Simon, Interesting commentary. The issue that Joe and I have both looked at, with the splitting of metadata and content extraction, is that if they're split then the underlying Tika extraction has to process the file twice: once to pull out the attributes and once to pull out the content. Perhaps it may be good to add ExtractMetadata and ExtractTextContent in addition to ExtractMediaAttributes - ? Seems kind of an overkill but I may be wrong.
It seems prudent to provide one wholesome, out-of-the-box extractor processor with options to extract just metadata, just content, or both metadata and content. I think what I'm hearing is that we need to allow for checking somewhere for whether text/content has already been extracted by the time we get to the ExtractMediaAttributes processor - ? If that is the issue then I believe the user would use RouteOnAttribute and if the content is already filled in then they'd not route to ExtractMediaAttributes. As far as the OCR. Tika internally supports OCR by directing image files to Tesseract (if Tesseract is installed and configured properly). We've started talking about how this could be reconciled in the ExtractMediaAttributes. I think that once we have the basic ExtractMediaAttributes, we could add filters for what files to enable the OCR on, and we'd need to expose a few config parameters specific to OCR, such as e.g. the location of the Tesseract installation and the maximum file size on which to attempt the OCR. Perhaps there can also be a RunOCR processor which would be dedicated to running OCR. But since Tika already has OCR integrated we'd probably want to take care of that in the ExtractMediaAttributes configuration. Additionally, I've proposed the idea of a ProcessPDF processor which would ascertain whether a PDF is 'text' or 'scanned'. If scanned, we would break it up into pages and run OCR on each page, then aggregate the extracted text. - Dmitry On Thu, Mar 31, 2016 at 3:19 PM, Simon Ball <[email protected]> wrote: > Just a thought… > > To keep consistent with other Nifi Parse patterns, would it make sense to > based the extraction of content on the presence of a relation. So your tika > processor would have an original relation which would have meta data > attached as attributed, and an extracted relation which would have the > metadata and the processed content (text from OCRed image for example). > That way you can just use context.hasConnection(relationship) to determine > whether to enable the tika content processing. > > This seems more idiomatic than a mode flag. > > Simon > > > On 31 Mar 2016, at 19:48, Joe Skora <[email protected]> wrote: > > > > Dmitry, > > > > I think we're good. I was confused because "XXX_METADATA MIMETYPE > FILTER" > > entries referred to some MIME type of the metadata, but you meant to use > > the file's MIME type to select what files have metadata extracted. > > > > Sorry, about that, I think we are on the same page. > > > > Joe > > > > On Thu, Mar 31, 2016 at 11:40 AM, Dmitry Goldenberg < > > [email protected]> wrote: > > > >> Hi Joe, > >> > >> I think if we have the filters in place then there's no need for the > 'mode' > >> enum, as the filters themselves guide the processor in deciding whether > >> metadata and/or content is extracted for a given input file. > >> > >> Agreed on the handling of archives as a separate processor (template, > seems > >> like). > >> > >> I think it's easiest to do both metadata and/or content in one processor > >> since it can tell Tika whether to extract metadata and/or content, in > one > >> pass over the file bytes (as you pointed out). > >> > >> Agreed on the exclusions trumping inclusions; I think that makes sense. > >> > >>>> We will only have a mimetype for the original flow file itself so I'm > >> not sure about the metadata mimetype filter. > >> > >> I'm not sure where there might be an issue here. The metadata MIME type > >> filter tells the processor for which MIME types to perform the metadata > >> extraction. For instance, extract metadata for images and videos, only. > >> This could possibly be coupled with an exclusion filter for content that > >> says, don't try to extract content from images and videos. > >> > >> I think with the six filters we get all the bases covered: > >> > >> 1. include metadata? -- > >> 1. yes -- > >> 1. determine the inclusion of metadata by filename pattern > >> 2. determine the inclusion of metadata by MIME type pattern > >> 2. no -- > >> 1. determine the exclusion of metadata by filename pattern > >> 2. determine the exclusion of metadata by MIME type pattern > >> 2. include content? -- > >> 1. yes -- > >> 1. determine the inclusion of content by filename pattern > >> 2. determine the inclusion of content by MIME type pattern > >> 2. no -- > >> 1. determine the exclusion of content by filename pattern > >> 2. determine the exclusion of content by MIME type pattern > >> > >> Does this work? > >> > >> Thanks, > >> - Dmitry > >> > >> > >> On Thu, Mar 31, 2016 at 9:27 AM, Joe Skora <[email protected]> wrote: > >> > >>> Dmitry, > >>> > >>> Looking at this and your prior email. > >>> > >>> > >>> 1. I can see "extract metadata only" being as popular as "extract > >>> metadata and content". It will all depend on the type of media, for > >>> audio/video files adding the metadata to the flow file is enough but > >> for > >>> Word, PDF, etc. files the content may be wanted as well. > >>> 2. After thinking about it, I agree on an enum for mode. > >>> 3. I think any handling of zips or archive files should be handled by > >>> another processor, that keeps this processor cleaner and improves its > >>> ability for re-use. > >>> 4. I like the addition of exclude filters but I'm not sure about > >> adding > >>> content filters. We will only have a mimetype for the original flow > >>> file > >>> itself so I'm not sure about the metadata mimetype filter. I think > >>> content > >>> filtering may be best left for another downstream processor, but it > >>> might > >>> be run faster if included here since the entire content will be > >> handled > >>> during extraction. If the content filters are implemented, for > >>> performance > >>> they need to short circuit so that if the property is not set or is > >> set > >>> to > >>> ".*" they don't evaluate the regex. > >>> 1. FILENAME_FILTER - selects flow files to process based on filename > >>> matching regex. (exists) > >>> 2. MIMETYPE_FILTER - selects flow files to process based on > >> mimetype > >>> matching regex. (exists) > >>> 3. FILENAME_EXCLUDE - excludes already selected flow files from > >>> processing based on filename matching regex. (new) > >>> 4. MIMETYPE_EXCLUDE - excludes already selected flow files from > >>> processing based on mimetype matching regex. (new) > >>> 5. CONTENT_FILTER (optional) - selects flow files for output based > >> on > >>> extracted content matching regex. (new) > >>> 6. CONTENT_EXCLUDE (optional) - excludes flow files from output > >> based > >>> on extracted content matching regex. (new) > >>> 5. As indicated in the descriptions in #4, I don't think overlapping > >>> filters are an error, instead excludes should take precedence over > >>> includes. Then I can include a domain (like A*) but exclude sub-sets > >>> (like > >>> AXYZ*). > >>> > >>> I'm sure there's something we missed, but I think that covers most of > it. > >>> > >>> Regards, > >>> Joe > >>> > >>> > >>> On Wed, Mar 30, 2016 at 1:56 PM, Dmitry Goldenberg < > >>> [email protected] > >>>> wrote: > >>> > >>>> Joe, > >>>> > >>>> Upon some thinking, I've started wondering whether all the cases can > be > >>>> covered by the following filters: > >>>> > >>>> INCLUDE_CONTENT_FILENAME_FILTER - defines any patterns for which input > >>>> files get their content extracted, by file name > >>>> INCLUDE_METADATA_FILENAME_FILTER - defines any patterns for which > input > >>>> files get their metadata extracted, by file name > >>>> INCLUDE_CONTENT_MIMETYPE_FILTER - defines any patterns for which input > >>>> files get their content extracted, by MIME type > >>>> INCLUDE_METADATA_MIMETYPE_FILTER - defines any patterns for which > input > >>>> files get their metadata extracted, by MIME type > >>>> > >>>> EXCLUDE_CONTENT_FILENAME_FILTER - defines any patterns for which input > >>>> files do NOT get their content extracted, by file name > >>>> EXCLUDE_METADATA_FILENAME_FILTER - defines any patterns for which > input > >>>> files do NOT get their metadata extracted, by file name > >>>> EXCLUDE_CONTENT_MIMETYPE_FILTER - defines any patterns for which input > >>>> files do NOT get their content extracted, by MIME type > >>>> EXCLUDE_METADATA_MIMETYPE_FILTER - defines any patterns for which > input > >>>> files do NOT get their metadata extracted, by MIME type > >>>> > >>>> I believe this gets all the bases covered. At processor init time, we > >> can > >>>> analyze the inclusions vs. exclusions; any overlap would cause a > >>>> configuration error. > >>>> > >>>> Let me know what you think, thanks. > >>>> - Dmitry > >>>> > >>>> On Wed, Mar 30, 2016 at 10:41 AM, Dmitry Goldenberg < > >>>> [email protected]> wrote: > >>>> > >>>>> Hi Joe, > >>>>> > >>>>> I follow your reasoning on the semantics of "media". One might argue > >>>> that > >>>>> media files are a case of "document" or that a document is a case of > >>>>> "media". > >>>>> > >>>>> I'm not proposing filters for the mode of processing, I'm proposing a > >>>>> flag/enum with 3 values: > >>>>> > >>>>> A) extract metadata only; > >>>>> B) extract content only and place it into the flowfile content; > >>>>> C) extract both metadata and content. > >>>>> > >>>>> I think the default should be C, to extract both. At least in my > >>>>> experience most flows I've dealt with were interested in extracting > >>> both. > >>>>> > >>>>> I don't see how this mode would benefit from being expression driven > >> - > >>> ? > >>>>> > >>>>> I think we can add this enum mode and have the basic use case > >> covered. > >>>>> > >>>>> Additionally, further down the line, I was thinking we could ponder > >> the > >>>>> following (these have been essential in search engine ingestion): > >>>>> > >>>>> 1. Extraction from compressed files/archives. How would > >>> UnpackContent > >>>>> work with ExtractMediaAttributes? Use-case being, we've got a zip > >>>> file as > >>>>> input and want to crack it open and unravel it recursively; it may > >>>> have > >>>>> other, nested zips inside, along with other documents. One way to > >>>> handle > >>>>> this is to treat the whole archive as one document and merge all > >>>> attributes > >>>>> into one FlowFile. The other way would be to treat each archive > >>>> entry as > >>>>> its own flow file and keep a pointer back at the parent archive. > >>> Yet > >>>>> another case is when the user might want to only extract the > >> 'leaf' > >>>> entries > >>>>> and discard any parent container archives. > >>>>> > >>>>> 2. Attachments and embeddings. Users may want to treat any > >> attached > >>> or > >>>>> embedded files as separate flowfiles with perhaps pointers back to > >>> the > >>>>> parent files. This definitely warrants a filter. Oftentimes Office > >>>>> documents have 'media' embeddings which are often not of interest, > >>>>> especially for the case of ingesting into a search engine. > >>>>> > >>>>> 3. PDF. For PDF's, we can do OCR. This is important for the > >>>>> 'image'/scanned PDF's for which Tika won't extract text. > >>>>> > >>>>> I'd like to understand how much of this is already supported in NiFi > >>> and > >>>>> if not I'd volunteer/collaborate to implement some of this. > >>>>> > >>>>> - Dmitry > >>>>> > >>>>> > >>>>> On Wed, Mar 30, 2016 at 9:03 AM, Joe Skora <[email protected]> wrote: > >>>>> > >>>>>> Dmitry, > >>>>>> > >>>>>> Are you proposing separate filters that determine the mode of > >>>> processing, > >>>>>> metadata/content/metadataAndContent? I was thinking of one > >> selection > >>>>>> filters and a static mode switch at the processor instance level, to > >>>> make > >>>>>> configuration more obvious such that one instance of the processor > >>> will > >>>>>> handle a known set of files regardless of the processing mode. > >>>>>> > >>>>>> I was thinking it would be useful for the mode switch to support > >>>>>> expression > >>>>>> language, but I'm not sure about that since the selection filters > >> will > >>>>>> control what files get processed and it would be harder to configure > >>> if > >>>>>> the > >>>>>> output flow file could vary between source format and extracted > >> text. > >>>> So, > >>>>>> while it might be easy to do, and occasionally useful, I think in > >>> normal > >>>>>> use I'd never have a varying mode but would more likely have > >> multiple > >>>>>> processor instances with some routing or selection going on further > >>>>>> upstream. > >>>>>> > >>>>>> I wrestled with the naming issue too. I went with > >>>>>> "ExtractMediaAttributes" > >>>>>> over "ExtractDocumentAttributes" because it seemed to represent the > >>>>>> broader > >>>>>> context better. In reality, media files and documents and documents > >>> are > >>>>>> media files, but in the end it's all just semantics. > >>>>>> > >>>>>> I don't think I would change the NAR bundle name, because I think > >>>>>> "nifi-media-nar" establishes it as a place to collect this and other > >>>> media > >>>>>> related processors in the future. > >>>>>> > >>>>>> Regards, > >>>>>> Joe > >>>>>> > >>>>>> On Tue, Mar 29, 2016 at 3:09 PM, Dmitry Goldenberg < > >>>>>> [email protected] > >>>>>>> wrote: > >>>>>> > >>>>>>> Hi Joe, > >>>>>>> > >>>>>>> Thanks for all the details. > >>>>>>> > >>>>>>> I wanted to propose that I do some of this work so as to go > >> through > >>>> the > >>>>>>> full cycle of developing a processor and committing it. > >>>>>>> > >>>>>>> Once your changes are merged, I could extend your > >>>> 'ExtractMediaMetadata' > >>>>>>> processor to handle the content, in addition to the metadata. > >>>>>>> > >>>>>>> We could keep the FILENAME_FILTER and MIMETYPE_FILTER but add a > >> mode > >>>>>> with 3 > >>>>>>> values: metadataOnly, contentOnly, metadataAndContent. > >>>>>>> > >>>>>>> One thing that looks to be a design issue right now is, your > >> changes > >>>> and > >>>>>>> the 'nomenclature' seem media-oriented ("nifi-media-nar" etc.) > >>>>>>> > >>>>>>> Would it make sense to have a generic processor > >>>>>>> ExtractDocumentMetadataAndContent? Are there enough specifics in > >>> the > >>>>>>> image/video processing stuff to warrant that to be a separate > >> layer; > >>>>>>> perhaps a subclass of ExtractDocumentMetadataAndContent ? Might > >> it > >>>> make > >>>>>>> sense to rename nifi-media-nar into nifi-text-extract-nar ? > >>>>>>> > >>>>>>> Thanks, > >>>>>>> - Dmitry > >>>>>>> > >>>>>>> > >>>>>>> > >>>>>>> On Tue, Mar 29, 2016 at 2:36 PM, Joe Skora <[email protected]> > >>> wrote: > >>>>>>> > >>>>>>>> Dmitry, > >>>>>>>> > >>>>>>>> Yeah, I agree, Tika is pretty impressive. The original ticket, > >>>>>> NIFI-615 > >>>>>>>> <https://issues.apache.org/jira/browse/NIFI-615>, wanted > >>> extraction > >>>>>> of > >>>>>>>> metadata from WAV files, but as I got into it I found Tika so > >> for > >>>> the > >>>>>>> same > >>>>>>>> effort it supports the 1,000+ file formats Tika understands. > >> That > >>>> new > >>>>>>>> processor called "ExtractMediaMetadata", you can pull that pull > >>>> PR-252 > >>>>>>>> <https://github.com/apache/nifi/pull/252> from GitHub if you > >> want > >>>> to > >>>>>>> give > >>>>>>>> it a try before it's merged. > >>>>>>>> > >>>>>>>> Extraction content for those 1,000+ formats would be a valuable > >>>>>> addition. > >>>>>>>> I see two possible approaches, 1) create a new > >>> "ExtractMediaContent" > >>>>>>>> processor that would put the document content in a new flow > >> file, > >>>> and > >>>>>> 2) > >>>>>>>> extend the new "ExtractMediaMetadata" processor so it can > >> extract > >>>>>>> metadata, > >>>>>>>> content, or both. One combined processor makes sense if it can > >>>>>> provide a > >>>>>>>> performance gain, otherwise two complementary processors may > >> make > >>>>>> usage > >>>>>>>> easier. > >>>>>>>> > >>>>>>>> I'm glad to help if you want to take a cut at the processor > >>>> yourself, > >>>>>> or > >>>>>>> I > >>>>>>>> can take a crack at it myself if you'd prefer. > >>>>>>>> > >>>>>>>> Don't hesitate to ask questions or share comments and feedback > >>>>>> regarding > >>>>>>>> the ExtractMediaMetadata processor or the addition of content > >>>>>> handling. > >>>>>>>> > >>>>>>>> Regards, > >>>>>>>> Joe Skora > >>>>>>>> > >>>>>>>> On Thu, Mar 24, 2016 at 11:40 AM, Dmitry Goldenberg < > >>>>>>>> [email protected]> wrote: > >>>>>>>> > >>>>>>>>> Thanks, Joe! > >>>>>>>>> > >>>>>>>>> Hi Joe S. - I'm definitely up for discussing and contributing. > >>>>>>>>> > >>>>>>>>> While building search-related ingestion systems, I've seen > >>>> metadata > >>>>>> and > >>>>>>>>> text extraction being done all the time; it's always there and > >>>>>> always > >>>>>>> has > >>>>>>>>> to be done for building search indexes. Beyond that, > >>> OCR-related > >>>>>>>>> capabilities are often requested, and the advantage of Tika is > >>>> that > >>>>>> it > >>>>>>>>> supports OCR out of the box. > >>>>>>>>> > >>>>>>>>> - Dmitry > >>>>>>>>> > >>>>>>>>> On Thu, Mar 24, 2016 at 11:36 AM, Joe Witt < > >> [email protected]> > >>>>>> wrote: > >>>>>>>>> > >>>>>>>>>> Dmitry, > >>>>>>>>>> > >>>>>>>>>> Another community member (Joe Skora) has a PR outstanding > >> for > >>>>>>>>>> extracting metadata from media files using Tika. Perhaps it > >>>> makes > >>>>>>>>>> sense to broaden that to in general extract what Tika can > >>> find. > >>>>>> Joe > >>>>>>> - > >>>>>>>>>> perhaps you can discuss your ideas with Dmitry and see if > >>>>>> broadening > >>>>>>>>>> is a good idea or if rather domain specific ones make more > >>>> sense. > >>>>>>>>>> > >>>>>>>>>> This concept of extracting metadata from documents/text > >> files, > >>>>>> etc.. > >>>>>>>>>> using something like Tika is certainly useful as that then > >> can > >>>>>> drive > >>>>>>>>>> nice automated routing decisions. > >>>>>>>>>> > >>>>>>>>>> Thanks > >>>>>>>>>> Joe > >>>>>>>>>> > >>>>>>>>>> On Thu, Mar 24, 2016 at 9:28 AM, Dmitry Goldenberg > >>>>>>>>>> <[email protected]> wrote: > >>>>>>>>>>> Hi, > >>>>>>>>>>> > >>>>>>>>>>> I see that the ExtractText processor extracts text using > >>>> regex. > >>>>>>>>>>> > >>>>>>>>>>> What about a processor that extracts text and metadata > >> from > >>>>>>> incoming > >>>>>>>>>>> files? That doesn't seem to exist - but perhaps I didn't > >>>> quite > >>>>>>> look > >>>>>>>> in > >>>>>>>>>> the > >>>>>>>>>>> right spots. > >>>>>>>>>>> > >>>>>>>>>>> If that doesn't exist I'd like to implement and commit it, > >>>> using > >>>>>>>> Apache > >>>>>>>>>>> Tika. There may also be a couple of related processors to > >>>> that. > >>>>>>>>>>> > >>>>>>>>>>> Thoughts? > >>>>>>>>>>> > >>>>>>>>>>> Thanks, > >>>>>>>>>>> - Dmitry > >>>>>>>>>> > >>>>>>>>> > >>>>>>>> > >>>>>>> > >>>>>> > >>>>> > >>>>> > >>>> > >>> > >> > >
