gte 0.0.12-x86_64-linux → 0.0.13-x86_64-linux

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 23560cdb3c9c908ba6edb00720125fd719a13c8f6f36baa3352e069bd8784bb9
-  data.tar.gz: cd439d8937edf771a123b9f64e54a8254c4eec3b432947392077c582589726de
+  metadata.gz: d91f5c0d89ad9b423d8cc0aa6beb349740d087dedc85fd2c8a6cc57762edaabf
+  data.tar.gz: d46f37e1e3ebfef95bc29eff93904c7195d9e11fa6b57896031efdbf2346dbdd
 SHA512:
-  metadata.gz: fcd7afb00b292447a6906a599e598ad8cef8751082372dea8d7031010eefc21c0dc0af6cd1870078799152a54b9bb14f795b1a335973cdb0ef58cf556ab81c39
-  data.tar.gz: 5c1b2e6b4a3bfb5f647da1786939f329d462f03fa209558a53ea3165af57cbfb33b4a5712df718dc1f2d36967e384784a23e94e0e792e707eb34de700a5430a0
+  metadata.gz: 8bb9df2128011282a0b8ebcde4d885be18b204e7659ab108d0aa1dbcd660f405f359cc7e3ce20447cfe8e8a145e65e9d49995b21a6c948f3a44d8a102d68533b
+  data.tar.gz: d8cdc887855b3d8cff93b9134435609fd1fc77bf22290dec68d85b0e9dae785511c6d73c05e142061b3543dde1aa01a56a399cb3a6129b1812bf71733d58e5ee

data/README.md

@@ -59,6 +59,43 @@ Notes:
 - Return a `Config::Text` from the block (for example, `config.with(...)`).
 - Model instances are cached by full config key; different config values create different cached instances.
 
+Common model presets:
+
+```ruby
+e5 = GTE.config(ENV.fetch("GTE_MODEL_DIR")) do |config|
+  config.with(
+    model_name: "model.onnx",
+    output_tensor: "last_hidden_state",
+    max_length: 512,
+    execution_providers: "cpu"
+  )
+end
+
+siglip2 = GTE.config(ENV.fetch("GTE_SIGLIP2_DIR")) do |config|
+  config.with(
+    model_name: "text_model_int8.onnx",
+    output_tensor: "pooler_output",
+    max_length: 64,
+    execution_providers: "cpu"
+  )
+end
+
+clip = GTE.config(ENV.fetch("GTE_CLIP_DIR")) do |config|
+  config.with(
+    output_tensor: "sentence_embedding",
+    max_length: 512,
+    execution_providers: "cpu"
+  )
+end
+```
+
+Picking a specific layer:
+
+- Use `output_tensor:` to request a named model output.
+- `last_hidden_state` gives token-level hidden states and is mean-pooled by `gte` when the tensor is rank 3.
+- `pooler_output`, `sentence_embedding`, and similar 2D tensors are returned directly and then L2-normalized by default.
+- If the requested tensor is not present in the model, `gte` raises an error instead of silently falling back.
+
 Low-level embedder setup (without model cache):
 
 ```ruby

data/VERSION

@@ -1 +1 @@
-0.0.12
+0.0.13

data/ext/gte/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "gte"
-version = "0.0.12"
+version = "0.0.13"
 edition = "2021"
 authors = ["elcuervo <elcuervo@elcuervo.net>"]
 license = "MIT"

data/ext/gte/src/session.rs

@@ -7,7 +7,7 @@ use ndarray::{Array2, ArrayView2, ArrayViewD, Ix2};
 use ort::execution_providers::{
     CoreMLExecutionProvider, ExecutionProviderDispatch, XNNPACKExecutionProvider,
 };
-use ort::session::Session;
+use ort::session::{OutputSelector, RunOptions, Session};
 use std::path::{Path, PathBuf};
 use std::sync::atomic::{AtomicUsize, Ordering};
 use std::sync::{Condvar, Mutex};
@@ -216,8 +216,11 @@ pub fn run_session(
     config: &ModelConfig,
 ) -> Result<Array2<f32>> {
     let input_tensors = InputTensors::from_tokenized(tokenized, config.with_attention_mask)?;
+    let run_opts = RunOptions::new()
+        .map_err(|e| GteError::Ort(e.to_string()))?
+        .with_outputs(OutputSelector::no_default().with(config.output_tensor.as_str()));
     let outputs = session
-        .run(input_tensors.inputs)
+        .run_with_options(input_tensors.inputs, &run_opts)
         .map_err(|e| GteError::Ort(e.to_string()))?;
     let array = extract_output_tensor(&outputs, config.output_tensor.as_str())?;
 

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: gte
 version: !ruby/object:Gem::Version
-  version: 0.0.12
+  version: 0.0.13
 platform: x86_64-linux
 authors:
 - elcuervo
 autorequire: 
 bindir: bin
 cert_chain: []
-date: 2026-05-13 00:00:00.000000000 Z
+date: 2026-05-14 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: rake