"card":518},"41048194","Zerox","Document OCR with GPT-mini","title","Show HN: Zerox – Document OCR with GPT-mini","https:\u002F\u002Fgithub.com\u002Fgetomni-ai\u002Fzerox","github.com","themanmaran",1721753393,2024,246,99,298,3515,9,"notable","ai","document ocr","library",[24,25,26,27],"ocr","gpt-4o-mini","pdf to markdown","vision model","curious",-0.092,{"positive":31,"neutral":32,"mixed":33,"critical":34},8,27,1,17,53,{"pricing":37,"performance":38,"design":40,"execution":42,"idea":44},{"positive":33,"critical":33},{"positive":33,"critical":39},6,{"positive":41,"critical":33},0,{"positive":33,"critical":43},2,{"positive":45,"critical":33},3,[45,45,45,33,45,41,33,41,45,33,45,41,33,41,41,41,47,41,33,33,33,48,41,45,33,47,33,33,45,48,45,45,33,43,33,48,33,45,41,48],4,5,"This started out as a weekend hack with gpt-4-mini, using the very basic strategy of \"just ask the ai to ocr the document\".\n\nBut this turned out to be better performing than our current implementation of Unstructured\u002FTextract. At pretty much the same cost.\n\nI've tested almost every variant of document OCR over the past year, especially trying things like table \u002F chart extraction. I've found the rules based extraction has always been lacking. Documents are meant to be a visual representation after all. With weird layouts, tables, charts, etc. Using a vision model just make sense!\n\nIn general, I",{"problem":51,"audience":57,"pricing":58,"motivation":62,"time_to_build":57,"stack":66,"openness":74,"stage":75,"team_size":76},{"value":52,"source":53,"evidence":54},"Extracting text and tables from documents with complex layouts","maker",[55],{"id":56},"post",null,{"value":59,"source":53,"evidence":60},"Roughly same cost as current Unstructured\u002FTextract implementation",[61],{"id":56},{"value":63,"source":53,"evidence":64},"Existing rules-based OCR lacked table and chart extraction",[65],{"id":56},[67],{"value":68,"source":53,"evidence":69},"GPT-4o-mini vision model",[70,71],{"id":56},{"id":72,"by":11,"text":73},"41051319","Interesting. It didn't seem like gpt-4o-mini was priced the same as gpt-4o during our testing. We're relying on OpenAI usage page of course, which doesn't give as much request by…","open_source","beta","team",{"praise":78,"objections":91,"requests":128,"questions":154,"alternatives":166},[79],{"value":80,"source":81,"evidence":82},"Vision-model approach handles layouts and tables well","commenter",[83,87],{"id":84,"by":85,"text":86},"41050233","hugodutka","I used this approach extensively over the past couple of months with GPT-4 and GPT-4o while building https:\u002F\u002Fhotseatai.com. Two things that helped me: 1. Prompt with examples. I…",{"id":88,"by":89,"text":90},"41053168","lootsauce","In my own experiments I have had major failures where much of the text is fabricated by the LLM to the point where I just find it hard to trust even with great prompt…",[92,105,117],{"value":93,"source":81,"evidence":94},"Name too close to Xerox trademark",[95,99,102],{"id":96,"by":97,"text":98},"41049569","8organicbits","I'm surprised by the name choice, there's a large company with an almost identical name that has products that do this. May be worth changing it sooner rather than later.…",{"id":100,"by":97,"text":101},"41051364","Even if they don't sue, do you really want to deal with people getting confused and thinking you mean one of the many pre-existing OCR tools that Xerox produces? A search for…",{"id":103,"by":97,"text":104},"41062185","> Are we really suggesting this issue is worth defending and spending resources on? Absolutely. Sure, sometimes non-competing products have the same name. Or products sold…",{"value":106,"source":81,"evidence":107},"LLM OCR can hallucinate or omit text",[108,112,116],{"id":109,"by":110,"text":111},"41051901","tensor","Traditional OCR neural networks like tesseract crucially they have strong measures of their accuracy levels, including when they employ dictionaries or the like to help with…",{"id":113,"by":114,"text":115},"41050720","surfingdino","It's all fun and games until you need to prove something in court or to the tax office. I don't think that throwing an LLM into this mix helps.",{"id":88,"by":89,"text":90},{"value":118,"source":81,"evidence":119},"OpenAI image pricing for mini is high",[120,124],{"id":121,"by":122,"text":123},"41051038","serjester","It should be noted for some reason OpenAI prices GPT-4o-mini image requests at the same price as GPT-4o. I have a similar library but we found OpenAI has subtle OCR…",{"id":125,"by":126,"text":127},"41050973","Dkuku","Check gpt-4o, gpt-4o-mini uses around 20 times more tokens for the same image: https:\u002F\u002Fyoutu.be\u002FZWxBHTgIRa0?si=yjPB1FArs2DS_Rc9&t=655",[129,140,147],{"value":130,"source":81,"evidence":131},"Support for local models",[132,136],{"id":133,"by":134,"text":135},"41050161","ravetcofx","I'd be more curious to see the performance over local models like LLaVa etc.",{"id":137,"by":138,"text":139},"41051401","daft_pink","I would really love something like this that could be run locally.",{"value":141,"source":81,"evidence":142},"Batch API support to halve costs",[143],{"id":144,"by":145,"text":146},"41049507","beklein","Very interesting project, thank you for sharing. Are you supporting the Batch API from OpenAI? This would lower costs by 50%. Many OCR tasks are not time-sensitive, so this might…",{"value":148,"source":81,"evidence":149},"Better handling of multi-column magazine layouts",[150],{"id":151,"by":152,"text":153},"41057203","samuell","The problem I've not found one OCR solution to handle well is complex column based layouts in magazines. Perhaps one problem is that there are often images spanning anything from…",[155,162],{"value":156,"source":81,"evidence":157},"Does it produce a confidence score?",[158],{"id":159,"by":160,"text":161},"41049662","downrightmike","Does it also produce a confidence number?",{"value":163,"source":81,"evidence":164},"How does it compare with Gemini Flash pricing?",[165],{"id":121,"by":122,"text":123},[167,178,185,192],{"value":168,"source":81,"evidence":169},"Tesseract",[170,174],{"id":171,"by":172,"text":173},"41053840","ipkstef","I think i'm missing something.. why would i pay to ocr the images when i can do it locally for free? Tesseract runs pretty well on just cpu, wouldn't even need something crazy…",{"id":175,"by":176,"text":177},"41052553","bearjaws","I did this for images using Tesseract for OCR + Ollama for AI. Check it out, https:\u002F\u002Fcluttr.ai Runs entirely in browser, using OPFS + WASM.",{"value":179,"source":81,"evidence":180},"AWS Textract",[181],{"id":182,"by":183,"text":184},"41055217","aman2k4","I am using AWS Textract + LLM (OpenAI\u002FClaude) to read grocery receipts for \u003Chttps:\u002F\u002Fwww.5outapp.com> So far, I have collected over 500 receipts from around 10 countries with 30…",{"value":186,"source":81,"evidence":187},"Azure Document AI",[188],{"id":189,"by":190,"text":191},"41056501","jerrygenser","Azure document AI accuracy I would categorize as high not \"mid\". Including hand writing. However for the $1.5\u002F1000 pages, it doesn't include layout detection. The $10\u002F1000 pages…",{"value":193,"source":81,"evidence":194},"LLMWhisperer",[195],{"id":196,"by":197,"text":198},"41055446","constantinum","If you want to do document OCR\u002FPDF text extraction with decent accuracy without using an LLM, do give LLMWhisperer[1] a try. Try with any PDF document in the playground -…",{"comments":31,"answered":39,"key":200},[201,202,205],{"id":72,"text":73},{"id":203,"text":204},"41052231","ZerOCR maybe!",{"id":206,"text":207},"41050762","Yup definitely a play on the name. Also the idea of photocopying a page, since we do pdf => image => markdown. We're not planning to name a company after it or anything, just the…",{"commenters":209,"max_depth":39,"maker_replies":31,"first_comment_mins":210,"lifespan_hours":211},69,33,230,[213,215,216,218,221,225,229,231,234,236,240,242,245,249,252,253,254,257,260,263,267,271,274,277,280,283,286,289,292,293,295,296,299,301,304,307,309,312,314,316,319,322,324,326,328,331,335,338,339,342,343,345,348,352,354,355,358,362,364,367,369,373,376,381,385,387,390,394,396,401,405,406,411,414,417,420,423,426,428,430,433,435,437,439,442,445,448,451,454,458,461,463,466,471,474,477,479,481,485],[121,41,45,214,122,41],"project\u002Fneutral\u002Fpricing",[72,41,41,214,11,33],[217,41,45,214,122,41],"41051399",[219,41,45,214,220,41],"41051624","elvennn",[222,41,41,223,224,41],"41054472","other\u002Fneutral\u002Fnone","yfontana",[226,41,45,227,228,41],"41055044","project\u002Fneutral\u002Fperformance","raffraffraff",[230,41,41,223,183,41],"41057345",[232,41,41,223,233,41],"41059075","ianhawes",[96,33,33,235,97,41],"project\u002Fcritical\u002Fnone",[237,33,45,238,239,41],"41049703","project\u002Fneutral\u002Fnone","pkaye",[203,33,41,241,11,33],"maker\u002Fneutral\u002Fnone",[243,33,45,238,244,41],"41049919","froh",[246,33,41,247,248,41],"41050498","commenter\u002Fneutral\u002Fnone","ot",[250,33,41,247,251,41],"41050542","UncleOxidant",[100,33,33,235,97,41],[206,33,41,241,11,33],[255,33,33,235,256,41],"41051283","wewtyflakes",[258,33,41,223,259,41],"41053104","qingcharles",[261,33,41,223,262,41],"41066250","austinjp",[264,33,48,265,266,41],"41051944","maker\u002Fpositive\u002Fnone","ned_at_codomain",[268,33,33,269,270,41],"41055203","maker\u002Fcritical\u002Fnone","ssl-3",[272,33,41,223,273,41],"41056118","neilv",[275,33,33,269,276,41],"41056660","haswell",[278,33,45,238,279,41],"41051314","blacksmith_tb",[281,33,45,241,282,41],"41051333","kevin_thibedeau",[284,33,41,223,285,41],"41051932","actionfromafar",[287,33,41,247,288,41],"41057550","627467",[290,33,41,247,291,41],"41057851","HumblyTossed",[103,33,41,247,97,41],[294,33,33,269,291,41],"41057806",[84,43,45,238,85,41],[297,43,41,247,298,41],"41050296","sidmitra",[300,43,41,247,85,41],"41050385",[302,43,41,303,11,33],"41051781","project\u002Fneutral\u002Fdesign",[305,43,41,247,306,41],"41062978","nbbaier",[308,43,41,247,85,41],"41081043",[310,43,41,223,311,41],"41093512","jimmySixDOF",[189,45,41,313,190,41],"other\u002Fneutral\u002Fperformance",[315,45,41,223,233,41],"41059187",[317,47,45,238,318,41],"41058004","ndr_",[320,47,45,238,321,41],"41058248","markous",[144,48,48,323,145,41],"project\u002Fpositive\u002Fpricing",[325,48,41,214,11,33],"41050707",[327,39,41,223,114,41],"41050160",[329,39,41,223,330,41],"41050225","merb",[332,39,33,333,334,41],"41050301","project\u002Fcritical\u002Fperformance","yjftsjthsd-h",[336,39,41,247,337,41],"41050380","mlyle",[113,39,33,333,114,41],[340,39,41,223,341,41],"41051115","wmf",[109,39,33,333,110,41],[344,39,41,247,337,41],"41071129",[346,39,48,347,259,41],"41053121","project\u002Fpositive\u002Fperformance",[349,39,41,350,351,41],"41055602","other\u002Fneutral\u002Fdesign","ctm92",[175,353,41,223,176,41],7,[196,31,41,223,197,41],[356,18,41,223,357,41],"41058638","binalpatel",[359,360,45,303,361,41],"41055027",10,"amluto",[182,363,41,223,183,41],11,[365,363,41,350,366,41],"41055576","sleno",[368,363,41,350,183,41],"41057315",[370,363,33,371,372,41],"41108072","project\u002Fcritical\u002Fdesign","sumedh",[374,363,45,375,183,41],"41136600","maker\u002Fneutral\u002Fdesign",[377,378,48,379,380,41],"41049511",12,"project\u002Fpositive\u002Fexecution","refulgentis",[382,378,45,383,384,41],"41051927","project\u002Fneutral\u002Fexecution","davedunkin",[386,378,45,383,380,41],"41059197",[88,388,33,389,89,41],13,"project\u002Fcritical\u002Fexecution",[391,392,45,303,393,41],"41052957",14,"jimmyechan",[395,392,41,375,11,33],"41054072",[397,398,33,399,400,41],"41050392",15,"project\u002Fcritical\u002Fidea","josefritzishere",[402,403,41,223,404,41],"41084877",16,"ReD_CoDE",[151,34,45,383,152,41],[407,408,48,409,410,41],"41054640",18,"project\u002Fpositive\u002Fidea","jagermo",[412,413,45,383,97,41],"41049653",19,[415,413,41,416,11,33],"41054135","maker\u002Fneutral\u002Fexecution",[418,413,41,419,97,41],"41057234","commenter\u002Fneutral\u002Fexecution",[421,413,45,383,422,41],"41057391","raisedbyninjas",[424,413,45,383,425,41],"41055450","throwthrowuknow",[159,427,45,227,160,41],20,[429,427,41,247,134,41],"41050137",[431,427,33,333,432,41],"41050219","wildzzz",[434,427,45,227,110,41],"41051915",[436,427,45,227,318,41],"41058049",[125,438,33,333,126,41],21,[133,440,45,441,134,41],22,"project\u002Fneutral\u002Fidea",[171,443,33,444,172,41],23,"project\u002Fcritical\u002Fpricing",[446,443,33,389,447,41],"41054461","gregolo",[449,443,41,419,450,41],"41055066","s5ma6n",[452,443,45,441,453,41],"41057669","daemonologist",[455,456,48,409,457,41],"41048570",24,"cmpaul",[459,456,33,333,460,41],"41049750","hpen",[462,456,41,247,460,41],"41057732",[464,465,45,383,425,41],"41055461",25,[467,468,45,469,470,41],"41055108",26,"project\u002Fmixed\u002Fexecution","jdthedisciple",[472,32,45,441,473,41],"41050917","fudged71",[475,32,41,476,11,33],"41051816","maker\u002Fneutral\u002Fidea",[478,32,41,223,453,41],"41057402",[137,480,48,409,138,41],28,[482,483,48,265,484,41],"41077372",29,"murmansk",[486,483,41,247,484,41],"41077379",[],[489,494,499,503,507,511],{"id":490,"name":491,"points":492,"category":493},"41007403","Sendune",429,"web_saas",{"id":495,"name":496,"points":497,"category":498},"41045834","Briefer",265,"data",{"id":500,"name":501,"points":502,"category":493},"41031975","billabear.com",260,{"id":504,"name":505,"points":506,"category":493},"41048196","Overlap",229,{"id":508,"name":509,"points":510,"category":20},"40998497","Mutable.ai Auto Wiki",162,{"id":512,"name":513,"points":514,"category":515},"41024964","Moocable",161,"consumer_social",false,"v2",{"id":519,"name":5,"line":6,"ai":516,"cat":20,"year":13,"points":14,"comments":15,"rec":28,"s":520,"top":18,"kw":521,"maker":11,"domain":10,"tier":19,"attempts":41,"strip":522,"url":9,"mix":523,"praise":80,"objection":93,"sub":21},41048194,-0.09,[24,25,26,27],"3331301031301000401115031411353312151305",[31,32,33,34],1793991096428]