feat: support Jina AI v5 Matryoshka embedding models (nano + small)

CHANGELOG.md

@@ -0,0 +1,15 @@
+# Changelog
+
+## [Unreleased]
+
+### Added
+- Jina AI embeddings v5 text backends (`jina-v5-nano`, `jina-v5-small`) via local ONNX inference
+- Matryoshka representation learning: configurable truncate_dim for jina-v5 backends
+- Asymmetric retrieval: `retrieval.query:` / `retrieval.passage:` instruction prefixes
+- Auto re-embed on embedder dimension change (`--no-auto-reembed` to opt out)
+- `embed_query` / `embed_document` distinction on the `Embedder` trait
+- `icm config show` now displays active backend name and license tag
+- `icm recall` now shows active model name in output header
+
+### License note
+Jina v5 model weights are CC BY-NC 4.0 (non-commercial). Commercial use requires a license from Jina AI.

Cargo.toml

@@ -24,6 +24,18 @@ zerocopy = { version = "0.8", features = ["derive"] }
 # Embeddings (optional)
 fastembed = "4"
 
+# Jina v5 embedder. These are workspace version pins only; each
+# consumer crate must mark them `optional = true` and gate them behind
+# the `jina-v5` feature. Default builds must NOT pull these crates.
+#
+# `ort` defaults pull `download-binaries` which fetches the ONNX Runtime
+# at build time. We disable defaults and load the system runtime via
+# `load-dynamic`; `ndarray` is required for `try_extract_tensor`.
+hf-hub = "0.4"
+ort = { version = "2.0.0-rc.9", default-features = false, features = ["load-dynamic", "ndarray"] }
+tokenizers = "0.21"
+ndarray = "0.16"
+
 # Serialization
 serde = { version = "1", features = ["derive"] }
 serde_json = { version = "1", features = ["preserve_order"] }

README.md

@@ -392,6 +392,36 @@ C:\Users\<user>\AppData\Roaming\icm\icm\config\config.toml              # Window
 
 See [config/default.toml](config/default.toml) for all options.
 
+## Embedder backends
+
+ICM supports three local ONNX embedding backends — no external API required.
+
+| Backend | Dims | License | Notes |
+|---------|------|---------|-------|
+| `fastembed` (default) | 384 / 768 / 1024 (model-dependent) | Apache-2.0 | multilingual-e5-base and others via fastembed |
+| `jina-v5-nano` | 32, 64, 128, 256, 512, 768 (default: 768) | CC BY-NC 4.0 | jinaai/jina-embeddings-v5-text-nano-retrieval |
+| `jina-v5-small` | 32, 64, 128, 256, 512, 768, 1024 (default: 1024) | CC BY-NC 4.0 | jinaai/jina-embeddings-v5-text-small-retrieval (Qwen3-based) |
+
+> **IMPORTANT — Non-commercial restriction:** Jina v5 model weights are licensed under
+> [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). **Use in commercial products
+> requires a commercial Jina AI license.** See https://jina.ai/contact-sales for details.
+> The default `fastembed` backend (Apache-2.0) has no such restriction.
+
+Weights download automatically to `~/.cache/huggingface/` on first run. No account or API key needed.
+
+```toml
+# config.toml
+[embeddings]
+backend = "jina-v5-nano"
+truncate_dim = 512  # optional Matryoshka dim (omit to use model default)
+```
+
+Matryoshka truncation lets you trade accuracy for speed and storage. Valid dims per backend:
+
+- `jina-v5-nano`: 32, 64, 128, 256, 512, 768 (default: 768)
+- `jina-v5-small`: 32, 64, 128, 256, 512, 768, 1024 (default: 1024)
+- `fastembed`: no truncation; dim is fixed by the chosen model
+
 ## Auto-extraction
 
 ICM extracts memories automatically via three layers:

config/default.toml

@@ -24,7 +24,15 @@ prune_threshold = 0.1
 # Set to false to disable embeddings entirely (no model download, keyword search only)
 # enabled = false
 
-# Embedding model (fastembed model_code). Default: multilingual-e5-small (384d, 100+ languages)
+# Embedder backend: "fastembed" (default, Apache-2.0) | "jina-v5-nano" | "jina-v5-small"
+#
+# Jina v5 backends: CC BY-NC 4.0 (non-commercial). For commercial use, acquire a
+# commercial license from Jina AI before deploying. Weights download to
+# ~/.cache/huggingface/ on first run.
+#
+# backend = "jina-v5-nano"
+
+# Embedding model (fastembed model_code). Default: multilingual-e5-base (768d, 100+ languages)
 # Other options:
 #   "BAAI/bge-small-en-v1.5"              — 384d, English-only (fastest)
 #   "Xenova/bge-small-en-v1.5"            — 384d, quantized English-only
@@ -34,6 +42,11 @@ prune_threshold = 0.1
 #   "jinaai/jina-embeddings-v2-base-code" — 768d, code-optimized
 model = "intfloat/multilingual-e5-base"
 
+# Matryoshka truncation (jina-v5 backends only). Valid dims:
+#   jina-v5-nano:  32, 64, 128, 256, 512, 768 (default: 768)
+#   jina-v5-small: 32, 64, 128, 256, 512, 768, 1024 (default: 1024)
+# truncate_dim = 512
+
 [extraction]
 # Layer 0: rule-based fact extraction (zero LLM cost)
 enabled = true

crates/icm-cli/Cargo.toml

@@ -15,6 +15,7 @@ path = "src/main.rs"
 [features]
 default = ["embeddings", "tui"]
 embeddings = ["icm-core/embeddings", "icm-mcp/embeddings"]
+jina-v5 = ["icm-core/jina-v5"]
 tui = ["dep:ratatui", "dep:crossterm"]
 web = ["dep:axum", "dep:tokio", "dep:tower-http", "dep:rust-embed", "dep:mime_guess", "dep:getrandom"]
 vendored-openssl = ["openssl/vendored"]

crates/icm-cli/src/config.rs

@@ -46,21 +46,42 @@ pub struct MemoryConfig {
     pub auto_consolidate_threshold: usize,
 }
 
+/// Which embedder backend to use for `EmbeddingsConfig.backend`.
+#[derive(Debug, Deserialize, Default, PartialEq, Eq, Clone)]
+#[serde(rename_all = "kebab-case")]
+pub enum EmbedderBackend {
+    /// fastembed (default) — multilingual-e5-base etc., Apache-2.0 weights.
+    #[default]
+    Fastembed,
+    /// jina-embeddings-v5-text-nano-retrieval — local ONNX, CC-BY-NC-4.0.
+    JinaV5Nano,
+    /// jina-embeddings-v5-text-small-retrieval (Qwen3-based) — local ONNX, CC-BY-NC-4.0.
+    JinaV5Small,
+}
+
 /// Embedding model settings.
 #[derive(Debug, Deserialize)]
 #[serde(default)]
 pub struct EmbeddingsConfig {
     /// Enable embeddings (set to false to skip model download entirely).
     pub enabled: bool,
-    /// Model identifier (fastembed model_code, e.g. "intfloat/multilingual-e5-small").
+    /// Which embedder backend to use.
+    pub backend: EmbedderBackend,
+    /// Model identifier for the fastembed backend
+    /// (e.g. "intfloat/multilingual-e5-base"). Ignored by other backends.
     pub model: String,
+    /// Matryoshka truncation dimension. `None` = use the model's default
+    /// output dimension. Consumed by the jina-v5-nano and jina-v5-small backends.
+    pub truncate_dim: Option<usize>,
 }
 
 impl Default for EmbeddingsConfig {
     fn default() -> Self {
         Self {
             enabled: true,
+            backend: EmbedderBackend::Fastembed,
             model: "intfloat/multilingual-e5-base".into(),
+            truncate_dim: None,
         }
     }
 }

crates/icm-cli/src/learn_tests.rs

@@ -9,7 +9,7 @@ mod tests {
     fn test_store() -> (TempDir, SqliteStore) {
         let tmp = TempDir::new().expect("failed to create temp dir");
         let db_path = tmp.path().join("test.db");
-        let store = SqliteStore::with_dims(&db_path, 384).expect("failed to create store");
+        let (store, _) = SqliteStore::with_dims(&db_path, 384).expect("failed to create store");
         (tmp, store)
     }