Metadata-Version: 2.5
Name: smart-slice
Version: 0.5.1
Summary: Fidelity-first document slicing for RAG pipelines: ~30 formats in, retrieval-ready paragraphs out.
Project-URL: Homepage, https://github.com/guangxiangdebizi/smart-slice
Project-URL: Repository, https://github.com/guangxiangdebizi/smart-slice
Project-URL: Issues, https://github.com/guangxiangdebizi/smart-slice/issues
Project-URL: Changelog, https://github.com/guangxiangdebizi/smart-slice/blob/main/CHANGELOG.md
License:                     GNU GENERAL PUBLIC LICENSE
                               Version 3, 29 June 2007
        
         Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
         Everyone is permitted to copy and distribute verbatim copies
         of this license document, but changing it is not allowed.
        
                                    Preamble
        
          The GNU General Public License is a free, copyleft license for
        software and other kinds of works.
        
          The licenses for most software and other practical works are designed
        to take away your freedom to share and change the works.  By contrast,
        the GNU General Public License is intended to guarantee your freedom to
        share and change all versions of a program--to make sure it remains free
        software for all its users.  We, the Free Software Foundation, use the
        GNU General Public License for most of our software; it applies also to
        any other work released this way by its authors.  You can apply it to
        your programs, too.
        
          When we speak of free software, we are referring to freedom, not
        price.  Our General Public Licenses are designed to make sure that you
        have the freedom to distribute copies of free software (and charge for
        them if you wish), that you receive source code or can get it if you
        want it, that you can change the software or use pieces of it in new
        free programs, and that you know you can do these things.
        
          To protect your rights, we need to prevent others from denying you
        these rights or asking you to surrender the rights.  Therefore, you have
        certain responsibilities if you distribute copies of the software, or if
        you modify it: responsibilities to respect the freedom of others.
        
          For example, if you distribute copies of such a program, whether
        gratis or for a fee, you must pass on to the recipients the same
        freedoms that you received.  You must make sure that they, too, receive
        or can get the source code.  And you must show them these terms so they
        know their rights.
        
          Developers that use the GNU GPL protect your rights with two steps:
        (1) assert copyright on the software, and (2) offer you this License
        giving you legal permission to copy, distribute and/or modify it.
        
          For the developers' and authors' protection, the GPL clearly explains
        that there is no warranty for this free software.  For both users' and
        authors' sake, the GPL requires that modified versions be marked as
        changed, so that their problems will not be attributed erroneously to
        authors of previous versions.
        
          Some devices are designed to deny users access to install or run
        modified versions of the software inside them, although the manufacturer
        can do so.  This is fundamentally incompatible with the aim of
        protecting users' freedom to change the software.  The systematic
        pattern of such abuse occurs in the area of products for individuals to
        use, which is precisely where it is most unacceptable.  Therefore, we
        have designed this version of the GPL to prohibit the practice for those
        products.  If such problems arise substantially in other domains, we
        stand ready to extend this provision to those domains in future versions
        of the GPL, as needed to protect the freedom of users.
        
          Finally, every program is threatened constantly by software patents.
        States should not allow patents to restrict development and use of
        software on general-purpose computers, but in those that do, we wish to
        avoid the special danger that patents applied to a free program could
        make it effectively proprietary.  To prevent this, the GPL assures that
        patents cannot be used to render the program non-free.
        
          The precise terms and conditions for copying, distribution and
        modification follow.
        
                               TERMS AND CONDITIONS
        
          0. Definitions.
        
          "This License" refers to version 3 of the GNU General Public License.
        
          "Copyright" also means copyright-like laws that apply to other kinds of
        works, such as semiconductor masks.
        
          "The Program" refers to any copyrightable work licensed under this
        License.  Each licensee is addressed as "you".  "Licensees" and
        "recipients" may be individuals or organizations.
        
          To "modify" a work means to copy from or adapt all or part of the work
        in a fashion requiring copyright permission, other than the making of an
        exact copy.  The resulting work is called a "modified version" of the
        earlier work or a work "based on" the earlier work.
        
          A "covered work" means either the unmodified Program or a work based
        on the Program.
        
          To "propagate" a work means to do anything with it that, without
        permission, would make you directly or secondarily liable for
        infringement under applicable copyright law, except executing it on a
        computer or modifying a private copy.  Propagation includes copying,
        distribution (with or without modification), making available to the
        public, and in some countries other activities as well.
        
          To "convey" a work means any kind of propagation that enables other
        parties to make or receive copies.  Mere interaction with a user through
        a computer network, with no transfer of a copy, is not conveying.
        
          An interactive user interface displays "Appropriate Legal Notices"
        to the extent that it includes a convenient and prominently visible
        feature that (1) displays an appropriate copyright notice, and (2)
        tells the user that there is no warranty for the work (except to the
        extent that warranties are provided), that licensees may convey the
        work under this License, and how to view a copy of this License.  If
        the interface presents a list of user commands or options, such as a
        menu, a prominent item in the list meets this criterion.
        
          1. Source Code.
        
          The "source code" for a work means the preferred form of the work
        for making modifications to it.  "Object code" means any non-source
        form of a work.
        
          A "Standard Interface" means an interface that either is an official
        standard defined by a recognized standards body, or, in the case of
        interfaces specified for a particular programming language, one that
        is widely used among developers working in that language.
        
          The "System Libraries" of an executable work include anything, other
        than the work as a whole, that (a) is included in the normal form of
        packaging a Major Component, but which is not part of that Major
        Component, and (b) serves only to enable use of the work with that
        Major Component, or to implement a Standard Interface for which an
        implementation is available to the public in source code form.  A
        "Major Component", in this context, means a major essential component
        (kernel, window system, and so on) of the specific operating system
        (if any) on which the executable work runs, or a compiler used to
        produce the work, or an object code interpreter used to run it.
        
          The "Corresponding Source" for a work in object code form means all
        the source code needed to generate, install, and (for an executable
        work) run the object code and to modify the work, including scripts to
        control those activities.  However, it does not include the work's
        System Libraries, or general-purpose tools or generally available free
        programs which are used unmodified in performing those activities but
        which are not part of the work.  For example, Corresponding Source
        includes interface definition files associated with source files for
        the work, and the source code for shared libraries and dynamically
        linked subprograms that the work is specifically designed to require,
        such as by intimate data communication or control flow between those
        subprograms and other parts of the work.
        
          The Corresponding Source need not include anything that users
        can regenerate automatically from other parts of the Corresponding
        Source.
        
          The Corresponding Source for a work in source code form is that
        same work.
        
          2. Basic Permissions.
        
          All rights granted under this License are granted for the term of
        copyright on the Program, and are irrevocable provided the stated
        conditions are met.  This License explicitly affirms your unlimited
        permission to run the unmodified Program.  The output from running a
        covered work is covered by this License only if the output, given its
        content, constitutes a covered work.  This License acknowledges your
        rights of fair use or other equivalent, as provided by copyright law.
        
          You may make, run and propagate covered works that you do not
        convey, without conditions so long as your license otherwise remains
        in force.  You may convey covered works to others for the sole purpose
        of having them make modifications exclusively for you, or provide you
        with facilities for running those works, provided that you comply with
        the terms of this License in conveying all material for which you do
        not control copyright.  Those thus making or running the covered works
        for you must do so exclusively on your behalf, under your direction
        and control, on terms that prohibit them from making any copies of
        your copyrighted material outside their relationship with you.
        
          Conveying under any other circumstances is permitted solely under
        the conditions stated below.  Sublicensing is not allowed; section 10
        makes it unnecessary.
        
          3. Protecting Users' Legal Rights From Anti-Circumvention Law.
        
          No covered work shall be deemed part of an effective technological
        measure under any applicable law fulfilling obligations under article
        11 of the WIPO copyright treaty adopted on 20 December 1996, or
        similar laws prohibiting or restricting circumvention of such
        measures.
        
          When you convey a covered work, you waive any legal power to forbid
        circumvention of technological measures to the extent such circumvention
        is effected by exercising rights under this License with respect to
        the covered work, and you disclaim any intention to limit operation or
        modification of the work as a means of enforcing, against the work's
        users, your or third parties' legal rights to forbid circumvention of
        technological measures.
        
          4. Conveying Verbatim Copies.
        
          You may convey verbatim copies of the Program's source code as you
        receive it, in any medium, provided that you conspicuously and
        appropriately publish on each copy an appropriate copyright notice;
        keep intact all notices stating that this License and any
        non-permissive terms added in accord with section 7 apply to the code;
        keep intact all notices of the absence of any warranty; and give all
        recipients a copy of this License along with the Program.
        
          You may charge any price or no price for each copy that you convey,
        and you may offer support or warranty protection for a fee.
        
          5. Conveying Modified Source Versions.
        
          You may convey a work based on the Program, or the modifications to
        produce it from the Program, in the form of source code under the
        terms of section 4, provided that you also meet all of these conditions:
        
            a) The work must carry prominent notices stating that you modified
            it, and giving a relevant date.
        
            b) The work must carry prominent notices stating that it is
            released under this License and any conditions added under section
            7.  This requirement modifies the requirement in section 4 to
            "keep intact all notices".
        
            c) You must license the entire work, as a whole, under this
            License to anyone who comes into possession of a copy.  This
            License will therefore apply, along with any applicable section 7
            additional terms, to the whole of the work, and all its parts,
            regardless of how they are packaged.  This License gives no
            permission to license the work in any other way, but it does not
            invalidate such permission if you have separately received it.
        
            d) If the work has interactive user interfaces, each must display
            Appropriate Legal Notices; however, if the Program has interactive
            interfaces that do not display Appropriate Legal Notices, your
            work need not make them do so.
        
          A compilation of a covered work with other separate and independent
        works, which are not by their nature extensions of the covered work,
        and which are not combined with it such as to form a larger program,
        in or on a volume of a storage or distribution medium, is called an
        "aggregate" if the compilation and its resulting copyright are not
        used to limit the access or legal rights of the compilation's users
        beyond what the individual works permit.  Inclusion of a covered work
        in an aggregate does not cause this License to apply to the other
        parts of the aggregate.
        
          6. Conveying Non-Source Forms.
        
          You may convey a covered work in object code form under the terms
        of sections 4 and 5, provided that you also convey the
        machine-readable Corresponding Source under the terms of this License,
        in one of these ways:
        
            a) Convey the object code in, or embodied in, a physical product
            (including a physical distribution medium), accompanied by the
            Corresponding Source fixed on a durable physical medium
            customarily used for software interchange.
        
            b) Convey the object code in, or embodied in, a physical product
            (including a physical distribution medium), accompanied by a
            written offer, valid for at least three years and valid for as
            long as you offer spare parts or customer support for that product
            model, to give anyone who possesses the object code either (1) a
            copy of the Corresponding Source for all the software in the
            product that is covered by this License, on a durable physical
            medium customarily used for software interchange, for a price no
            more than your reasonable cost of physically performing this
            conveying of source, or (2) access to copy the
            Corresponding Source from a network server at no charge.
        
            c) Convey individual copies of the object code with a copy of the
            written offer to provide the Corresponding Source.  This
            alternative is allowed only occasionally and noncommercially, and
            only if you received the object code with such an offer, in accord
            with subsection 6b.
        
            d) Convey the object code by offering access from a designated
            place (gratis or for a charge), and offer equivalent access to the
            Corresponding Source in the same way through the same place at no
            further charge.  You need not require recipients to copy the
            Corresponding Source along with the object code.  If the place to
            copy the object code is a network server, the Corresponding Source
            may be on a different server (operated by you or a third party)
            that supports equivalent copying facilities, provided you maintain
            clear directions next to the object code saying where to find the
            Corresponding Source.  Regardless of what server hosts the
            Corresponding Source, you remain obligated to ensure that it is
            available for as long as needed to satisfy these requirements.
        
            e) Convey the object code using peer-to-peer transmission, provided
            you inform other peers where the object code and Corresponding
            Source of the work are being offered to the general public at no
            charge under subsection 6d.
        
          A separable portion of the object code, whose source code is excluded
        from the Corresponding Source as a System Library, need not be
        included in conveying the object code work.
        
          A "User Product" is either (1) a "consumer product", which means any
        tangible personal property which is normally used for personal, family,
        or household purposes, or (2) anything designed or sold for incorporation
        into a dwelling.  In determining whether a product is a consumer product,
        doubtful cases shall be resolved in favor of coverage.  For a particular
        product received by a particular user, "normally used" refers to a
        typical or common use of that class of product, regardless of the status
        of the particular user or of the way in which the particular user
        actually uses, or expects or is expected to use, the product.  A product
        is a consumer product regardless of whether the product has substantial
        commercial, industrial or non-consumer uses, unless such uses represent
        the only significant mode of use of the product.
        
          "Installation Information" for a User Product means any methods,
        procedures, authorization keys, or other information required to install
        and execute modified versions of a covered work in that User Product from
        a modified version of its Corresponding Source.  The information must
        suffice to ensure that the continued functioning of the modified object
        code is in no case prevented or interfered with solely because
        modification has been made.
        
          If you convey an object code work under this section in, or with, or
        specifically for use in, a User Product, and the conveying occurs as
        part of a transaction in which the right of possession and use of the
        User Product is transferred to the recipient in perpetuity or for a
        fixed term (regardless of how the transaction is characterized), the
        Corresponding Source conveyed under this section must be accompanied
        by the Installation Information.  But this requirement does not apply
        if neither you nor any third party retains the ability to install
        modified object code on the User Product (for example, the work has
        been installed in ROM).
        
          The requirement to provide Installation Information does not include a
        requirement to continue to provide support service, warranty, or updates
        for a work that has been modified or installed by the recipient, or for
        the User Product in which it has been modified or installed.  Access to a
        network may be denied when the modification itself materially and
        adversely affects the operation of the network or violates the rules and
        protocols for communication across the network.
        
          Corresponding Source conveyed, and Installation Information provided,
        in accord with this section must be in a format that is publicly
        documented (and with an implementation available to the public in
        source code form), and must require no special password or key for
        unpacking, reading or copying.
        
          7. Additional Terms.
        
          "Additional permissions" are terms that supplement the terms of this
        License by making exceptions from one or more of its conditions.
        Additional permissions that are applicable to the entire Program shall
        be treated as though they were included in this License, to the extent
        that they are valid under applicable law.  If additional permissions
        apply only to part of the Program, that part may be used separately
        under those permissions, but the entire Program remains governed by
        this License without regard to the additional permissions.
        
          When you convey a copy of a covered work, you may at your option
        remove any additional permissions from that copy, or from any part of
        it.  (Additional permissions may be written to require their own
        removal in certain cases when you modify the work.)  You may place
        additional permissions on material, added by you to a covered work,
        for which you have or can give appropriate copyright permission.
        
          Notwithstanding any other provision of this License, for material you
        add to a covered work, you may (if authorized by the copyright holders of
        that material) supplement the terms of this License with terms:
        
            a) Disclaiming warranty or limiting liability differently from the
            terms of sections 15 and 16 of this License; or
        
            b) Requiring preservation of specified reasonable legal notices or
            author attributions in that material or in the Appropriate Legal
            Notices displayed by works containing it; or
        
            c) Prohibiting misrepresentation of the origin of that material, or
            requiring that modified versions of such material be marked in
            reasonable ways as different from the original version; or
        
            d) Limiting the use for publicity purposes of names of licensors or
            authors of the material; or
        
            e) Declining to grant rights under trademark law for use of some
            trade names, trademarks, or service marks; or
        
            f) Requiring indemnification of licensors and authors of that
            material by anyone who conveys the material (or modified versions of
            it) with contractual assumptions of liability to the recipient, for
            any liability that these contractual assumptions directly impose on
            those licensors and authors.
        
          All other non-permissive additional terms are considered "further
        restrictions" within the meaning of section 10.  If the Program as you
        received it, or any part of it, contains a notice stating that it is
        governed by this License along with a term that is a further
        restriction, you may remove that term.  If a license document contains
        a further restriction but permits relicensing or conveying under this
        License, you may add to a covered work material governed by the terms
        of that license document, provided that the further restriction does
        not survive such relicensing or conveying.
        
          If you add terms to a covered work in accord with this section, you
        must place, in the relevant source files, a statement of the
        additional terms that apply to those files, or a notice indicating
        where to find the applicable terms.
        
          Additional terms, permissive or non-permissive, may be stated in the
        form of a separately written license, or stated as exceptions;
        the above requirements apply either way.
        
          8. Termination.
        
          You may not propagate or modify a covered work except as expressly
        provided under this License.  Any attempt otherwise to propagate or
        modify it is void, and will automatically terminate your rights under
        this License (including any patent licenses granted under the third
        paragraph of section 11).
        
          However, if you cease all violation of this License, then your
        license from a particular copyright holder is reinstated (a)
        provisionally, unless and until the copyright holder explicitly and
        finally terminates your license, and (b) permanently, if the copyright
        holder fails to notify you of the violation by some reasonable means
        prior to 60 days after the cessation.
        
          Moreover, your license from a particular copyright holder is
        reinstated permanently if the copyright holder notifies you of the
        violation by some reasonable means, this is the first time you have
        received notice of violation of this License (for any work) from that
        copyright holder, and you cure the violation prior to 30 days after
        your receipt of the notice.
        
          Termination of your rights under this section does not terminate the
        licenses of parties who have received copies or rights from you under
        this License.  If your rights have been terminated and not permanently
        reinstated, you do not qualify to receive new licenses for the same
        material under section 10.
        
          9. Acceptance Not Required for Having Copies.
        
          You are not required to accept this License in order to receive or
        run a copy of the Program.  Ancillary propagation of a covered work
        occurring solely as a consequence of using peer-to-peer transmission
        to receive a copy likewise does not require acceptance.  However,
        nothing other than this License grants you permission to propagate or
        modify any covered work.  These actions infringe copyright if you do
        not accept this License.  Therefore, by modifying or propagating a
        covered work, you indicate your acceptance of this License to do so.
        
          10. Automatic Licensing of Downstream Recipients.
        
          Each time you convey a covered work, the recipient automatically
        receives a license from the original licensors, to run, modify and
        propagate that work, subject to this License.  You are not responsible
        for enforcing compliance by third parties with this License.
        
          An "entity transaction" is a transaction transferring control of an
        organization, or substantially all assets of one, or subdividing an
        organization, or merging organizations.  If propagation of a covered
        work results from an entity transaction, each party to that
        transaction who receives a copy of the work also receives whatever
        licenses to the work the party's predecessor in interest had or could
        give under the previous paragraph, plus a right to possession of the
        Corresponding Source of the work from the predecessor in interest, if
        the predecessor has it or can get it with reasonable efforts.
        
          You may not impose any further restrictions on the exercise of the
        rights granted or affirmed under this License.  For example, you may
        not impose a license fee, royalty, or other charge for exercise of
        rights granted under this License, and you may not initiate litigation
        (including a cross-claim or counterclaim in a lawsuit) alleging that
        any patent claim is infringed by making, using, selling, offering for
        sale, or importing the Program or any portion of it.
        
          11. Patents.
        
          A "contributor" is a copyright holder who authorizes use under this
        License of the Program or a work on which the Program is based.  The
        work thus licensed is called the contributor's "contributor version".
        
          A contributor's "essential patent claims" are all patent claims
        owned or controlled by the contributor, whether already acquired or
        hereafter acquired, that would be infringed by some manner, permitted
        by this License, of making, using, or selling its contributor version,
        but do not include claims that would be infringed only as a
        consequence of further modification of the contributor version.  For
        purposes of this definition, "control" includes the right to grant
        patent sublicenses in a manner consistent with the requirements of
        this License.
        
          Each contributor grants you a non-exclusive, worldwide, royalty-free
        patent license under the contributor's essential patent claims, to
        make, use, sell, offer for sale, import and otherwise run, modify and
        propagate the contents of its contributor version.
        
          In the following three paragraphs, a "patent license" is any express
        agreement or commitment, however denominated, not to enforce a patent
        (such as an express permission to practice a patent or covenant not to
        sue for patent infringement).  To "grant" such a patent license to a
        party means to make such an agreement or commitment not to enforce a
        patent against the party.
        
          If you convey a covered work, knowingly relying on a patent license,
        and the Corresponding Source of the work is not available for anyone
        to copy, free of charge and under the terms of this License, through a
        publicly available network server or other readily accessible means,
        then you must either (1) cause the Corresponding Source to be so
        available, or (2) arrange to deprive yourself of the benefit of the
        patent license for this particular work, or (3) arrange, in a manner
        consistent with the requirements of this License, to extend the patent
        license to downstream recipients.  "Knowingly relying" means you have
        actual knowledge that, but for the patent license, your conveying the
        covered work in a country, or your recipient's use of the covered work
        in a country, would infringe one or more identifiable patents in that
        country that you have reason to believe are valid.
        
          If, pursuant to or in connection with a single transaction or
        arrangement, you convey, or propagate by procuring conveyance of, a
        covered work, and grant a patent license to some of the parties
        receiving the covered work authorizing them to use, propagate, modify
        or convey a specific copy of the covered work, then the patent license
        you grant is automatically extended to all recipients of the covered
        work and works based on it.
        
          A patent license is "discriminatory" if it does not include within
        the scope of its coverage, prohibits the exercise of, or is
        conditioned on the non-exercise of one or more of the rights that are
        specifically granted under this License.  You may not convey a covered
        work if you are a party to an arrangement with a third party that is
        in the business of distributing software, under which you make payment
        to the third party based on the extent of your activity of conveying
        the work, and under which the third party grants, to any of the
        parties who would receive the covered work from you, a discriminatory
        patent license (a) in connection with copies of the covered work
        conveyed by you (or copies made from those copies), or (b) primarily
        for and in connection with specific products or compilations that
        contain the covered work, unless you entered into that arrangement,
        or that patent license was granted, prior to 28 March 2007.
        
          Nothing in this License shall be construed as excluding or limiting
        any implied license or other defenses to infringement that may
        otherwise be available to you under applicable patent law.
        
          12. No Surrender of Others' Freedom.
        
          If conditions are imposed on you (whether by court order, agreement or
        otherwise) that contradict the conditions of this License, they do not
        excuse you from the conditions of this License.  If you cannot convey a
        covered work so as to satisfy simultaneously your obligations under this
        License and any other pertinent obligations, then as a consequence you may
        not convey it at all.  For example, if you agree to terms that obligate you
        to collect a royalty for further conveying from those to whom you convey
        the Program, the only way you could satisfy both those terms and this
        License would be to refrain entirely from conveying the Program.
        
          13. Use with the GNU Affero General Public License.
        
          Notwithstanding any other provision of this License, you have
        permission to link or combine any covered work with a work licensed
        under version 3 of the GNU Affero General Public License into a single
        combined work, and to convey the resulting work.  The terms of this
        License will continue to apply to the part which is the covered work,
        but the special requirements of the GNU Affero General Public License,
        section 13, concerning interaction through a network will apply to the
        combination as such.
        
          14. Revised Versions of this License.
        
          The Free Software Foundation may publish revised and/or new versions of
        the GNU General Public License from time to time.  Such new versions will
        be similar in spirit to the present version, but may differ in detail to
        address new problems or concerns.
        
          Each version is given a distinguishing version number.  If the
        Program specifies that a certain numbered version of the GNU General
        Public License "or any later version" applies to it, you have the
        option of following the terms and conditions either of that numbered
        version or of any later version published by the Free Software
        Foundation.  If the Program does not specify a version number of the
        GNU General Public License, you may choose any version ever published
        by the Free Software Foundation.
        
          If the Program specifies that a proxy can decide which future
        versions of the GNU General Public License can be used, that proxy's
        public statement of acceptance of a version permanently authorizes you
        to choose that version for the Program.
        
          Later license versions may give you additional or different
        permissions.  However, no additional obligations are imposed on any
        author or copyright holder as a result of your choosing to follow a
        later version.
        
          15. Disclaimer of Warranty.
        
          THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
        APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
        HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
        OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
        THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
        PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
        IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
        ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
        
          16. Limitation of Liability.
        
          IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
        WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
        THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
        GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
        USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
        DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
        PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
        EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
        SUCH DAMAGES.
        
          17. Interpretation of Sections 15 and 16.
        
          If the disclaimer of warranty and limitation of liability provided
        above cannot be given local legal effect according to their terms,
        reviewing courts shall apply local law that most closely approximates
        an absolute waiver of all civil liability in connection with the
        Program, unless a warranty or assumption of liability accompanies a
        copy of the Program in return for a fee.
        
                             END OF TERMS AND CONDITIONS
        
                    How to Apply These Terms to Your New Programs
        
          If you develop a new program, and you want it to be of the greatest
        possible use to the public, the best way to achieve this is to make it
        free software which everyone can redistribute and change under these terms.
        
          To do so, attach the following notices to the program.  It is safest
        to attach them to the start of each source file to most effectively
        state the exclusion of warranty; and each file should have at least
        the "copyright" line and a pointer to where the full notice is found.
        
            <one line to give the program's name and a brief idea of what it does.>
            Copyright (C) <year>  <name of author>
        
            This program is free software: you can redistribute it and/or modify
            it under the terms of the GNU General Public License as published by
            the Free Software Foundation, either version 3 of the License, or
            (at your option) any later version.
        
            This program is distributed in the hope that it will be useful,
            but WITHOUT ANY WARRANTY; without even the implied warranty of
            MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
            GNU General Public License for more details.
        
            You should have received a copy of the GNU General Public License
            along with this program.  If not, see <https://www.gnu.org/licenses/>.
        
        Also add information on how to contact you by electronic and paper mail.
        
          If the program does terminal interaction, make it output a short
        notice like this when it starts in an interactive mode:
        
            <program>  Copyright (C) <year>  <name of author>
            This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
            This is free software, and you are welcome to redistribute it
            under certain conditions; type `show c' for details.
        
        The hypothetical commands `show w' and `show c' should show the appropriate
        parts of the General Public License.  Of course, your program's commands
        might be different; for a GUI interface, you would use an "about box".
        
          You should also get your employer (if you work as a programmer) or school,
        if any, to sign a "copyright disclaimer" for the program, if necessary.
        For more information on this, and how to apply and follow the GNU GPL, see
        <https://www.gnu.org/licenses/>.
        
          The GNU General Public License does not permit incorporating your program
        into proprietary programs.  If your program is a subroutine library, you
        may consider it more useful to permit linking proprietary applications with
        the library.  If this is what you want to do, use the GNU Lesser General
        Public License instead of this License.  But first, please read
        <https://www.gnu.org/licenses/why-not-lgpl.html>.
License-File: LICENSE
Keywords: chunking,document-parsing,embedding,image-extraction,knowledge-base,llm,markdown,multimodal,pdf,rag,text-splitting,vlm
Classifier: Development Status :: 4 - Beta
Classifier: Intended Audience :: Developers
Classifier: License :: OSI Approved :: GNU General Public License v3 (GPLv3)
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.9
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.13
Classifier: Topic :: Software Development :: Libraries :: Python Modules
Classifier: Topic :: Text Processing :: Indexing
Classifier: Typing :: Typed
Requires-Python: >=3.9
Requires-Dist: charset-normalizer>=3.0
Provides-Extra: accel
Requires-Dist: setuptools>=68; extra == 'accel'
Requires-Dist: ziglang>=0.11; extra == 'accel'
Provides-Extra: all
Requires-Dist: beautifulsoup4>=4.12; extra == 'all'
Requires-Dist: extract-msg>=0.48; extra == 'all'
Requires-Dist: jieba>=0.42; extra == 'all'
Requires-Dist: markdownify>=0.13; extra == 'all'
Requires-Dist: mobi>=0.3.3; extra == 'all'
Requires-Dist: olefile>=0.46; extra == 'all'
Requires-Dist: openpyxl>=3.1; extra == 'all'
Requires-Dist: pillow-heif>=0.13; extra == 'all'
Requires-Dist: pillow>=10.0; extra == 'all'
Requires-Dist: py7zr>=0.20; extra == 'all'
Requires-Dist: pypdf>=4.0; extra == 'all'
Requires-Dist: python-docx>=1.1; extra == 'all'
Requires-Dist: python-pptx>=1.0; extra == 'all'
Requires-Dist: striprtf>=0.0.26; extra == 'all'
Requires-Dist: uuid-utils>=0.6; extra == 'all'
Requires-Dist: xlrd>=2.0; extra == 'all'
Provides-Extra: archive
Requires-Dist: py7zr>=0.20; extra == 'archive'
Provides-Extra: dev
Requires-Dist: beautifulsoup4>=4.12; extra == 'dev'
Requires-Dist: extract-msg>=0.48; extra == 'dev'
Requires-Dist: jieba>=0.42; extra == 'dev'
Requires-Dist: markdownify>=0.13; extra == 'dev'
Requires-Dist: mobi>=0.3.3; extra == 'dev'
Requires-Dist: mypy>=1.8; extra == 'dev'
Requires-Dist: olefile>=0.46; extra == 'dev'
Requires-Dist: openpyxl>=3.1; extra == 'dev'
Requires-Dist: pillow-heif>=0.13; extra == 'dev'
Requires-Dist: pillow>=10.0; extra == 'dev'
Requires-Dist: py7zr>=0.20; extra == 'dev'
Requires-Dist: pypdf>=4.0; extra == 'dev'
Requires-Dist: pytest-cov>=4.0; extra == 'dev'
Requires-Dist: pytest>=7.0; extra == 'dev'
Requires-Dist: python-docx>=1.1; extra == 'dev'
Requires-Dist: python-pptx>=1.0; extra == 'dev'
Requires-Dist: ruff>=0.5; extra == 'dev'
Requires-Dist: setuptools>=68; extra == 'dev'
Requires-Dist: striprtf>=0.0.26; extra == 'dev'
Requires-Dist: uuid-utils>=0.6; extra == 'dev'
Requires-Dist: xlrd>=2.0; extra == 'dev'
Requires-Dist: xlwt>=1.3; extra == 'dev'
Requires-Dist: ziglang>=0.11; extra == 'dev'
Provides-Extra: ebook
Requires-Dist: mobi>=0.3.3; extra == 'ebook'
Provides-Extra: fastuuid
Requires-Dist: uuid-utils>=0.6; extra == 'fastuuid'
Provides-Extra: image
Requires-Dist: pillow-heif>=0.13; extra == 'image'
Requires-Dist: pillow>=10.0; extra == 'image'
Provides-Extra: keywords
Requires-Dist: jieba>=0.42; extra == 'keywords'
Provides-Extra: mail
Requires-Dist: extract-msg>=0.48; extra == 'mail'
Provides-Extra: markup
Requires-Dist: beautifulsoup4>=4.12; extra == 'markup'
Requires-Dist: markdownify>=0.13; extra == 'markup'
Provides-Extra: ocr
Requires-Dist: pillow>=10.0; extra == 'ocr'
Requires-Dist: rapidocr-onnxruntime>=1.3; extra == 'ocr'
Provides-Extra: office
Requires-Dist: olefile>=0.46; extra == 'office'
Requires-Dist: openpyxl>=3.1; extra == 'office'
Requires-Dist: python-docx>=1.1; extra == 'office'
Requires-Dist: python-pptx>=1.0; extra == 'office'
Requires-Dist: striprtf>=0.0.26; extra == 'office'
Requires-Dist: xlrd>=2.0; extra == 'office'
Provides-Extra: pdf
Requires-Dist: pillow>=10.0; extra == 'pdf'
Requires-Dist: pypdf>=4.0; extra == 'pdf'
Provides-Extra: test
Requires-Dist: beautifulsoup4>=4.12; extra == 'test'
Requires-Dist: extract-msg>=0.48; extra == 'test'
Requires-Dist: jieba>=0.42; extra == 'test'
Requires-Dist: markdownify>=0.13; extra == 'test'
Requires-Dist: mobi>=0.3.3; extra == 'test'
Requires-Dist: olefile>=0.46; extra == 'test'
Requires-Dist: openpyxl>=3.1; extra == 'test'
Requires-Dist: pillow-heif>=0.13; extra == 'test'
Requires-Dist: pillow>=10.0; extra == 'test'
Requires-Dist: py7zr>=0.20; extra == 'test'
Requires-Dist: pypdf>=4.0; extra == 'test'
Requires-Dist: python-docx>=1.1; extra == 'test'
Requires-Dist: python-pptx>=1.0; extra == 'test'
Requires-Dist: striprtf>=0.0.26; extra == 'test'
Requires-Dist: uuid-utils>=0.6; extra == 'test'
Requires-Dist: xlrd>=2.0; extra == 'test'
Requires-Dist: xlwt>=1.3; extra == 'test'
Description-Content-Type: text/markdown

# smart-slice

**Fidelity-first document slicing for RAG pipelines.** Feed it 197 file extensions across 30 handlers, get back retrieval-ready paragraphs that keep every original character reachable - plus the document's pictures, joined to the paragraphs they belong to.

[![License: GPL v3](https://img.shields.io/badge/License-GPLv3-blue.svg)](LICENSE)
[![Python](https://img.shields.io/badge/python-3.9%20%7C%203.10%20%7C%203.11%20%7C%203.12%20%7C%203.13-blue.svg)](https://www.python.org/downloads/)
[![Tests](https://img.shields.io/badge/tests-415%20passed-brightgreen.svg)]()
[![PyPI version](https://img.shields.io/pypi/v/smart-slice.svg)](https://pypi.org/project/smart-slice/)

```python
from smart_slice import slice_text

paragraphs = slice_text("# Chapter\n\nbody text", limit=1000)
# [{'title': 'Chapter', 'content': 'body text'}]
```

---

## Why another splitter?

Most text splitters treat a document as a bag of characters and chop it on a fixed grid. That loses the two things retrieval actually depends on: **where a passage sits in the document's structure**, and **the passage's exact original wording**. `smart-slice` is built around five rules.

### 1. Fidelity before cleverness

No cleaning step deletes source text.

- Heading markers (`#`) are stripped **only at a line start** and **only outside ` ``` ` code fences** — a `#RRGGBB` colour or a Python comment in a code block survives.
- A markdown table cut by the length budget gets its **header row appended** to the continuation chunk, not rewritten into it. The original rows are untouched.
- An oversized spreadsheet row is **re-split**, never truncated.

### 2. Structure before length

`limit` is a *budget*, not a grid. The slicer walks the document's own shape first:

1. the markdown heading tree (levels 1–6),
2. then blank-line paragraph breaks,
3. then sentence boundaries (`。 . ! ！ ? ？`),
4. and only as a last resort, a hard character cut.

Each paragraph carries a `title` field holding its **heading chain** (`"Part I  Chapter 3"`), so downstream chunks keep their section context.

### 3. Measurably fast - and honestly measured

Slicing a 550 KB structured document takes ~125 ms (≈4.4 MB/s) on a laptop CPU.
The hot path is a single linear scan that decides which heading levels a block can
possibly contain, so provably-empty regex passes are skipped; an optional C
extension (`pip install smart-slice[accel]`) runs that scan natively. Numbers and
the reasoning are in [`docs/PERFORMANCE.md`](docs/PERFORMANCE.md).

Speed is the KPI, and it is tracked release over release: **394 ms -> 124 ms
(~3.1x) from 0.1.0 to 0.5.0** on the same corpus, every step asserted
output-identical, with the batch scheduler adding a further 1.61x on a 48-document
/ 26.5 MB corpus. [`docs/BENCHMARK.md`](docs/BENCHMARK.md) has the per-change
breakdown, the peer throughput comparison, and what did *not* help.

Throughput is not the only axis, and on that one `smart-slice` is honestly beaten:
a plain character splitter runs ~36x faster. What the extra time buys is measured
against `chonkie` and `langchain-text-splitters` on the same corpus - `smart-slice`
is the only row that keeps every natural unit intact, carries its heading chain,
never cuts a code fence and never orphans a table row from its header. Retrieval
recall is deliberately **not** a KPI here: it needs an embedding model and a
labelled query set, so one archived run against the Chroma benchmark is recorded in
`docs/BENCHMARK.md` section 4 with its limitations stated, and is not repeated.

### 4. No framework

A pure library: no Django, no ORM, no network calls, no logging configuration (it emits records on the `smart_slice` logger and leaves routing to you). Errors are `SliceError` with an HTTP-style `.code` (`400` = input problem, `500` = parse failure). Images pulled out of a document are handed to a `save_image` **callback** — the library never persists anything, and a remote `<img src="https://...">` is recorded, never fetched.

### 5. Pictures are part of the document

A vision-language embedding model wants the figure next to the paragraph that
discusses it, not a text-only index of a document that was half diagrams.
`slice_multimodal` returns both, already joined — see
[Multimodal slicing](#multimodal-slicing-pictures--paragraphs).

---

## Installation

The core install is deliberately tiny (just a text decoder). Format parsers are optional extras, and a handler whose parser is missing simply reports "unsupported" instead of crashing.

```bash
pip install smart-slice                 # core: text/markdown/config/source files
pip install smart-slice[all]            # every format below
```

Published on PyPI as [`smart-slice`](https://pypi.org/project/smart-slice/). The
`0.4.0` wheel and sdist are also attached to the
[GitHub Release](https://github.com/guangxiangdebizi/smart-slice/releases/tag/v0.4.0),
built from the tagged commit by `.github/workflows/release.yml`.

Targeted extras:

| Extra | Unlocks |
|-------|---------|
| `markup` | HTML, MHTML, EPUB, EML, MSG, MOBI (markdownify + bs4) |
| `office` | DOCX, PPTX, PPT, XLSX, XLS, WPS, ET, RTF, ODF |
| `pdf` | PDF (pypdf + Pillow) |
| `mail` | MSG (extract-msg) |
| `archive` | 7z (zip/tar are stdlib) |
| `ebook` | MOBI/AZW |
| `image` | HEIC/HEIF images |
| `ocr` | local OCR for embedded images (RapidOCR, opt-in at runtime) |
| `keywords` | jieba-backed keyword helpers |
| `fastuuid` | time-sortable image ids (a pure-Python fallback is built in) |
| `test` | everything in `all` plus `xlwt`, which the test suite needs to *write* legacy `.xls` fixtures (`xlrd` only reads them) |
| `dev` | `test` + accel, pytest, pytest-cov, ruff, mypy |

Call `smart_slice.missing_dependencies()` at runtime to see exactly which extras would unlock more formats.

---

## Quickstart

### Slice a string

```python
from smart_slice import slice_text

rows = slice_text("# Guide\n\nIntro text.\n\n## Setup\n\nStep one.", limit=1000)
for r in rows:
    print(r["title"], "->", r["content"][:40])
# Guide -> Intro text.
# Guide  Setup -> Step one.
```

### Slice a file from disk

```python
from smart_slice import slice_path

rows = slice_path("report.pdf", limit=1000)        # name decides the handler
```

### Slice from bytes (with image capture)

```python
from smart_slice import slice_bytes, ImageAsset

images = []
def save_image(assets):           # assets: list[ImageAsset]
    images.extend(assets)

rows = slice_bytes(raw_bytes, "deck.pptx", limit=1000, save_image=save_image)
# each ImageAsset has .id, .file_name, .content (bytes); the paragraph text
# already references it as ![name](./oss/file/{id})
```

### Output shape

Every entry point returns a list of `{"title": str, "content": str}`:

```python
[
    {"title": "Part I  Chapter 3", "content": "the paragraph body..."},
    ...
]
```

`title` is the heading chain (empty when the document has no headings); `content` is the paragraph's verbatim text.

---

## Supported formats

30 handlers covering 197 declared extensions. Every parser except `charset-normalizer` (text decoding) is an **optional extra**: a handler whose parser is missing reports `400 Unsupported file format` instead of breaking `import smart_slice`.

| Category | Handler | Extensions | Parser |
|----------|---------|-----------|--------|
| Web | `HTMLSplitHandle` | `.html` `.htm` `.xhtml` `.shtml` | markdownify + bs4 |
| Web | `MhtmlSplitHandle` | `.mhtml` `.mht` | markdownify |
| Office | `DocSplitHandle` | `.docx` `.docm` `.doc` `.dotx` `.dotm` | python-docx |
| PDF | `PdfSplitHandle` | `.pdf` | pypdf + Pillow |
| Spreadsheet | `XlsxSplitHandle` | `.xlsx` `.xlsm` `.xltx` `.xltm` | openpyxl |
| Spreadsheet | `XlsSplitHandle` | `.xls` | xlrd |
| Spreadsheet | `CsvSplitHandle` | `.csv` `.tsv` `.tab` | stdlib csv |
| Archive | `ZipSplitHandle` | `.zip` | stdlib zipfile |
| Mind map | `XmindSplitHandle` | `.xmind` | stdlib zipfile + json |
| Office | `PptxSplitHandle` | `.pptx` `.pptm` `.ppsx` `.ppsm` `.potx` `.potm` | python-pptx |
| Office | `PptSplitHandle` | `.ppt` `.dps` | olefile / python-pptx |
| Office | `WpsSplitHandle` | `.wps` `.et` | python-docx / openpyxl |
| Office | `RtfSplitHandle` | `.rtf` | striprtf |
| Office | `OdfSplitHandle` | `.odt` `.ods` `.odp` `.fodt` `.fods` `.fodp` | stdlib zipfile + xml |
| Ebook | `EpubSplitHandle` | `.epub` | stdlib zipfile + bs4 |
| Mail | `EmlSplitHandle` | `.eml` | stdlib email + bs4 |
| Mail | `MsgSplitHandle` | `.msg` | extract-msg + bs4 |
| Ebook | `MobiSplitHandle` | `.mobi` `.azw` `.azw1` `.azw3` `.azw4` `.prc` | mobi + bs4 |
| Archive | `TarSplitHandle` | `.tar` `.tar.gz` `.tgz` `.tar.bz2` `.tar.xz` `.txz` | stdlib tarfile |
| Archive | `SevenZipSplitHandle` | `.7z` | py7zr |
| Image | `ImageSplitHandle` | `.jpg` `.jpeg` `.png` `.gif` `.bmp` `.tiff` `.tif` `.webp` `.heic` `.heif` | Pillow (+ pillow-heif) |
| Ebook | `Fb2SplitHandle` | `.fb2` `.fb2.zip` | stdlib zipfile + xml |
| Vector / structured | `SvgSplitHandle` | `.svg` `.svgz` | stdlib xml |
| Vector / structured | `IpynbSplitHandle` | `.ipynb` | stdlib json |
| Vector / structured | `SubtitleSplitHandle` | `.srt` `.vtt` `.ass` `.ssa` `.sub` | stdlib re |
| Mail | `MboxSplitHandle` | `.mbox` | stdlib mailbox |
| Vector / structured | `VcalendarSplitHandle` | `.ics` `.ifb` | stdlib |
| Vector / structured | `VcardSplitHandle` | `.vcf` | stdlib |
| Image | `ExtendedImageSplitHandle` | `.ico` `.cur` `.tga` `.pcx` `.dds` `.sgi` `.ppm` `.pgm` `.pbm` `.pnm` `.pfm` `.im` `.icns` `.qoi` `.jfif` `.jpe` `.apng` `.xbm` `.psd` (+ aliases) | Pillow |
| Text / source | `TextSplitHandle` | `.txt` `.md` `.markdown` `.log` `.json` `.jsonl` `.ndjson` `.yaml` `.yml` `.toml` `.ini` `.cfg` `.conf` `.rst` `.tex` `.sql` and 60+ source/config extensions (`.py` `.js` `.ts` `.java` `.go` `.c` `.cpp` `.rs` `.rb` `.php` `.cs` `.swift` `.kt` `.scala` `.lua` `.r` `.sh` `.ps1` `.css` `.scss` `.xml` `.proto` `.graphql` `.adoc` `.org` `.diff` `.patch` `.po` `.properties` `.env` ...) | charset-normalizer |

Archives are unpacked recursively; each inner file goes through the same dispatch. The text handler is the fallback for any decodable content whose extension is not otherwise claimed.

**Pictures** come out of every one of these containers - `docx`/`pptx`/`pdf`/`xlsx`/`zip` through their handlers, and `html`/`epub`/`eml`/`mhtml`/`odf`/`fb2`/standalone image files through the multimodal probe. See [Multimodal slicing](#multimodal-slicing-pictures--paragraphs).

### Not supported (rejected with `400`)

These are deliberately rejected rather than force-decoded into garbage:

| Category | Extensions | Why |
|----------|-----------|-----|
| Video | `.mp4` `.avi` `.mov` `.mkv` `.flv` `.wmv` `.webm` `.mpeg` `.mpg` `.3gp` `.rmvb` | no transcription pipeline |
| Audio | `.mp3` `.wav` `.flac` `.aac` `.ogg` `.m4a` `.wma` `.opus` `.alac` `.aiff` `.amr` | no ASR |
| Archive | `.rar` | no pure-Python decoder (rarfile needs the system `unrar` binary) |
| Apple iWork | `.key` `.pages` `.numbers` | proprietary IWA container |
| Legacy binary Office | old OLE `.wps` / `.et` / `.doc` | only the OOXML/zip variants are parsed; convert to `.docx` |
| Executable / image | `.exe` `.dll` `.msi` `.dmg` `.apk` `.iso` | not documents |
| Vector / raw photo | `.raw` | no universal decoder (vendor-private) |

**Content-level rejections** (right extension, unparsable content): encrypted PDFs and
image-only/scanned PDFs with no extractable text raise `500` naming the cause; encrypted or
oversized archives raise `ResourceLimitError`. Local OCR is **off by default**
(`SMART_SLICE_OCR_ENABLED=0`); enable it with the `ocr` extra to recover text from embedded
and scanned images.

---

## Multimodal slicing (pictures + paragraphs)

Text-only indexing throws away the figures, charts and screenshots in a document -
exactly the content a vision-language embedding model can use. `slice_multimodal`
returns the paragraphs *and* the pictures, already joined:

```python
from smart_slice import slice_multimodal

result = slice_multimodal(path="deck.pptx", limit=1000)

result.summary()          # '14 paragraphs, 6 images (6 attached, 0 unattached, 0.42 MB)'
result.coverage           # 1.0 - every image landed on a paragraph

for row in result.paired():            # only paragraphs that carry a picture
    print(row["title"], len(row["content"]), row["images"][0]["mime_type"])
    send_to_vl_model(row["title"], row["content"],
                     [i["data_uri"] for i in row["images"]])
```

Each entry of `result.records()` / `result.paired()` is plain JSON:

```python
{
  "index": 3,
  "title": "Q3 Results  Revenue by region",
  "content": "Revenue grew ... ![chart](./oss/file/0198f...)",
  "images": [{
      "id": "0198f...", "file_name": "chart1.png", "mime_type": "image/png",
      "bytes": 41233, "sha256": "9d1e...", "width": 940, "height": 512,
      "source": "zip", "member": "ppt/media/image3.png", "match": "reference",
      "data_uri": "data:image/png;base64,iVBORw0..."     # with with_data=True
  }]
}
```

### Where the pictures come from

Two independent channels, deduplicated against each other by content hash:

1. **the handlers** - `docx`, `pptx`, `pdf`, `xlsx` and `zip` already emit
   `ImageAsset` objects through `save_image`. `slice_multimodal` installs an
   `ImageCollector` as that callback, so they are captured instead of dropped.
2. **the probe** - when the handlers produced nothing for this document,
   `scan_media` opens the container itself. This is what recovers the formats that
   only ever yielded text:

| container | what the probe reads |
|-----------|----------------------|
| `zip` (OOXML / ODF / EPUB / any archive) | `word/media/*`, `ppt/media/*`, `xl/media/*`, `Pictures/*`, and every other image member; then text members for `<img src>` and `![]()` references, resolved relative to the member they were found in |
| `pdf` | `page.images`, tagged with the page number (needs the `pdf` extra) |
| `mime` (`.eml` `.mhtml`) | image parts, indexed by `Content-ID` / `Content-Location` so `cid:` references in the body resolve |
| `markup` (`.html` `.md` ...) | `data:` URIs decoded; remote URLs **recorded, never fetched** |
| standalone picture | the file itself, pinned to its one paragraph |

`probe="auto"` (the default) means "only when the handlers came up empty", so a
`.docx` is not read twice. `probe=True` always probes, `probe=False` never does.

### How a picture is joined to a paragraph

Four passes, strongest signal first; each placed image records how it matched in
`meta["match"]`, and anything left over lands in `result.unattached` rather than
being dropped:

| pass | signal | reliability |
|------|--------|-------------|
| `reference` | the `./oss/file/{id}` the handler wrote into the paragraph text | exact |
| `heading` | the heading the picture sat under in the source markup | strong |
| `anchor` | the ~96 characters of visible text that preceded the tag | heuristic |
| `document` | a standalone picture file *is* the document | exact by definition |

**The sliced text is never modified** - attachment is a join, not a rewrite, and
the test suite asserts `result.paragraphs == slice_bytes(...)` for every container.

### Using the collector on its own

```python
from smart_slice import slice_bytes, ImageCollector

collector = ImageCollector()                       # dedupes by sha256
rows = slice_bytes(data, "report.docx", save_image=collector)
collector.assets                                   # [ImageAsset, ...]
collector.duplicates                               # how many repeats it folded away
```

`ImageCollector` returns the `{new_id: kept_id}` map the `save_image` contract
expects, so duplicate pictures collapse onto one id *and* every
`./oss/file/{id}` reference in the paragraphs is rewritten to match.

`ImageAsset` grew computed views (the constructor is unchanged): `.content`,
`.size`, `.sha256`, `.mime_type`, `.suffix`, `.dimensions`, `.width`, `.height`,
`.data_uri()`, `.to_dict(with_data=...)`. MIME type and pixel size come from the
magic bytes and the file header - no Pillow required.

### Tuning

```python
slice_multimodal(
    data, "page.html",
    min_side=20,            # drop bullets, icons and tracking pixels
    include_vector=True,    # also return SVG / EMF / WMF (off: a VL model cannot eat them)
    include_external=False, # do not record remote <img src> URLs
    max_images=512,         # per-document ceiling
    max_image_bytes=64 << 20,
    keep_content=False,     # metadata only - frees the bytes after slicing
)
```

### A whole corpus at once

```python
from smart_slice import slice_many

report = slice_many(paths, backend="thread", concurrency=4, collect_images=True)
report.paragraphs        # unchanged: every document's paragraphs, in input order
report.images            # every picture in the batch, deduplicated across documents
```

`collect_images=True` needs a thread or serial backend - the collector holds a
lock, and a lock is not picklable - so `backend="process"` raises instead of
silently returning an empty list.

### On the command line

```bash
smart-slice media deck.pptx                  # what is in there, and where it belongs
smart-slice media page.html --json           # machine readable
smart-slice media book.epub --out-dir imgs/  # also write the picture bytes to disk
smart-slice media scan.pdf --min-side 20 --probe always
```

---
## Command line

```bash
smart-slice slice report.pdf --limit 1000 --format json -o out.json
smart-slice slice notes.md --overlap 150 --stats          # 15% overlap on a 1000-char budget
smart-slice slice notes.md --overlap-ratio 0.15 --format text
smart-slice chunk notes.md --chunk-size 256 --chunk-overlap 40 --carry-title
smart-slice detect mystery.bin
smart-slice formats            # list handlers + missing optional deps
smart-slice batch docs/*.pdf -j 4 --backend process --stats   # a whole corpus
smart-slice media deck.pptx --out-dir imgs/       # pictures out, and where they belong
smart-slice cores              # detected cores, derived width, affinity mask
python -m smart_slice slice notes.md --format md   # module form
```

- `slice` — `--format {json,jsonl,text,md}`, `--limit`, `--overlap`, `--overlap-ratio`, `--overlap-section-only`, `--no-filter`, `--title-prefix`, `-o/--output`, `--stats` (summary to stderr).
- `chunk` — slices first, then cuts paragraphs to an embedding window: `--chunk-size`, `--chunk-overlap`, `--carry-title`, `-o/--output`, `--stats`. Emits one JSON object per chunk.
- `batch` — slices many documents under one scheduling policy: `-j/--concurrency {N,auto,cores,serial,2x}`, `--backend {thread,process,serial}`, `--pin-cores`, `--max-concurrency`, `--error-policy {raise_first,collect}`, `--unordered`, `--timeout`, plus the `slice` options; `--format {json,jsonl,text,summary}`, `-o/--output`, `--stats`. Exit 1 names every failed document, exit 2 is a usage error.
- `media` — slices the document and reports every picture in it, joined to the paragraph it belongs to: `--limit`, `--probe {auto,always,never}`, `--include-vector`, `--min-side`, `--no-external`, `--out-dir` (write the image bytes to disk), `--json`.
- `cores` — the scheduling facts for this machine: usable cores, derived width, affinity mask, pinning support, and which environment variables are set (`--json` for machines).
- `detect` — prints the handler class that would claim the file (exit 2 if none).
- `formats` — supported extensions and which extras are not installed (`--json` for machines).

---

## Advanced usage

### Chunk size and overlap

Both stages default to sensible values and both are fully configurable — pass
nothing for defaults, or pass numbers/objects for control.

| Stage | Knob | Default | Customise |
|-------|------|---------|-----------|
| slicing (paragraphs) | `limit` | `1000` chars | `limit=800` |
| slicing (paragraphs) | `overlap` | `0` (none) | `overlap=150` or `overlap_ratio=0.15` |
| chunking (embeddings) | `chunk_size` | `256` chars | `chunk_size=512` |
| chunking (embeddings) | `chunk_overlap` | `0` (none) | `chunk_overlap=40` |

```python
from smart_slice import slice_bytes, chunk_paragraphs, ChunkingOptions

# defaults: 1000-char paragraphs, no overlap
rows = slice_bytes(data, "doc.md")

# custom: 800-char paragraphs carrying 15% of context forward
rows = slice_bytes(data, "doc.md", limit=800, overlap_ratio=0.15)

# reuse one configuration across many documents
opts = ChunkingOptions(limit=600, overlap=90, carry_title=True)
for name, blob in documents.items():
    rows = slice_bytes(blob, name, options=opts)

# second stage: embedding chunks with their own overlap
chunks = chunk_paragraphs(rows, chunk_size=256, chunk_overlap=40, carry_title=True)
```

**What overlap does to the output.** Paragraph *count* and *titles* are unchanged;
each paragraph simply gains the previous paragraph's tail:

```python
plain    = slice_bytes(data, "doc.md", limit=500)                # 42 paragraphs
overlapped = slice_bytes(data, "doc.md", limit=500, overlap=100) # 42 paragraphs
assert len(plain) == len(overlapped)
assert plain[0]["content"] in overlapped[0]["content"]   # additive, never lossy
```

With `overlap=0` (the default) chunks tile the source exactly once, so
concatenating them reproduces the original text. A positive overlap trades that
strict tiling property for cross-boundary retrieval recall — the right choice when
an answer spans a cut point, the wrong one when you need to reconstruct the
document from its chunks.

`ChunkingOptions` fields:

| Field | Default | Meaning |
|-------|---------|---------|
| `limit` | `1000` | maximum chunk size, measured with `length_fn` |
| `overlap` | `None` | characters of shared context; clamped to `limit // 2` |
| `overlap_ratio` | `None` | `overlap` as a fraction of `limit` (`0.15` = 15%) |
| `boundary` | `True` | snap cuts to sentence ends instead of cutting mid-sentence |
| `lookback` | `0.5` | fraction of `limit` searched backwards for a boundary |
| `overlap_boundary` | `True` | start carried context at a sentence/whitespace edge |
| `overlap_within_section` | `False` | only carry context inside one heading chain |
| `min_chunk` | `0` | merge chunks shorter than this into the previous one |
| `carry_title` | `False` | prefix **every** chunk with its heading chain |
| `length_fn` | `len` | size metric — pass a tokenizer to budget in tokens |

Token-based budgeting needs no extra plumbing:

```python
import tiktoken
enc = tiktoken.encoding_for_model("gpt-4o-mini")
opts = ChunkingOptions(limit=400, overlap=60, length_fn=lambda s: len(enc.encode(s)))
rows = slice_bytes(data, "doc.md", options=opts)   # 400-token paragraphs
```

### Custom heading patterns

```python
import re
from smart_slice import slice_bytes, MARKDOWN_HEADINGS, BLANK_LINE

# slice on a bespoke "SECTION n:" scheme, then blank lines
patterns = [re.compile(r"(?m)^SECTION \d+:.*"), *BLANK_LINE]
rows = slice_bytes(data, "spec.txt", limit=800, patterns=patterns)
```

### Raw (non-normalised) structure

`normalize=False` returns the handler's native shape — nested groups for multi-sheet workbooks and archives, which preview UIs want:

```python
raw = slice_bytes(data, "book.xlsx", limit=1000, normalize=False)
# [{"name": "Sheet1", "content": [...]}, {"name": "Sheet2", "content": [...]}]
```

### Image OCR hook

```python
from smart_slice import slice_bytes, build_image_text_extractor

ocr = build_image_text_extractor()   # None unless SMART_SLICE_OCR_ENABLED=1 and `ocr` extra installed
rows = slice_bytes(data, "scan.docx", limit=1000, image_text_extractor=ocr)
```

### Progress heartbeat

`progress_hook` is a zero-arg callback fired before each handler, each archive member, and each OCR call — wire it to a job heartbeat while parsing large files:

```python
slice_bytes(data, "huge.pdf", limit=1000, progress_hook=lambda: job.touch())
```

### Batch slicing and scheduling

One document is a `slice_path` call. A corpus is a scheduling problem, and the
scheduler that sat under the source platform's slicing path ships with the
package - every knob is a parameter, so you choose the policy per call:

```python
from smart_slice import slice_paths

report = slice_paths(["a.pdf", "b.docx", "c.md"], limit=1000, concurrency=4)
report.ok                 # True when every document sliced
report.results            # per-document paragraphs, in input order
report.paragraphs         # all of them concatenated, in input order
print(report.summary())   # 3/3 documents sliced in 412 ms (width=4, backend=thread, ...)
```

Width is a *variable*, not a constant. Accepted forms: an integer (`4`), or
`"auto"` (the default), `"cores"`, `"serial"`, or a multiple of the core count
(`"2x"`, `"0.5x"`):

```python
slice_paths(paths, concurrency="cores")     # every usable core
slice_paths(paths, concurrency=2)           # exactly two workers
slice_paths(paths, backend="serial")        # no pool at all
```

`"auto"` applies the allocation rule the scheduler was extracted from: 3 workers
on a machine with more than six cores, half the cores below that, never fewer
than one. `available_cores()` reports what the process may actually use - the
Linux affinity mask, so `taskset` and cgroup limits are honoured rather than the
host's core count:

```python
from smart_slice import available_cores, auto_concurrency
available_cores(), auto_concurrency()       # (12, 3) on a 12-core laptop
```

Core allocation is opt-in: `pin_cores=True` binds each task to one core of the
mask, round-robin, and restores the mask when the task ends (Linux
`sched_setaffinity`, Windows `SetThreadAffinityMask`, a logged no-op on macOS).
`TaskOutcome.core` records what each task actually got.

Batches can mix input shapes, and one policy object can be reused:

```python
from smart_slice import SchedulerPolicy, SliceJob, slice_many

policy = SchedulerPolicy(concurrency=4, pin_cores=True, error_policy="collect")
report = slice_many([
    "report.pdf",                                    # path
    ("upload.docx", uploaded_bytes),                  # (name, bytes)
    SliceJob.from_path("big.md", limit=4000),         # per-document override
    {"name": "notes.md", "content": raw},             # mapping
    upload_handle,                                    # anything with .name/.read()
], policy=policy, limit=1000)

for outcome in report.outcomes:
    if not outcome.ok:
        log.warning("%s failed: %s", outcome.name, outcome.error)   # code in .error.code
```

Failure semantics are the ported ones and they are a choice:

| `error_policy` | Behaviour |
|----------------|-----------|
| `"raise_first"` (default) | every document is still attempted, then the error of the **lowest failing index** is re-raised as the original object (traceback intact) and the remaining failures are logged |
| `"collect"` | nothing is raised; inspect `report.failures` / `outcome.error` |

Two backends, and the choice matters more than the width:

| Backend | Use it when | Cost |
|---------|-------------|------|
| `"thread"` (default) | the worker also does I/O (database writes, embedding calls, downloads), or you pass callbacks (`save_image`, `progress_hook`, a tokenizer in `length_fn`) | none; but the GIL means pure-Python slicing does not speed up |
| `"process"` | a CPU-bound corpus of plain files, big enough to amortise startup | interpreter startup plus pickling the bytes both ways; inputs must be picklable |

Measured on a 12-core laptop (551 KB structured markdown per document, CPython
3.12, Windows) - `python scripts/benchmark.py --batch`:

| Batch | serial | threads x4 | processes x4 |
|-------|--------|------------|--------------|
| 12 documents (6.6 MB) | 1.00x | 1.07x | 0.69x |
| 48 documents (26.5 MB) | 1.00x | 0.95x | **1.61x** |

So: threads buy you overlap only where the GIL is released; processes buy real
throughput on CPU-bound batches from a few dozen documents up. Details and the
reasoning are in [`docs/PERFORMANCE.md`](docs/PERFORMANCE.md).

Underneath, `run_parallel(items, worker, policy)` is the same ordered,
error-isolating parallel map without any slicing in it - use it when you want
the scheduling for your own worker.

### Second-stage chunking for embeddings

Slicing yields semantic paragraphs; embedding models still need fixed-size input.
`chunk_paragraphs` cuts paragraphs to a window while the paragraph stays the
retrieval unit. See [Chunk size and overlap](#chunk-size-and-overlap) for the
overlap and `carry_title` options.

### Error handling

```python
from smart_slice import slice_bytes, SliceError

try:
    slice_bytes(data, "movie.mp4")
except SliceError as e:
    print(e.code, e.message)   # 400 Unsupported file format: .mp4
```

---

## Configuration

Defensive parser caps (anti decompression-bomb guards) come from the environment, canonical name `SMART_SLICE_PARSER_<FIELD>`:

| Variable | Default | Meaning |
|----------|---------|---------|
| `SMART_SLICE_PARSER_MAX_INPUT_BYTES` | 134217728 | max input size |
| `SMART_SLICE_PARSER_MAX_XML_BYTES` | 33554432 | max XML member size |
| `SMART_SLICE_PARSER_MAX_PACKAGE_MEMBERS` | 10000 | max archive entries |
| `SMART_SLICE_PARSER_MAX_PACKAGE_BYTES` | 134217728 | max expanded archive size |
| `SMART_SLICE_PARSER_MAX_XML_ELEMENTS` | 100000 | max XML nodes |
| `SMART_SLICE_PARSER_MAX_XML_DEPTH` | 128 | max XML nesting |
| `SMART_SLICE_PARSER_MAX_TABLE_CELLS` | 1000000 | max spreadsheet cells |
| `SMART_SLICE_PARSER_MAX_MIME_PARTS` | 1000 | max MIME parts |
| `SMART_SLICE_PARSER_MAX_MIME_DEPTH` | 32 | max MIME nesting |
| `SMART_SLICE_PARSER_MAX_ODF_TEXT_BYTES` | 33554432 | max ODF extracted text |
| `SMART_SLICE_OCR_ENABLED` | `0` | enable local OCR (needs the `ocr` extra) |

Batch scheduling is configured the same way, and every variable is overridden by
the matching argument:

| Variable | Default | Meaning |
|----------|---------|---------|
| `SMART_SLICE_CONCURRENCY` | `auto` | batch width: an integer, `auto`, `cores`, `serial`, or a multiple (`2x`) |
| `SMART_SLICE_MAX_CONCURRENCY` | `32` | ceiling for a derived width |
| `SMART_SLICE_SCHEDULER_BACKEND` | `thread` | `thread`, `process` or `serial` |
| `SMART_SLICE_PIN_CORES` | `0` | allocate one core per task, round-robin |

Or override per call by passing a `ParserLimits` instance to the validators, or a
`SchedulerPolicy` to `slice_many` / `run_parallel`.

---

## API reference

| Function | Purpose |
|----------|---------|
| `slice_text(text, *, limit, patterns, with_filter, name)` | slice an in-memory string |
| `slice_bytes(content, name, *, limit, ...)` | slice raw bytes (full options) |
| `slice_path(path, *, limit, ...)` | read a file, then slice it |
| `extract_text(content, name, *, save_image)` | extract text only, no slicing |
| `detect_handler(name, content=None)` | which handler claims this input |
| `supported_extensions()` | handler → extensions map |
| `missing_dependencies()` | optional extras not installed |
| `chunk_paragraphs(paragraphs, *, chunk_size)` | second-stage embedding chunking |
| `chunk(text, *, chunk_size)` | convenience wrapper |
| `slice_many(inputs, *, concurrency, backend, pin_cores, collect_images, ...)` | slice a corpus in one call -> `BatchReport` |
| `slice_paths(paths, ...)` | `slice_many` over filesystem paths |
| `slice_multimodal(content=None, name=None, *, path, limit, probe, ...)` | paragraphs **and** pictures, already joined -> `MultimodalResult` |
| `slice_path_multimodal(path, ...)` | `slice_multimodal` for a file on disk |
| `extract_media(content, name, ...)` / `scan_media(...)` | pull the images out of a container, no slicing |
| `attach_images(paragraphs, assets)` | join images to paragraphs |
| `ImageCollector` | a `save_image` sink that dedupes and returns the id-remap contract |
| `MultimodalResult` / `MultimodalParagraph` / `MediaExtraction` | multimodal result types |
| `container_kind(content, name)` | which reader the probe would use |
| `run_parallel(items, worker, policy)` | the ordered, error-isolating parallel map |
| `SchedulerPolicy` / `SliceJob` / `BatchReport` / `TaskOutcome` | scheduling configuration and results |
| `available_cores()` / `auto_concurrency()` / `resolve_concurrency()` | core discovery and width derivation |
| `use_policy(policy)` | publish a policy for a block (contextvar-scoped) |
| `split_document(...)` | the low-level orchestration entry point |

Building blocks: `SplitModel`, `smart_split_paragraph`, `filter_special_char`, `MarkChunkHandle`, `ParserLimits`, `ImageAsset` (with `.mime_type`, `.sha256`, `.dimensions`, `.data_uri()`, `.to_dict()`), `SPLIT_HANDLERS`, `MARKDOWN_HEADINGS`, `DEFAULT_PATTERNS`, `BLANK_LINE`, `LITERAL_PATTERNS`.

See [`docs/ARCHITECTURE.md`](docs/ARCHITECTURE.md) for how slicing works internally, [`docs/BENCHMARK.md`](docs/BENCHMARK.md) for the comparison against other chunkers, and [`docs/PORTING.md`](docs/PORTING.md) for how this was extracted from the source platform.

---

## Development

```bash
git clone https://github.com/guangxiangdebizi/smart-slice.git
cd smart-slice
pip install -e ".[dev]"
pytest                                   # 415 tests, all offline
ruff check smart_slice tests

python scripts/benchmark.py              # throughput + profile + batch scaling
pip install chonkie langchain-text-splitters tiktoken
python scripts/benchmark_peers.py        # peer throughput + structure comparison
```

---

## License

GPL-3.0. See [LICENSE](LICENSE).
