Switch to magika+onxx instead of guesslang+tensorflow

[robherley]

Closes #245

We have pretty dramatic savings by ditching tensorflow (like 400 MB), but it looks like the guesslang model is faster.

Guesslang itself is a smaller, less accurate model and only supports ~54 languages with a stated 90% accuracy. Whereas Magika supports > 200 content types and states 99% accuracy.

IMO the difference in speed isn't dramatic enough to stick with guesslang.

---

Note, the data collected below was investigated and summarized by Claude.

Size

Image Size Comparison Summary

Architecture	Remote (TensorFlow)	Local (Magika)	Savings
amd64	529.28 MB	128.54 MB	400.74 MB (75.7%)
arm64	104.88 MB	115.24 MB	-10.36 MB (+9.9%)

Key Findings:

amd64 (x86_64):

• The old TensorFlow-based image had a massive 436 MB layer for TensorFlow libs
• The new Magika image uses only 22.3 MB for ONNX Runtime libs
• Total savings: ~401 MB (76% reduction)

arm64:

• The remote arm64 image didn't actually include TensorFlow (0B extra libs layer) - likely due to lack of TensorFlow arm64 support at the time
• The new local arm64 image is slightly larger (+10 MB) because it now includes ONNX Runtime (18.9 MB) where previously there was nothing
• This is actually a feature improvement - arm64 now has full ML inference support via Magika/ONNX

Performance

Initialization Time

Time to load the ML model and prepare for inference.

Library	Time	Memory	Allocations
Magika	3.29 ms	177 KB	2,745
Guesslang	34.08 ms	N/A*	N/A*

*Guesslang uses TensorFlow which manages memory internally.

Takeaway: Magika initializes ~10x faster, making it better for CLI tools or short-lived processes.

---

Average time to detect the language of a single file (after initialization).

Library	Avg Time	Throughput	Memory/op	Allocs/op
Magika	2.02 ms	0.58 MB/s	~21 KB	12
Guesslang	0.27 ms	4.30 MB/s	~6.5 KB	141

Per-Language Breakdown

Language	Magika (ns/op)	Guesslang (ns/op)	Guesslang Speedup
Go	1,995,021	235,510	8.5x
Python	1,993,068	252,971	7.9x
JavaScript	2,001,053	259,589	7.7x
Rust	2,080,715	265,405	7.8x
Java	2,051,423	288,144	7.1x
TypeScript	2,034,903	265,808	7.7x
Ruby	2,039,749	259,497	7.9x
C++	2,032,912	278,796	7.3x
C	2,022,978	276,233	7.3x
PHP	2,049,994	287,743	7.1x

Takeaway: Guesslang is ~7.5x faster for per-file detection.

[Copilot]

Pull request overview

This pull request switches the file type detection system from guesslang (based on TensorFlow) to magika (based on ONNX Runtime), enabling better cross-platform support including ARM64 architectures.

Changes:

• Replaces TensorFlow/guesslang-go dependencies with ONNX Runtime/magika-go
• Adds vendor-onnxruntime script to download and install ONNX Runtime binaries
• Updates build system with new env and build scripts for CGO/linker configuration
• Removes architecture-specific limitations (previously disabled on ARM64)

Reviewed changes

Copilot reviewed 18 out of 20 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
script/vendor-onnxruntime	New script to download and install ONNX Runtime for the platform
script/run	New script to run the application with proper environment setup
script/env	New script to configure CGO/linker flags for ONNX Runtime
script/build	New script to build the binary with cross-compilation support
script/install-libtensorflow	Removed TensorFlow installation script
internal/renderer/guess.go	Replaced guesslang with magika scanner implementation using lazy initialization
internal/renderer/guess_disabled.go	Updated build tags to remove arm64 restriction
internal/renderer/detect.go	Updated comments to reflect AI guessing instead of Guesslang
internal/config/config_guesser.go	Changed build tag from amd64 to cgo
internal/config/config_noguesser.go	Changed build tag from arm64 to !cgo
internal/config/config.go	Updated description to reference AI model instead of Guesslang
go.mod	Replaced guesslang-go with magika-go dependency
go.sum	Updated dependency checksums
docs/self-hosting.md	Updated documentation for new ONNX-based approach
docs/contributing.md	Updated setup instructions for local development
README.md	Updated technology credits
Dockerfile	Refactored to use ONNX Runtime instead of TensorFlow
.gitignore	Added third_party/ directory
.github/workflows/test.yml	Updated to use vendor-onnxruntime instead of install-libtensorflow
.github/workflows/lint.yml	Updated to use vendor-onnxruntime instead of install-libtensorflow

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.