Consume focused TinyInference LLM and embeddings crates

Cargo.toml

@@ -78,15 +78,19 @@ tinymemory-api = { path = "crates/tinymemory-api" }
 # source onto the same checkout `tinymemory-core` resolves through crates.io so
 # both crates share one trait identity.
 [patch."https://github.com/tinyhumansai/tinyinference"]
-tinyinference = { path = "vendor/tinyinference/crates/tinyinference" }
+tinyinference-core = { path = "vendor/tinyinference/crates/tinyinference-core" }
+tinyinference-embeddings = { path = "vendor/tinyinference/crates/tinyinference-embeddings" }
+tinyinference-llm = { path = "vendor/tinyinference/crates/tinyinference-llm" }
 
 [patch.crates-io]
 tinycortex = { path = "vendor/tinycortex" }
 tinycortex-api = { path = "vendor/tinycortex/api" }
 # `tinymemory-core` names the provider-neutral chat and embedding contracts
 # from TinyInference. It is not published, so standalone builds resolve the
 # version requirement to the pinned checkout here.
-tinyinference = { path = "vendor/tinyinference/crates/tinyinference" }
+tinyinference-core = { path = "vendor/tinyinference/crates/tinyinference-core" }
+tinyinference-embeddings = { path = "vendor/tinyinference/crates/tinyinference-embeddings" }
+tinyinference-llm = { path = "vendor/tinyinference/crates/tinyinference-llm" }
 
 [profile.release]
 # Cross-crate optimization and smaller, faster binaries for release builds.

crates/tinymemory-core/Cargo.toml

@@ -40,7 +40,8 @@ tinycortex-api = { version = "0.1" }
 # Provider-neutral chat-model and embedding primitives used by the tree
 # summarizer and embedding factory. Agent runtime and session behavior do not
 # belong in the memory layer.
-tinyinference = "0.2"
+tinyinference-embeddings = "0.2"
+tinyinference-llm = "0.2"
 anyhow = "1.0"
 async-trait = "0.1"
 # NO `git2` HERE, DELIBERATELY — do not add it back. `diff/` holds the ops and

crates/tinymemory-core/src/chat.rs

@@ -13,8 +13,8 @@ use async_trait::async_trait;
 
 use crate::chat_host::{create_chat_model_with_model_id, provider_for_role, UsageInfo};
 use crate::Config;
-use tinyinference::message::Message;
-use tinyinference::model::{ChatModel, ModelRequest};
+use tinyinference_llm::message::Message;
+use tinyinference_llm::model::{ChatModel, ModelRequest};
 
 /// One pair of prompt messages handed to the memory LLM backend.
 #[derive(Debug, Clone)]

crates/tinymemory-core/src/chat_host.rs

@@ -8,7 +8,7 @@
 //!
 //! # Why this trait is here and not in `tinymemory-api`
 //!
-//! It names [`tinyinference::model::ChatModel`], and the contract crate is
+//! It names [`tinyinference_llm::model::ChatModel`], and the contract crate is
 //! deliberately dependency-light — it must not pull in an inference SDK. This
 //! crate already depends on TinyInference, so it is the one place that can name
 //! both the model trait and the config seam. The host implements it here.
@@ -20,7 +20,7 @@
 use std::sync::Arc;
 
 use parking_lot::RwLock;
-use tinyinference::model::{ChatModel, ModelResponse};
+use tinyinference_llm::model::{ChatModel, ModelResponse};
 
 use crate::Config;
 

crates/tinymemory-core/src/embedding_adapter.rs

@@ -2,14 +2,14 @@
 //! model trait onto the seam's [`EmbeddingProvider`].
 //!
 //! It lives in this crate rather than in `tinymemory-api` because the contract
-//! crate must stay dependency-light and cannot name `tinyinference`; and rather
+//! crate must stay dependency-light and cannot name `tinyinference-llm`; and rather
 //! than in the host because the tree's embedder factory — which is core code —
 //! builds Ollama models directly and needs to wrap them. The host re-exports it
 //! from `inference::embeddings`, so every existing path there keeps resolving
 //! and keeps naming this one type.
 
 use async_trait::async_trait;
-use tinyinference::embeddings::EmbeddingModel;
+use tinyinference_embeddings::EmbeddingModel;
 
 pub use tinymemory_api::host::{format_embedding_signature, EmbeddingProvider};
 

crates/tinymemory-core/src/store/factories.rs

@@ -18,7 +18,7 @@ use rusqlite::Connection;
 use crate::embedding_host::require_embedding_host;
 use crate::store::namespace_store::UnifiedMemory;
 use crate::traits::Memory;
-use tinyinference::embeddings::{DEFAULT_OLLAMA_DIMENSIONS, DEFAULT_OLLAMA_MODEL};
+use tinyinference_embeddings::{DEFAULT_OLLAMA_DIMENSIONS, DEFAULT_OLLAMA_MODEL};
 use tinymemory_api::host::MemoryConfig;
 use tinymemory_api::host::{format_embedding_signature, EmbeddingProvider};
 use tinymemory_api::host::{EmbeddingRouteConfig, StorageProviderConfig};

crates/tinymemory-core/src/store/factories_tests.rs

@@ -91,7 +91,7 @@ fn embedding_settings_local_overrides_memory_config() {
     assert_eq!(model, "nomic-embed-text:latest");
     assert_eq!(
         dims,
-        tinyinference::embeddings::DEFAULT_OLLAMA_DIMENSIONS,
+        tinyinference_embeddings::DEFAULT_OLLAMA_DIMENSIONS,
         "dimensions must default to Ollama default"
     );
 }
@@ -105,10 +105,10 @@ fn embedding_settings_local_with_empty_model_uses_default() {
     assert_eq!(provider, "ollama");
     assert_eq!(
         model,
-        tinyinference::embeddings::DEFAULT_OLLAMA_MODEL,
+        tinyinference_embeddings::DEFAULT_OLLAMA_MODEL,
         "empty model ID must fall back to default Ollama model"
     );
-    assert_eq!(dims, tinyinference::embeddings::DEFAULT_OLLAMA_DIMENSIONS);
+    assert_eq!(dims, tinyinference_embeddings::DEFAULT_OLLAMA_DIMENSIONS);
 }
 
 #[test]
@@ -223,7 +223,7 @@ async fn start_mock_ollama() -> String {
 /// the legacy `local_ai.usage.embeddings = true` flag was set. Used so
 /// the existing test scenarios continue to drive the local code path.
 fn local_embedding_for_test() -> &'static str {
-    tinyinference::embeddings::DEFAULT_OLLAMA_MODEL
+    tinyinference_embeddings::DEFAULT_OLLAMA_MODEL
 }
 
 #[tokio::test]

crates/tinymemory-core/src/test_seams.rs

@@ -72,13 +72,13 @@ impl crate::chat_host::ChatHost for TestChatHost {
         _role: &str,
         _config: &Config,
         _temperature: f64,
-    ) -> Result<(Arc<dyn tinyinference::model::ChatModel<()>>, String), String> {
+    ) -> Result<(Arc<dyn tinyinference_llm::model::ChatModel<()>>, String), String> {
         Err("TestChatHost does not build models — model routing is host behaviour".to_string())
     }
 
     fn usage_from_response(
         &self,
-        _response: &tinyinference::model::ModelResponse,
+        _response: &tinyinference_llm::model::ModelResponse,
     ) -> Option<tinymemory_api::host::UsageInfo> {
         None
     }

crates/tinymemory-core/src/tree/score/embed/factory.rs

@@ -39,7 +39,7 @@ use tinymemory_api::host::test_support::TestHostConfig;
 use super::{Embedder, InertEmbedder, ProviderEmbedder, EMBEDDING_DIM};
 use crate::embedding_host::require_embedding_host;
 use crate::Config;
-use tinyinference::embeddings::{OllamaEmbeddingModel, RECOMMENDED_OLLAMA_CONTEXT_TOKENS};
+use tinyinference_embeddings::{OllamaEmbeddingModel, RECOMMENDED_OLLAMA_CONTEXT_TOKENS};
 
 /// Cheap heuristic for "is a backend session reachable?" — the cloud
 /// embedder needs one and bails on first embed call without it. We use

crates/tinymemory-core/src/tree/score/embed/mod.rs

@@ -87,7 +87,7 @@ pub trait Embedder: Send + Sync {
 
 /// Adapts the canonical host embedding-provider contract to the legacy
 /// memory-tree embedder shape. Concrete network implementations live in
-/// `tinyinference::embeddings`; this bridge owns only dimension checks
+/// `tinyinference_llm::embeddings`; this bridge owns only dimension checks
 /// and the memory tree's per-position batch fallback contract.
 pub struct ProviderEmbedder {
     inner: Box<dyn tinymemory_api::host::EmbeddingProvider>,

crates/tinymemory-core/src/tree/tree_runtime/engine.rs

@@ -8,8 +8,8 @@ use crate::engine::backend::tree::runtime::{
 use anyhow::{Context, Result};
 use async_trait::async_trait;
 use chrono::{DateTime, Timelike, Utc};
-use tinyinference::message::Message;
-use tinyinference::model::{ChatModel, ModelRequest};
+use tinyinference_llm::message::Message;
+use tinyinference_llm::model::{ChatModel, ModelRequest};
 
 use crate::engine::engine_config;
 use crate::Config;

crates/tinymemory-core/src/tree/tree_runtime/engine_tests.rs

@@ -5,7 +5,7 @@ use crate::tree::tree_runtime::store;
 use chrono::TimeZone;
 use std::sync::Mutex;
 use tempfile::TempDir;
-use tinyinference::model::ModelResponse;
+use tinyinference_llm::model::ModelResponse;
 use tinymemory_api::host::test_support::TestHostConfig;
 
 struct RecordingModel {
@@ -28,11 +28,11 @@ impl ChatModel<()> for RecordingModel {
         &self,
         _state: &(),
         request: ModelRequest,
-    ) -> tinyinference::Result<ModelResponse> {
+    ) -> tinyinference_llm::Result<ModelResponse> {
         self.requests.lock().unwrap().push(request);
         match &self.reply {
             Ok(reply) => Ok(ModelResponse::assistant(reply.clone())),
-            Err(message) => Err(tinyinference::Error::Model(message.clone())),
+            Err(message) => Err(tinyinference_llm::Error::Model(message.clone())),
         }
     }
 }