feat(search): support Ollama embedding models

README.md

@@ -241,34 +241,50 @@ The large default context can exceed memory for big models on constrained hardwa
 
 ### Embedding Model
 
-`skilld search` uses a local embedding model. It runs offline through transformers.js. It needs no API key or network after the first download. Pick one under **Embedding model** in `skilld config`:
+`skilld search` is powered by a local embedding model. It runs offline through transformers.js. It needs no API key or network traffic after the first download. Pick one under **Embedding model** in `skilld config`:
 
 | Model | Dimensions | Notes |
 |-------|-----------:|-------|
 | `bge-small-en-v1.5` | 384 | Default. Fastest to index, smallest download. |
 | `bge-base-en-v1.5` | 768 | Balanced accuracy and speed. |
+| `Xenova/bge-large-en-v1.5` | 1024 | Most accurate English retrieval, slowest to index. |
 | `bge-m3` | 1024 | Multilingual, 8192-token context. |
 
-Larger models retrieve more accurately but cost more time and memory when indexing. Set `SKILLD_EMBED_MODEL` to override the saved setting:
+Larger models retrieve more accurately but cost more time and memory when indexing. Locally-pulled [Ollama](#ollama-embedding-models) models can be used too. Export `SKILLD_EMBED_MODEL` to override the saved setting for every index and search command in the current shell:
 
 ```bash
 export SKILLD_EMBED_MODEL=bge-m3
-skilld update --force
+skilld update
 ```
 
-Each search index belongs to one model and device. Keep environment overrides set for both indexing and querying. Rebuild indexes after either setting changes:
+Search indexes must use one embedding model and device. Rebuild them after switching either setting:
 
 ```bash
-skilld update --force
+skilld update
 ```
 
+### Ollama Embedding Models
+
+If [Ollama](https://ollama.com) is running, locally-pulled embedding models appear in the **Embedding model** picker alongside the built-in ones. They are addressed as `ollama:<name>`, matching the `-m ollama:<name>` syntax used for enhancement models:
+
+```bash
+ollama pull qwen3-embedding
+SKILLD_EMBED_MODEL=ollama:qwen3-embedding skilld add npm:vue
+```
+
+Only models that advertise the `embedding` capability are listed, so chat models cannot be selected by mistake. Dimensions and context length are read from Ollama. skilld validates each vector before indexing.
+
+This talks to Ollama's HTTP API directly. It needs no additional dependency or API key. Set `OLLAMA_HOST` to point at a non-default daemon. If Ollama is not running, the picker simply shows the built-in models.
+
+Ollama manages its own execution device, so **Embedding device** does not apply to `ollama:` models.
+
 ### Embedding Device
 
 The embedding model runs on the CPU by default. **Embedding device** in `skilld config` moves it onto a GPU backend, which can be substantially faster:
 
 | Device | Notes |
 |--------|-------|
-| `auto` | Default. Lets transformers.js choose, CPU under Node. |
+| `auto` | Default. Lets transformers.js choose; CPU under Node. |
 | `cpu` | Always available, predictable. |
 | `webgpu` | Fastest on Apple Silicon in testing. |
 | `coreml` | Apple Neural Engine. Measured slower than CPU for these models. |
@@ -281,13 +297,13 @@ Measured on an Apple M5 Max, 120 documents, best of 3 after warm-up (docs/sec):
 | `bge-base-en-v1.5` | 198 | 68 | **580** |
 | `Xenova/bge-large-en-v1.5` | 71 | 9 | **201** |
 
-WebGPU was 2.6 to 2.9 times faster than CPU at every size. `bge-large` on WebGPU indexed faster than `bge-base` on CPU. CoreML was consistently slower.
+WebGPU was 2.6 to 2.9 times faster than CPU at every size. On WebGPU, `bge-large` indexed faster than `bge-base` on CPU and improved retrieval. CoreML was consistently slower.
 
-The ranking is hardware-specific, so benchmark before trusting a device on other machines. Set `SKILLD_EMBED_DEVICE` to override the saved setting:
+The ranking is hardware-specific, so benchmark before trusting a device on other machines. Export `SKILLD_EMBED_DEVICE` to override the saved setting in the current shell:
 
 ```bash
 export SKILLD_EMBED_DEVICE=cpu
-skilld update --force
+skilld update
 ```
 
 If a backend is unavailable, indexing fails to start. Switch back to `auto`.

package.json

@@ -84,6 +84,7 @@
   "devDependencies": {
     "@antfu/eslint-config": "catalog:dev-lint",
     "@types/node": "catalog:dev-build",
+    "@typescript/native": "catalog:",
     "@vitest/coverage-v8": "catalog:dev-test",
     "bumpp": "catalog:",
     "eslint": "catalog:dev-lint",

packages/protocol/package.json

@@ -59,6 +59,7 @@
     "@antfu/eslint-config": "catalog:dev-lint",
     "@arethetypeswrong/cli": "catalog:",
     "@types/node": "catalog:dev-build",
+    "@typescript/native": "catalog:",
     "eslint": "catalog:dev-lint",
     "obuild": "catalog:dev-build",
     "publint": "catalog:",