Encoder path, quantized router index, and a catalog SPI - #111

Merged

bsbodden merged 28 commits into

mainfrom

feat/router-classifier

Aug 8, 2026

Commits on Aug 7, 2026

Add a benchmark-farmed corpus for router task classification
bsbodden
committed
57833b2
Add TaskIndexBuilder: writes the classifier index as a persistent collection
bsbodden
committed
d160a73
Add TaskIndex and PretrainedTaskClassifier over the persistent index
bsbodden
committed
498bc21
Add TaskIndexCli: builds the router task index and gates its accuracy
bsbodden
committed
8c0be50
Stream bake-off results to disk after each model
bsbodden
committed
ea5634f
Fix four task sources that were labelling the corpus wrongly
bsbodden
committed
2caa365
Add threshold calibration measuring in-domain against out-of-domain
bsbodden
committed
ee1b9db
Add TaskIndexResource: unpack the shipped index so it can be mapped
bsbodden
committed
219b84f
Keep corpus iteration order stable, and digest the corpus into the manifest
bsbodden
committed
56cba8d
Add the llama.cpp reference for EmbeddingGemma-300M Q8_0
bsbodden
committed
17708af
Wire task-index into the benchmark CLI dispatcher
bsbodden
committed
6d9225a
Implement the gemma-embedding encoder path
bsbodden
committed
3051d9b
Leave pooling to encoders when building the task index
bsbodden
committed
0579392
Support Matryoshka truncation when embedding
bsbodden
committed
9a5ee9b
Record the index quantizer, and cache embeddings between builds
bsbodden
committed
f0a8c14

Commits on Aug 8, 2026

Store the router index as quantized codes only
bsbodden
committed
3580b30
Parallelise the encoder, and ship the index in models-router
bsbodden
committed
20293e2
Move the training corpus to its own repository
bsbodden
committed
9c13b80
Rebuild the shipped index with instruction-only prompts
bsbodden
committed
542efa1
Add a catalog SPI so local models register themselves
bsbodden
committed
154ccbe
Discover Apple on-device intelligence, behind an opt-in
bsbodden
committed
7cdfc2a
Estimate throughput instead of dropping unprofiled models
bsbodden
committed
93f37e7
Merge origin/main into feat/router-classifier
bsbodden
committed
f28fb50
Bump to 0.3.1
bsbodden
committed
8a33385
Actually commit the classifier index
bsbodden
committed
bfdfd78
Keep the catalog hint out of the architecture guard
bsbodden
committed
8489891
Carry 0.3.1 through the release metadata
bsbodden
committed
c4fccfd
Keep Antora's component version symbolic
bsbodden
committed
8e6aece