fix(embeddings): Ollama embeddings via native client; ollama in base deps

The default zero-config stack crashed on a lean install: `memorizz` ->
"langchain_ollama is required for Ollama embeddings". Two missing-dependency
problems:

1. The Ollama embedding provider used `langchain_ollama` (only in the `[ollama]`
   extra). Rewrite it to use the native `ollama` client's embeddings endpoint —
   the same SDK the LLM side already uses. langchain_ollama is dropped entirely.
2. The `ollama` client wasn't a base dependency, so even without (1) a bare
   `pip install memorizz` couldn't run the default Ollama stack. Move
   `ollama>=0.6.0` into base (small; httpx already comes via openai). No torch.

Why this slipped through testing: my dev env (memorizz_local) had
langchain_ollama installed (from the [all] extra), and I tested via that
interpreter rather than the shipped CLI's lean install — so the missing dep
never surfaced. Now verified in a clean 3.11 venv from the wheel: ollama present,
langchain_ollama/torch absent, build_session_agent + a 768-dim embedding work.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: RichmondAlake

CHANGELOG.md

@@ -22,6 +22,12 @@
 
 ### Fixes
 
+* **Default local stack works on a lean install.** Ollama embeddings required
+  `langchain_ollama` (an optional extra) and the `ollama` client wasn't a base
+  dependency, so a plain `pip install memorizz` + `memorizz` crashed with
+  "langchain_ollama is required for Ollama embeddings". Embeddings now use the
+  native `ollama` client, and `ollama` is a base dependency (still no langchain,
+  no torch).
 * Agents executing a scheduled automation no longer wander into managing
   automations — automation-management tools are suppressed for the run (fixes
   off-task / empty outputs from smaller models).

pyproject.toml

@@ -46,6 +46,11 @@ dependencies = [
     "rich>=13.7.0",
     "prompt_toolkit>=3.0.0",
     "platformdirs>=4.0.0",
+    # Ollama daemon SDK (LLM *and* embeddings via the native client). Kept in
+    # base, not an extra, so the default zero-config local stack works on a bare
+    # `pip install memorizz` with a running Ollama daemon. Small (httpx, already
+    # pulled by openai) — none of the heavy local-ML stack.
+    "ollama>=0.6.0",
     # NOTE: the heavy local-ML stack (transformers / sentence-transformers /
     # accelerate / huggingface-hub) moved OUT of base into the `huggingface`
     # extra so the default install stays lean (no multi-GB torch download).
@@ -62,10 +67,8 @@ cli = []
 # Fully-local stack: Ollama LLM + Ollama embeddings + FAISS filesystem memory.
 local = ["memorizz[ollama]", "memorizz[filesystem]"]
 anthropic = ["anthropic>=0.26.0"]
-# langchain-ollama 1.x dropped the `timeout=` constructor kwarg; the embedding
-# provider has a fallback for it but we still want the ollama daemon SDK at a
-# version that knows about Gemma 4 tags (>=0.6.0).
-ollama = ["langchain_ollama>=1.0.0", "ollama>=0.6.0"]
+# Ollama ships in base now; this alias keeps `pip install memorizz[ollama]` valid.
+ollama = ["ollama>=0.6.0"]
 voyageai = ["voyageai"]
 huggingface = [
     "numpy>=1.21,<2",
@@ -113,7 +116,6 @@ all = [
     "anthropic>=0.26.0",
     "pymongo>=4.0.0",
     "oracledb>=2.0.0",
-    "langchain_ollama>=1.0.0",
     "ollama>=0.6.0",
     "voyageai",
     "faiss-cpu>=1.7.4",

src/memorizz/embeddings/ollama/provider.py

@@ -54,30 +54,38 @@ def __init__(self, config: Dict[str, Any] = None):
         )
 
     def _init_client(self):
-        """Initialize the Ollama embeddings client."""
+        """Initialize the native Ollama client (no langchain dependency)."""
         try:
-            from langchain_ollama import OllamaEmbeddings
+            import ollama
         except ImportError:
             raise ImportError(
-                "langchain_ollama is required for Ollama embeddings. "
-                "Install it with: pip install langchain-ollama"
+                "The 'ollama' package is required for Ollama embeddings. Install "
+                "it with: pip install ollama (or: pip install memorizz[ollama])"
             )
 
-        # langchain-ollama >=1.0 dropped the `timeout` constructor arg.
+        client_kwargs: Dict[str, Any] = {"host": self.base_url}
+        if self.timeout:
+            client_kwargs["timeout"] = self.timeout
         try:
-            self.embeddings = OllamaEmbeddings(
-                model=self.model, base_url=self.base_url, timeout=self.timeout
-            )
-        except (TypeError, ValueError):
-            self.embeddings = OllamaEmbeddings(model=self.model, base_url=self.base_url)
+            self.client = ollama.Client(**client_kwargs)
+        except TypeError:
+            self.client = ollama.Client(host=self.base_url)
+
+    def _embed_one(self, text: str) -> List[float]:
+        """Embed a single string via the native client (attr- or dict-style resp)."""
+        resp = self.client.embeddings(model=self.model, prompt=text)
+        vector = getattr(resp, "embedding", None)
+        if vector is None and isinstance(resp, dict):
+            vector = resp.get("embedding")
+        return list(vector or [])
 
     def _probe_dimensions(self):
         """Probe the actual dimensions of the model by generating a test embedding."""
         if self._dimensions_probed:
             return
 
         try:
-            test_embedding = self.embeddings.embed_query("test")
+            test_embedding = self._embed_one("test")
             if test_embedding:
                 actual_dimensions = len(test_embedding)
                 if actual_dimensions != self.dimensions:
@@ -126,7 +134,7 @@ def get_embedding(self, text: str, **kwargs) -> List[float]:
                 self._probe_dimensions()
 
             # Generate embedding
-            embedding = self.embeddings.embed_query(text)
+            embedding = self._embed_one(text)
             return embedding
         except Exception as e:
             logger.error(f"Error generating Ollama embedding: {str(e)}")