Consume focused TinyInference LLM and embeddings crates

Cargo.toml

@@ -121,7 +121,8 @@ thiserror = "2"
 # published; embedding hosts may patch this exact git source so the host and
 # TinyCortex share one `EmbeddingModel` trait identity without pulling in an
 # agent runtime.
-tinyinference = { git = "https://github.com/tinyhumansai/tinyinference", rev = "cc8aca484bb995fbab3b7358f0d72ab5056b587c" }
+tinyinference-embeddings = { git = "https://github.com/tinyhumansai/tinyinference", rev = "bb0f820" }
+tinyinference-llm = { git = "https://github.com/tinyhumansai/tinyinference", rev = "bb0f820" }
 # The stable contract crate (value types, error enum, `Memory` trait). Declared
 # as a path dependency, not a registry requirement, so embedding hosts resolve
 # it through this checkout without needing their own `[patch.crates-io]` entry.

examples/persona_agent.rs

@@ -5,7 +5,7 @@
 //!
 //! 1. [`PersonaRetriever`] ranks persisted observations with deterministic,
 //!    network-free BM25 retrieval weighted by evidence tier.
-//! 2. A provider-neutral [`tinyinference`] model receives only those ranked
+//! 2. A provider-neutral [`tinyinference-llm`] model receives only those ranked
 //!    observations and the person's explicit directives, then writes one
 //!    evidence-cited decision.
 //!
@@ -25,9 +25,9 @@
 //! - `TINYCORTEX_LLM_MODEL` — model id (default `deepseek/deepseek-v4-flash`).
 //! - `PERSONA_IDENTITY` — identity label in the prompt (default `this developer`).
 
-use tinyinference::message::Message;
-use tinyinference::model::{ChatModel, ModelRequest};
-use tinyinference::providers::openai::OpenAiModel;
+use tinyinference_llm::message::Message;
+use tinyinference_llm::model::{ChatModel, ModelRequest};
+use tinyinference_llm::providers::openai::OpenAiModel;
 
 use tinycortex::memory::config::MemoryConfig;
 use tinycortex::memory::persona::compile::read_directives;

src/memory/chunks/signature.rs

@@ -22,10 +22,10 @@
 ///
 /// Delegates to the same formatter the namespace store's embedding providers
 /// use ([`crate::memory::store::vectors::format_embedding_signature`], itself a
-/// re-export from `tinyinference`), so the two stores cannot drift into
+/// re-export from `tinyinference-embeddings`), so the two stores cannot drift into
 /// byte-different spellings of the same space again.
 pub fn format_signature(provider: &str, model: &str, dims: usize) -> String {
-    tinyinference::embeddings::format_embedding_signature(provider, model, dims)
+    tinyinference_embeddings::format_embedding_signature(provider, model, dims)
 }
 
 /// A signature decomposed into the parts that identify its vector space.

src/memory/health.rs

@@ -221,7 +221,7 @@ pub fn classify_embed_error_str(msg: &str) -> PipelineFailure {
 
     // #5354 — the local Ollama runtime is not usable: the daemon is not
     // listening, or the configured embedding model was never pulled. Both are
-    // emitted by `tinyinference::embeddings::ollama` with the fix already
+    // emitted by `tinyinference_embeddings::ollama` with the fix already
     // in the text:
     //
     //   "ollama embed request failed (is Ollama running at <base>?): …"

src/memory/health_tests.rs

@@ -266,7 +266,7 @@ fn classify_transport_error_is_transient() {
 }
 
 /// #5354 — the Ollama daemon is not listening. Verbatim wording from
-/// `tinyinference::embeddings::OllamaEmbeddingModel` request path.
+/// `tinyinference_embeddings::OllamaEmbeddingModel` request path.
 /// Note the parenthesised hint: `parse_http_status` reads the first `(`, so
 /// without an explicit match this fell through to `Transient` and the panel
 /// told the user to wait for a retry that can never start their daemon.

src/memory/store/vectors/embedding.rs

@@ -20,7 +20,7 @@ use async_trait::async_trait;
 /// instantiated backend. Drift between the two would silently split one
 /// embedding space into two.
 pub fn format_embedding_signature(name: &str, model_id: &str, dims: usize) -> String {
-    tinyinference::embeddings::format_embedding_signature(name, model_id, dims)
+    tinyinference_embeddings::format_embedding_signature(name, model_id, dims)
 }
 
 /// Interface for embedding backends that convert text into numerical vectors.
@@ -63,30 +63,30 @@ pub trait EmbeddingBackend: Send + Sync {
 #[async_trait]
 impl<T> EmbeddingBackend for T
 where
-    T: tinyinference::embeddings::EmbeddingModel + ?Sized,
+    T: tinyinference_embeddings::EmbeddingModel + ?Sized,
 {
     fn name(&self) -> &str {
-        tinyinference::embeddings::EmbeddingModel::name(self)
+        tinyinference_embeddings::EmbeddingModel::name(self)
     }
 
     fn model_id(&self) -> &str {
-        tinyinference::embeddings::EmbeddingModel::model_id(self)
+        tinyinference_embeddings::EmbeddingModel::model_id(self)
     }
 
     fn dimensions(&self) -> usize {
-        tinyinference::embeddings::EmbeddingModel::dimensions(self)
+        tinyinference_embeddings::EmbeddingModel::dimensions(self)
     }
 
     fn signature(&self) -> String {
-        tinyinference::embeddings::EmbeddingModel::signature(self)
+        tinyinference_embeddings::EmbeddingModel::signature(self)
     }
 
     async fn embed(&self, texts: &[&str]) -> anyhow::Result<Vec<Vec<f32>>> {
         let owned = texts
             .iter()
             .map(|text| (*text).to_owned())
             .collect::<Vec<_>>();
-        tinyinference::embeddings::EmbeddingModel::embed(self, &owned)
+        tinyinference_embeddings::EmbeddingModel::embed(self, &owned)
             .await
             .map_err(|error| anyhow::anyhow!(error))
     }