anydoc

README

go-anydoc

English | 简体中文

⚠️ Experimental branch

This branch replaces wazero with wazy, a pure-Go runtime descended from it that is considerably faster on this workload. Do not use it in production. wazy has no stable release: its published versions are retracted betas plus a pseudo-version, it has one author, and it states that it makes no stability promise.

It exists so the difference can be measured rather than argued about, on the same 7.5 MB PDF and the same anydoc.wasm:

converting	main (wazero)	this branch (wazy)
1 KB docx, interpreted	3.5 ms	2.7 ms	1.3×
1 KB docx, compiled	0.4 ms	0.62 ms	0.6×
5 MB docx body, interpreted	11.1 s	8.6 s	1.3×
5 MB docx body, compiled	0.86 s	0.19 s	4.5×
7.5 MB PDF, interpreted	41.4 s	32.6 s	1.3×
7.5 MB PDF, compiled	3.4 s	0.75 s	4.5×

The win is on long compute; tiny documents are a wash or slightly worse, where per-call overhead dominates. Output is byte-identical and the full test suite passes, including cross-compilation to riscv64, ppc64le, 386 and s390x.

The API, the embedded module and the exit-code ABI are unchanged; only the runtime differs. main stays on wazero so that no downstream user has a beta dependency forced into their module graph.

This branch is tagged v0.1.4-experiment.1. Take it with a replace, not a require:

go mod edit -replace github.com/xusenlin/go-anydoc=github.com/xusenlin/go-anydoc@v0.1.4-experiment.1
go mod tidy

Same import path, same API.

It has to be a replace because a pre-release sorts below the release it is based on: a require here loses to main's next release, and go get -u swaps the runtime back with no signal. go get never rewrites a replace, so the pin holds on its own and the tag never needs re-cutting.

Use replace, because no version number can hold this branch. Go turns the commit into a pseudo-version based on the nearest tag, which carries a pre-release suffix — and a pre-release always sorts below the release it is based on. So however this branch is tagged, a require on it loses to main's next release and go get -u swaps the runtime back with no signal. A replace is the one thing go get will not rewrite: it may bump the require line above it, and the build still uses this branch.

That holds because a replace applies in the main module. If you are a library rather than an application, you cannot make this choice for your dependents — stay on main.

Go bindings for anydoc — convert Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV and text-based PDF to Markdown.

No cgo. No subprocesses. No system dependencies. The Rust library is compiled to WebAssembly and embedded, so a binary using this package is self-contained and cross-compiles wherever Go does:

CGO_ENABLED=0 GOOS=windows GOARCH=arm64 go build ./...

Install

go get github.com/xusenlin/go-anydoc

Requires Go 1.25 or newer, which is what the runtime requires.

Usage

c, err := anydoc.New()
if err != nil {
    return err
}
defer c.Close(ctx)

md, err := c.Convert(ctx, docBytes, "docx")

Pass "" as the hint to detect the format from content. CSV has no signature and must be named explicitly.

Errors are matched with errors.Is:

switch {
case errors.Is(err, anydoc.ErrEncrypted):    // password-protected
case errors.Is(err, anydoc.ErrUnsupported):  // not a format anydoc parses
case errors.Is(err, anydoc.ErrMalformed):    // recognised but corrupt
}

Design notes

Interpreter, not the optimising compiler. wazy can either translate the module to native machine code up front (fast to run, expensive to load, amd64 and arm64 only) or interpret it instruction by instruction (cheap to load, portable everywhere, slower to run). This package interprets by default, because a library that ships inside other people's binaries cannot assume it may write a compilation cache somewhere, cannot assume the host allows the executable pages the compiler needs — macOS hardened runtime and some seccomp profiles refuse them — and cannot assume the target is amd64 or arm64. WithCompiler opts into the other side of that trade.

An application does know where its own data lives, and that changes the arithmetic: WithCompilationCache makes the compiler's cost a one-time 2.4 s and 622 MB instead of a per-start one, and every start after that is 6.6 ms at 36 MB — cheaper than interpreting, and an order of magnitude faster to convert with. The defaults below assume no cache, because a library cannot assume one.

The trade is real and it scales with document size, so measure against your own corpus before assuming it is free:

	interpreter (default)	compiler	compiler + warm cache
New() — once per process	~86 ms	~2.4 s	6.6 ms
1 KB docx	2.7 ms	0.62 ms	0.62 ms
docx with a 5 MB uncompressed body	8.6 s	0.19 s	0.19 s
7.5 MB PDF	32.6 s	0.75 s	0.75 s
RSS after New()	122 MB	622 MB	36 MB

The third column is the second one after WithCompilationCache has a directory to read from — same execution, none of the startup. Conversion figures are identical because the cache changes how the machine code is obtained, not what it is. Only the first run on a machine pays the second column.

Ordinary office documents are in the second row's territory and cost nothing worth optimising. Multi-megabyte ones are ~12× slower than they would be compiled, so if you convert those, either bound the tail with WithMaxInputBytes and a context deadline — cancellation interrupts the guest mid-conversion — or opt into WithCompiler.

The slowness is the runtime, not wasm. It is tempting to read the numbers above as the cost of compiling to WebAssembly. Measured on the same PDF and the same anydoc.wasm, it is not — the format costs almost nothing, and nearly all of the gap is how a particular runtime turns wasm into machine code:

running the same 7.5 MB PDF		vs native
native Rust binary	0.28 s	1×
the same wasm, wasmtime with Cranelift	0.37 s	1.3×
the same wasm, wasmtime with Winch, a deliberately baseline compiler	0.90 s	3.2×
the same wasm, wazy with WithCompiler (this branch)	0.75 s	2.7×
the same wasm, wazero with WithCompiler (main)	3.4 s	12×
the same wasm, wazy interpreted	33 s	118×

So the things usually blamed — 32-bit pointers, bounds-checked linear memory, no native SIMD — account for the 1.3×. Everything beyond that is code generation — and, as the wazy row shows, a pure-Go runtime can close most of it. wazero's compiler is fast, simple and portable rather than optimising; wazy descends from it and spends the effort on the memory-access paths that dominate this workload.

That is the trade, stated plainly: wasmtime's Go bindings are cgo, so adopting them would cost CGO_ENABLED=0, free cross-compilation, and the single self-contained binary. If your workload values native speed above those, a cgo binding is the honest answer and this package is the wrong tool.

_{Measured on Apple M5 Pro, macOS 26.5, Go 1.26.1, wazy (pseudo-version cd2607360a17), against anydoc.wasm 6,542,355 bytes (anydoc 0.1.7). The 5 MB figure is a 25,000-row table: 42 KB zipped, since the size that matters is the uncompressed body.}

A fresh guest per document. Each Convert instantiates its own linear memory, so a large document cannot leave memory permanently claimed and a malformed one cannot leak state into the next call. Compilation, the expensive part, happens once in New.

A command module, not exported functions. The guest reads stdin and writes stdout, so the Go side never touches linear memory, pointers, or UTF-8 boundaries. The exit code carries the error kind. See rust/src/main.rs — the code table there and in errors.go is a two-language contract.

Sandboxed. WASI is instantiated for stdio and random_get (which HashMap seeding reaches for). No filesystem, no clock, no sockets.

Controlling size

The module is embedded by default so go get and New() just work. Builds that would rather ship it out of band:

go build -tags anydoc_nowasm    # 6.59 MB smaller

embeddedWASM is then nil and New requires WithWASM(r) or WithWASMBytes(b). Useful for container layering, serverless size limits, pinning a different anydoc build, or environments that forbid opaque embedded blobs.

The saving is the 6.54 MB module plus ~47 KB of embed machinery. For scale, examples/convert is 14.1 MB built normally and 7.6 MB with the tag.

The build tag affects the compiled binary, not go get: the module is in the Go module either way.

Tuning

anydoc.New(
    anydoc.WithConcurrency(4),        // simultaneous guests; the main memory lever
    anydoc.WithMemoryLimitPages(1024), // 64 MiB per guest
    anydoc.WithMaxInputBytes(64<<20),
    anydoc.WithCompiler(),            // throughput over startup cost; see below
    anydoc.WithCompilationCache(dir), // and pay that cost only once; see below
)

WithCompiler

Compiles the module to native code instead of interpreting it, turning the table above from the left column into the right one. The cost is paid once, in New — Convert only instantiates the already-compiled module — so it pays off in a long-lived process that reuses one Converter, and is a pure loss in a short-lived one that converts a single small document and exits (~2.1 s bought to save ~4 ms).

It is a request, not a guarantee. The backend needs amd64 or arm64 on a mainstream OS, plus a host that permits mmap'd executable pages. Where that does not hold — riscv64, ppc64le, 386, macOS hardened runtime, some seccomp profiles — wazy falls back to the interpreter silently and the conversion still happens, just at interpreter speed. Cross-compilation is unaffected either way: wazy is pure Go, and this option changes no build constraints.

WithCompilationCache

Persists the compiler's output under a directory you own, so WithCompiler costs what it costs once per machine rather than once per process:

New() with WithCompiler	time	peak RSS
cold — compiling	2.4 s	622 MB
warm — reading the result back	6.6 ms	36 MB

That gap is the whole argument against WithCompiler disappearing. The memory it is expensive for is the compiler working, not the compiled module sitting there; a hit loads machine code instead of producing it. The directory holds about 14 MB.

This only affects WithCompiler. The interpreter emits no machine code, so there is nothing to persist and the option does nothing — the directory stays empty, and the test suite asserts it.

An entry is keyed by the module, the CPU's feature bits, the runtime version and the target platform. Anything else is a miss, and a miss just compiles again and writes a new entry; it will never hand a host machine code it cannot run. So the directory is disposable — deleting it costs one recompilation — and it holds machine code, never anything a caller passed to Convert.

Two things follow for anyone shipping this. Give it somewhere that survives a restart, or the point is lost. And prefer to let it fill at runtime rather than baking it into an image: the CPU feature bits are in the key, and a CI builder rarely shares them with wherever the image ends up.

WithMemoryLimitPages is validated against the module's declared minimum at New time, not at conversion time — setting it too low fails fast with a clear message rather than surfacing later as a mysterious conversion error.

How much a guest actually needs tracks the uncompressed content, not the size of the file on disk — a zip bomb is small on disk and large in memory, which is why the limit exists:

document	pages needed
module minimum, converts nothing	64 (4 MiB)
docx, 0.4 MB body	192 (12 MiB)
docx, 2 MB body	576 (36 MiB)
PDF, 7.5 MB file	512–1024 (32–64 MiB)
docx, 5 MB body	1280 (80 MiB)

The default of 1024 pages converts an ordinary 7.5 MB PDF. Raise it if conversions fail on inputs that convert fine elsewhere — the error says so explicitly when the guest ran out of memory, and names this option. Each concurrent guest can claim up to this much, so WithConcurrency multiplied by this value is your real ceiling.

Development

Tasks are run with Task; task --list shows them all.

task test      # builds the wasip1 test stub, runs the harness suite
task verify    # everything CI runs, and the checklist before tagging
task wasm      # rebuilds the embedded module; needs Rust 1.88 + wasm-opt

There are two suites, split by build tag:

• anydoc_test.go runs under -tags anydoc_nowasm against a stub built with Go's own wasip1 target. The stub speaks the same ABI without converting anything, so the harness — stdio wiring, exit-code mapping, concurrency, cancellation, memory containment — is testable with no Rust toolchain in the loop. See testdata/README.md.
• anydoc_real_test.go is tagged !anydoc_nowasm, so it compiles exactly when anydoc.wasm is in the tree, and covers what the harness is for: real conversion output, format detection, and the error codes the crate actually emits.

anydoc.wasm is a build artifact committed to the repo, because Go modules have no build step: whatever is committed is what go get delivers. It is built from the pinned crate by task wasm, run by hand rather than by CI, and anydoc.wasm.sha256 records the checksum of the committed copy. See anydoc.wasm.README for the toolchain versions needed to reproduce it.

Versioning

This module's version is independent of the crate's. The crate is pinned with = in rust/Cargo.toml, and moving to a new upstream release is a deliberate act: edit the pin, run task wasm, review what changed in the conversion output, then tag. An upstream release on its own changes nothing here.

anydoc.EmbeddedAnydocVersion reports which crate version the embedded module was built from, and anydoc.Info() prints it with the payload size. It is copied from rust/Cargo.lock by task sync-version, and task check-version fails the build if the two ever disagree — so it cannot quietly misreport what is actually embedded.

rust/Cargo.lock is committed too. The = pin only fixes anydoc itself; the lockfile is what stops a transitive dependency from silently changing the conversion output between one rebuild and the next.

License

MIT for this package — see LICENSE. The embedded module is built from anydoc, also MIT — see LICENSE-anydoc.

Documentation

Overview

Package anydoc converts office documents to Markdown by running the anydoc Rust library as WebAssembly, with no cgo and no external processes.

The wasm module is embedded, so a binary built with this package is self-contained and cross-compiles anywhere Go does:

CGO_ENABLED=0 GOOS=windows GOARCH=arm64 go build ./...

Basic use:

c, err := anydoc.New()
defer c.Close(ctx)
md, err := c.Convert(ctx, docBytes, "docx")

Index

• Constants
• Variables
• func Info() string
• type ConvertError
  • func (e *ConvertError) Error() string
  • func (e *ConvertError) Unwrap() error
• type Converter
  • func New(opts ...Option) (*Converter, error)
  • func (c *Converter) Close(ctx context.Context) error
  • func (c *Converter) Convert(ctx context.Context, doc []byte, hint string) (string, error)
  • func (c *Converter) ConvertReader(ctx context.Context, r io.Reader, hint string) (string, error)
• type Option
  • func WithCompilationCache(dirname string) Option
  • func WithCompiler() Option
  • func WithConcurrency(n int) Option
  • func WithMaxInputBytes(n int) Option
  • func WithMemoryLimitPages(pages uint32) Option
  • func WithWASM(r io.Reader) Option
  • func WithWASMBytes(b []byte) Option

Constants

const EmbeddedAnydocVersion = "0.1.7"

EmbeddedAnydocVersion is the upstream crate version this wasm was built from. Copied from rust/Cargo.lock by `task sync-version`; `task check-version` fails the build if the two drift apart.

Variables

var (
	// ErrUnsupported means the input was not a format anydoc can parse, either
	// because the content signature matched nothing or the hint named a format
	// that does not exist.
	ErrUnsupported = errors.New("anydoc: unsupported format")

	// ErrMalformed means the format was recognised but the bytes were corrupt.
	ErrMalformed = errors.New("anydoc: malformed document")

	// ErrEncrypted means the document is password-protected. anydoc does not
	// decrypt; obtain a decrypted copy first.
	ErrEncrypted = errors.New("anydoc: document is encrypted")

	// ErrResourceLimit means a guard inside anydoc tripped, e.g. a zip bomb or
	// an implausible declared size. Distinct from the host-side memory limit,
	// which surfaces as ErrWASM.
	ErrResourceLimit = errors.New("anydoc: resource limit exceeded")

	// ErrMissingPart means a required part of a container format was absent,
	// e.g. a .docx with no word/document.xml.
	ErrMissingPart = errors.New("anydoc: missing required part")

	// ErrWASM means the guest trapped, ran out of memory, or otherwise failed
	// in a way that is not a document problem. Treat as an internal error.
	ErrWASM = errors.New("anydoc: wasm execution failed")

	// ErrNoWASM means no module was available: the build used the
	// anydoc_nowasm tag and no WithWASM option was supplied.
	ErrNoWASM = errors.New("anydoc: no wasm module (build without anydoc_nowasm, or pass WithWASM)")
)

Sentinel errors returned by Convert. Match them with errors.Is.

The numeric codes behind these live in rust/src/main.rs and are part of the module ABI: a wasm built from a different revision of that file may not agree with this table. That is what EmbeddedAnydocVersion is for.

Functions

func Info

func Info() string

Info reports what this build is carrying, for diagnostics and version reporting in host applications.

Types

type ConvertError

type ConvertError struct {
	Kind   error  // one of the sentinels above
	Detail string // guest stderr, may be empty
	Code   int    // raw exit code, for debugging
}

ConvertError carries the guest's stderr alongside a sentinel, so callers can branch on the class of failure and still surface anydoc's own message.

func (*ConvertError) Error

func (e *ConvertError) Error() string

func (*ConvertError) Unwrap

func (e *ConvertError) Unwrap() error

type Converter

type Converter struct {
	// contains filtered or unexported fields
}

Converter runs conversions. It is safe for concurrent use and should be created once and reused: New compiles the module, which is the expensive part, and Convert then instantiates a cheap short-lived guest per document.

func New

func New(opts ...Option) (*Converter, error)

New compiles the wasm module and prepares the runtime.

Compilation happens once, here, and costs a few hundred milliseconds. Do it at startup, not per request.

func (*Converter) Close

func (c *Converter) Close(ctx context.Context) error

Close releases the runtime, and the compilation cache when there is one. The Converter is unusable afterwards. Cached entries stay on disk: that is the point of them.

func (*Converter) Convert

func (c *Converter) Convert(ctx context.Context, doc []byte, hint string) (string, error)

Convert turns a document into Markdown.

hint is a bare extension such as "docx" that selects the parser. Pass "" to detect from content, which works for every format except CSV, which has no signature and must be named.

Each call gets a fresh guest instance with its own linear memory. A large document therefore cannot leave memory permanently claimed, and a malformed one cannot leak state into the next call.

func (*Converter) ConvertReader

func (c *Converter) ConvertReader(ctx context.Context, r io.Reader, hint string) (string, error)

ConvertReader is a convenience wrapper that reads r fully before converting. The whole document has to be in memory regardless: it is copied into the guest's linear memory.

type Option

type Option func(*config)

Option configures a Converter.

func WithCompilationCache

func WithCompilationCache(dirname string) Option

WithCompilationCache persists the compiler's output under dirname, so the cost of WithCompiler is paid once per machine instead of once per process.

It changes what WithCompiler costs more than it changes what it does. Measured on the real module, a hit turns New from 2.8s and 647 MB of peak RSS into 34ms and 50 MB, because the machine code is read back rather than produced -- the memory WithCompiler is expensive for is the compiler working, not the compiled module sitting there. The directory holds ~23 MB.

This only affects WithCompiler. The interpreter emits no machine code, so there is nothing to persist and this option does nothing.

An entry is keyed by the module, the CPU's feature bits, the runtime version and the target platform. Anything else is a miss, and a miss simply compiles again and writes a new entry -- it will not hand the host machine code it cannot run. That also makes the directory disposable: deleting it costs one recompilation and nothing else, and what it holds is machine code, never anything a caller passed to Convert.

Give it a directory the process owns and that survives restarts; a temporary directory wastes the point. It is created if missing, and Close releases it.

func WithCompiler

func WithCompiler() Option

WithCompiler trades startup cost and memory for throughput, by having wazy compile the module to native code instead of interpreting it.

Measured on the real module, per document: a 1 KB docx goes from 3.5ms to 0.4ms, one with a 5 MB uncompressed body from 11.1s to 0.86s, and a 7.5 MB PDF from 41s to 3.4s. The price is paid once in New, which goes from ~100ms and 182 MB of RSS to ~2.7s and 638 MB. That 638 MB is an OOM kill in a 512 MB container, which is why this is opt-in rather than the default.

The compilation is paid once, in New, not per document: Convert only instantiates the already-compiled module. So this pays off in a long-lived process that reuses one Converter, and is a pure loss in a short-lived one that converts a single small document and exits -- there it buys ~3ms for ~2.7s. Leave it off for tight containers too, unless you can also give it WithCompilationCache, which removes both of those objections after the first run.

This is a request, not a guarantee. The compiler backend needs amd64 or arm64 (with SSE4.1 on amd64) on a mainstream OS, and needs the host to allow mmap'd executable pages -- macOS hardened runtime and some seccomp profiles refuse them. Where it is unavailable, wazy falls back to the interpreter silently: the conversion still happens, at interpreter speed. Nothing here affects cross-compilation, since wazy is pure Go.

func WithConcurrency

func WithConcurrency(n int) Option

WithConcurrency caps simultaneous guest instances. Each holds its own linear memory, so this is the main lever on peak memory. Default 4.

func WithMaxInputBytes

func WithMaxInputBytes(n int) Option

WithMaxInputBytes rejects documents larger than n before any wasm runs. Zero disables the check. Default 64 MiB.

func WithMemoryLimitPages

func WithMemoryLimitPages(pages uint32) Option

WithMemoryLimitPages caps one guest's linear memory in 64 KiB pages. Default 1024 (64 MiB).

What a guest needs tracks the uncompressed content rather than the input's size on disk, so a compact file can still be expensive. Measured against the real module: 64 pages is the module's own floor and converts nothing, a docx with a 0.4 MB body needs 192, a 2 MB body 576, a 5 MB body 1280, and a 7.5 MB PDF needs between 512 and 1024.

Raise this when conversions fail with ErrWASM on inputs that convert fine elsewhere; the error says so explicitly when the guest ran out of memory. Each concurrent guest can claim up to this much, so this multiplied by WithConcurrency is the real ceiling.

Validated at New time against the module's declared minimum, so too low a value fails immediately rather than at conversion.

func WithWASM

func WithWASM(r io.Reader) Option

WithWASM supplies the module instead of using the embedded one.

This is the counterpart to building with -tags anydoc_nowasm, where nothing is embedded and this option becomes mandatory. It is also useful without the tag: pinning a different anydoc build, loading from a shared volume, or keeping the payload in its own container layer.

The module is read immediately; r is not retained.

func WithWASMBytes

func WithWASMBytes(b []byte) Option

WithWASMBytes is WithWASM for a module already in memory.