graph-mind-memory 0.3.3__tar.gz → 0.4.0__tar.gz

{graph_mind_memory-0.3.3/graph_mind_memory.egg-info → graph_mind_memory-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: graph-mind-memory
-Version: 0.3.3
+Version: 0.4.0
 Summary: Local-first memory for AI coding assistants: verbatim, on your own PC, no LLM calls at write time.
 Author-email: Yohan Ko <goyohan0611@gmail.com>
 License-Expression: AGPL-3.0-only
@@ -13,11 +13,11 @@ Classifier: Operating System :: OS Independent
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: torch<3,>=2.2
 Requires-Dist: mcp<3,>=2
 Requires-Dist: numpy
-Requires-Dist: transformers
-Requires-Dist: tiktoken
+Requires-Dist: onnxruntime>=1.17
+Requires-Dist: tokenizers>=0.15
+Requires-Dist: huggingface_hub>=0.20
 Requires-Dist: pgserver; python_version < "3.13"
 Requires-Dist: psycopg[binary]
 Requires-Dist: psutil
@@ -48,8 +48,9 @@ Local-first memory for AI coding assistants. Verbatim storage, on your own PC:
 </div>
 
 > [!NOTE]
-> **Alpha.** Used daily on Windows 11, and checked in a fresh-install, two-PC end-to-end run.
-> The test suite also passes on Linux and macOS in CI, but nobody has used it day to day there yet.
+> **Alpha.** Windows: used daily, with full end-to-end checks (fresh install, two PCs, real
+> Claude and Codex). macOS and Linux: tests and a clean install from PyPI pass in CI on every
+> release and every day, but nobody has used them day to day yet; consider them experimental.
 
 ---
 
@@ -125,8 +126,8 @@ graph-mind-install
 
 If `graph-mind-install` is "not recognized", pip put it in a Scripts folder that is not on your
 PATH (common with the Windows Python install manager). `python -m install` runs the same thing.
-If it stops with `WinError 1114` loading `c10.dll`, Windows is missing the Microsoft Visual C++
-runtime that PyTorch needs: install [vc_redist.x64.exe](https://aka.ms/vs/17/release/vc_redist.x64.exe),
+If it stops with a DLL load error, Windows is missing the Microsoft Visual C++ runtime that the
+embedding runtime needs: install [vc_redist.x64.exe](https://aka.ms/vs/17/release/vc_redist.x64.exe),
 restart, and run the installer again.
 
 or from source:
@@ -139,12 +140,12 @@ python install.py
 
 The installer:
 
-- installs the packages (PyTorch CPU is the large one, about 2 GB);
+- installs the packages (no PyTorch: the embedding model runs on ONNX Runtime);
 - registers the MCP server with every app it finds: Claude Code, Claude Desktop (including the
   Microsoft Store build) and Codex;
 - starts the capture service at login (Windows Startup folder, a macOS LaunchAgent, or an XDG
   autostart entry on Linux);
-- downloads the embedding model.
+- downloads the embedding model (about 470 MB, once).
 
 Then restart your AI apps. The server is also listed in the
 [MCP Registry](https://registry.modelcontextprotocol.io) as `io.github.goyohan0611-png/graph-mind`. Running the installer again is safe. Run it again if you move the folder.
@@ -258,8 +259,9 @@ python -m unittest discover -p "test_*.py"
 
 ## Known limits
 
-- **The first question after an AI app starts waits for the embedding model to load** (a few
-  seconds; 30-50 s on a slow or synced disk). Later questions take well under a second.
+- **For the first few seconds after an AI app starts, answers use word search only** while the
+  embedding model loads in the background (6-9 s on a slow, synced disk; it was 50-70 s before
+  0.4). The first answer never waits for it; later answers take well under a second.
 - **Daily use on Windows only so far.** macOS and Linux pass the tests in CI but have not seen real
   use.
 - **Capture follows each app's transcript format,** which is not a public interface. An app update

{graph_mind_memory-0.3.3 → graph_mind_memory-0.4.0}/README.md

@@ -22,8 +22,9 @@ Local-first memory for AI coding assistants. Verbatim storage, on your own PC:
 </div>
 
 > [!NOTE]
-> **Alpha.** Used daily on Windows 11, and checked in a fresh-install, two-PC end-to-end run.
-> The test suite also passes on Linux and macOS in CI, but nobody has used it day to day there yet.
+> **Alpha.** Windows: used daily, with full end-to-end checks (fresh install, two PCs, real
+> Claude and Codex). macOS and Linux: tests and a clean install from PyPI pass in CI on every
+> release and every day, but nobody has used them day to day yet; consider them experimental.
 
 ---
 
@@ -99,8 +100,8 @@ graph-mind-install
 
 If `graph-mind-install` is "not recognized", pip put it in a Scripts folder that is not on your
 PATH (common with the Windows Python install manager). `python -m install` runs the same thing.
-If it stops with `WinError 1114` loading `c10.dll`, Windows is missing the Microsoft Visual C++
-runtime that PyTorch needs: install [vc_redist.x64.exe](https://aka.ms/vs/17/release/vc_redist.x64.exe),
+If it stops with a DLL load error, Windows is missing the Microsoft Visual C++ runtime that the
+embedding runtime needs: install [vc_redist.x64.exe](https://aka.ms/vs/17/release/vc_redist.x64.exe),
 restart, and run the installer again.
 
 or from source:
@@ -113,12 +114,12 @@ python install.py
 
 The installer:
 
-- installs the packages (PyTorch CPU is the large one, about 2 GB);
+- installs the packages (no PyTorch: the embedding model runs on ONNX Runtime);
 - registers the MCP server with every app it finds: Claude Code, Claude Desktop (including the
   Microsoft Store build) and Codex;
 - starts the capture service at login (Windows Startup folder, a macOS LaunchAgent, or an XDG
   autostart entry on Linux);
-- downloads the embedding model.
+- downloads the embedding model (about 470 MB, once).
 
 Then restart your AI apps. The server is also listed in the
 [MCP Registry](https://registry.modelcontextprotocol.io) as `io.github.goyohan0611-png/graph-mind`. Running the installer again is safe. Run it again if you move the folder.
@@ -232,8 +233,9 @@ python -m unittest discover -p "test_*.py"
 
 ## Known limits
 
-- **The first question after an AI app starts waits for the embedding model to load** (a few
-  seconds; 30-50 s on a slow or synced disk). Later questions take well under a second.
+- **For the first few seconds after an AI app starts, answers use word search only** while the
+  embedding model loads in the background (6-9 s on a slow, synced disk; it was 50-70 s before
+  0.4). The first answer never waits for it; later answers take well under a second.
 - **Daily use on Windows only so far.** macOS and Linux pass the tests in CI but have not seen real
   use.
 - **Capture follows each app's transcript format,** which is not a public interface. An app update

{graph_mind_memory-0.3.3 → graph_mind_memory-0.4.0}/graph_mind_mcp_server.py

@@ -218,7 +218,15 @@ def semantic_turns(database, query, *, scopes=None, since=None, until=None, limi
     # new pieces are embedded inline; a backlog is filled in the background while this question
     # is answered from what is already embedded (word search covers the rest meanwhile).
     missing = [p for p in pieces if embedder._key(p.strip()[:2000]) not in embedder.cache]
-    if len(missing) > INLINE_EMBED_LIMIT:
+    # Nor wait for the model itself: an app's first question used to sit through the torch import
+    # (38.8 s measured). If it is not loaded within WAIT_FOR_MODEL seconds, this question is
+    # answered from word search plus what is embedded already, and loading carries on.
+    asked = embedder._key(query.strip()[:2000]) in embedder.cache
+    wait = getattr(embedder, "WAIT_FOR_MODEL", None)
+    loading = (wait is not None and (missing or not asked) and not embedder.ready(wait))
+    if loading and not asked:
+        return []
+    if len(missing) > INLINE_EMBED_LIMIT or (loading and missing):
         embed_in_background(embedder, missing)
         keep = [i for i, p in enumerate(pieces)
                 if embedder._key(p.strip()[:2000]) in embedder.cache]
@@ -741,6 +749,8 @@ def preload(database):
 
 
 def main():
+    from local_embedder import LocalEmbedder
+    LocalEmbedder.WAIT_FOR_MODEL = 0.5      # an app's question never waits out a model load
     preload(configured_db_path())
     server.run(transport="stdio")
 

{graph_mind_memory-0.3.3 → graph_mind_memory-0.4.0/graph_mind_memory.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: graph-mind-memory
-Version: 0.3.3
+Version: 0.4.0
 Summary: Local-first memory for AI coding assistants: verbatim, on your own PC, no LLM calls at write time.
 Author-email: Yohan Ko <goyohan0611@gmail.com>
 License-Expression: AGPL-3.0-only
@@ -13,11 +13,11 @@ Classifier: Operating System :: OS Independent
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown
 License-File: LICENSE
-Requires-Dist: torch<3,>=2.2
 Requires-Dist: mcp<3,>=2
 Requires-Dist: numpy
-Requires-Dist: transformers
-Requires-Dist: tiktoken
+Requires-Dist: onnxruntime>=1.17
+Requires-Dist: tokenizers>=0.15
+Requires-Dist: huggingface_hub>=0.20
 Requires-Dist: pgserver; python_version < "3.13"
 Requires-Dist: psycopg[binary]
 Requires-Dist: psutil
@@ -48,8 +48,9 @@ Local-first memory for AI coding assistants. Verbatim storage, on your own PC:
 </div>
 
 > [!NOTE]
-> **Alpha.** Used daily on Windows 11, and checked in a fresh-install, two-PC end-to-end run.
-> The test suite also passes on Linux and macOS in CI, but nobody has used it day to day there yet.
+> **Alpha.** Windows: used daily, with full end-to-end checks (fresh install, two PCs, real
+> Claude and Codex). macOS and Linux: tests and a clean install from PyPI pass in CI on every
+> release and every day, but nobody has used them day to day yet; consider them experimental.
 
 ---
 
@@ -125,8 +126,8 @@ graph-mind-install
 
 If `graph-mind-install` is "not recognized", pip put it in a Scripts folder that is not on your
 PATH (common with the Windows Python install manager). `python -m install` runs the same thing.
-If it stops with `WinError 1114` loading `c10.dll`, Windows is missing the Microsoft Visual C++
-runtime that PyTorch needs: install [vc_redist.x64.exe](https://aka.ms/vs/17/release/vc_redist.x64.exe),
+If it stops with a DLL load error, Windows is missing the Microsoft Visual C++ runtime that the
+embedding runtime needs: install [vc_redist.x64.exe](https://aka.ms/vs/17/release/vc_redist.x64.exe),
 restart, and run the installer again.
 
 or from source:
@@ -139,12 +140,12 @@ python install.py
 
 The installer:
 
-- installs the packages (PyTorch CPU is the large one, about 2 GB);
+- installs the packages (no PyTorch: the embedding model runs on ONNX Runtime);
 - registers the MCP server with every app it finds: Claude Code, Claude Desktop (including the
   Microsoft Store build) and Codex;
 - starts the capture service at login (Windows Startup folder, a macOS LaunchAgent, or an XDG
   autostart entry on Linux);
-- downloads the embedding model.
+- downloads the embedding model (about 470 MB, once).
 
 Then restart your AI apps. The server is also listed in the
 [MCP Registry](https://registry.modelcontextprotocol.io) as `io.github.goyohan0611-png/graph-mind`. Running the installer again is safe. Run it again if you move the folder.
@@ -258,8 +259,9 @@ python -m unittest discover -p "test_*.py"
 
 ## Known limits
 
-- **The first question after an AI app starts waits for the embedding model to load** (a few
-  seconds; 30-50 s on a slow or synced disk). Later questions take well under a second.
+- **For the first few seconds after an AI app starts, answers use word search only** while the
+  embedding model loads in the background (6-9 s on a slow, synced disk; it was 50-70 s before
+  0.4). The first answer never waits for it; later answers take well under a second.
 - **Daily use on Windows only so far.** macOS and Linux pass the tests in CI but have not seen real
   use.
 - **Capture follows each app's transcript format,** which is not a public interface. An app update

{graph_mind_memory-0.3.3 → graph_mind_memory-0.4.0}/graph_mind_memory.egg-info/requires.txt

@@ -1,8 +1,8 @@
-torch<3,>=2.2
 mcp<3,>=2
 numpy
-transformers
-tiktoken
+onnxruntime>=1.17
+tokenizers>=0.15
+huggingface_hub>=0.20
 psycopg[binary]
 psutil
 fasteners

{graph_mind_memory-0.3.3 → graph_mind_memory-0.4.0}/install.py

@@ -180,35 +180,32 @@ def capture_service(start: bool) -> str:
 
 def embedding_model():
     sys.path.insert(0, str(HERE))
-    from local_embedder import LOCAL_MODEL, own_hf_home
+    from local_embedder import LOCAL_MODEL, MODEL_FILES, model_paths, own_hf_home
     try:
-        from transformers import AutoModel, AutoTokenizer
-    except OSError as error:
-        if "1114" not in str(error):
+        import onnxruntime  # noqa: F401  (a DLL failure here would leave search on words only)
+    except ImportError as error:
+        if platform.system() != "Windows":
             raise
-        # WinError 1114 on c10.dll: PyTorch needs the Microsoft Visual C++ runtime, which a fresh
-        # Windows often lacks. Everything else is installed; recall uses word search until then.
-        sys.exit(f"\n{error}\n\nPyTorch could not load. Install the Microsoft Visual C++ "
-                 "Redistributable (x64), restart, and run this again:\n"
+        sys.exit(f"\n{error}\n\nThe embedding runtime could not load. Install the Microsoft "
+                 "Visual C++ Redistributable (x64), restart, and run this again:\n"
                  "    https://aka.ms/vs/17/release/vc_redist.x64.exe\n"
                  "The apps are already connected; until then recall works by word search only.")
-
     try:
-        AutoTokenizer.from_pretrained(LOCAL_MODEL)
-        AutoModel.from_pretrained(LOCAL_MODEL)
+        model_paths()
     except OSError as error:
         # ~/.cache refused (access denied): move Hugging Face's whole home under ~/.graph-mind,
         # where every Graph-MIND process looks from now on (local_embedder.use_own_hf_home)
-        if "PermissionError" not in str(error) and not isinstance(error.__cause__, PermissionError):
+        if "PermissionError" not in str(error) and not isinstance(error.__cause__, PermissionError) \
+                and not isinstance(error, PermissionError):
             raise
         own = own_hf_home()
         own.mkdir(parents=True, exist_ok=True)
         step(f"the usual model folder is not writable; using {own}")
         # huggingface_hub fixed its folders when it was imported: download in a fresh process
         subprocess.run([sys.executable, "-c",
-                        "from transformers import AutoModel, AutoTokenizer; "
-                        f"AutoTokenizer.from_pretrained({LOCAL_MODEL!r}); "
-                        f"AutoModel.from_pretrained({LOCAL_MODEL!r})"],
+                        "from huggingface_hub import hf_hub_download; "
+                        + "; ".join(f"hf_hub_download({LOCAL_MODEL!r}, {name!r})"
+                                    for name in MODEL_FILES)],
                        env={**os.environ, "HF_HOME": str(own),
                             "HF_HUB_DISABLE_SYMLINKS_WARNING": "1"}, check=True)
 

{graph_mind_memory-0.3.3 → graph_mind_memory-0.4.0}/local_embedder.py

@@ -36,20 +36,43 @@ def use_own_hf_home() -> None:
 use_own_hf_home()
 
 
+MODEL_FILES = ("tokenizer.json", "onnx/model.onnx")
+
+
 def model_on_disk() -> bool:
+    """Both files the embedder reads are already downloaded (an install from before 0.4 has the
+    torch weights in the same folder, but not the ONNX file)."""
     hf_home = Path(os.environ.get("HF_HOME") or Path.home() / ".cache" / "huggingface")
-    return (hf_home / "hub" / ("models--" + LOCAL_MODEL.replace("/", "--"))).is_dir()
+    snapshots = hf_home / "hub" / ("models--" + LOCAL_MODEL.replace("/", "--")) / "snapshots"
+    return any(all((snapshot / name).exists() for name in MODEL_FILES)
+               for snapshot in (snapshots.iterdir() if snapshots.is_dir() else []))
+
+
+def model_paths(model=LOCAL_MODEL) -> tuple:
+    """Local paths of the tokenizer and the ONNX model, downloaded on first use."""
+    from huggingface_hub import hf_hub_download
+    return tuple(hf_hub_download(model, name) for name in MODEL_FILES)
 
 
 class LocalEmbedder:
-    """On-device multilingual embeddings via raw transformers (mean pooling, L2-normalized).
-    No text leaves the machine — required by the user-owned local-brain promise."""
+    """On-device multilingual embeddings: the model's own ONNX export run by ONNX Runtime,
+    mean pooling, L2-normalized. No text leaves the machine.
+
+    Until 0.4 this ran on torch + transformers: same model, same vectors (on 1,000 texts the
+    largest difference was 2.4e-7, cosine 1.000000), but a 2 GB install, the Visual C++ runtime,
+    and 48-71 s to load on a OneDrive-synced Windows PC against 6-9 s now."""
 
-    _model = None
-    _tok = None
+    _model = None            # an onnxruntime.InferenceSession
+    _tok = None              # a tokenizers.Tokenizer
     _lock = threading.RLock()  # model load + cache writes are not thread-safe
 
     FLUSH_EVERY = 5000   # vectors between writes: bounds lost work without rewriting per call
+    # Seconds a question may wait for the model to load before it is answered by word search
+    # alone. None waits (benchmarks, tests, the capture service); the MCP server sets 0.5: loading
+    # torch took 38.8 s for an app's first question on a OneDrive-synced Windows PC; now 2.4 s.
+    WAIT_FOR_MODEL = None
+    _loaded = threading.Event()
+    _loading = threading.Lock()
 
     def __init__(self, cache_path, model=LOCAL_MODEL):
         # float32 rows in a sibling directory: the old JSON cache turned 1.1GB on disk into 3-4GB
@@ -63,26 +86,52 @@ class LocalEmbedder:
 
     def _load(self):
         if LocalEmbedder._model is None:
-            import torch
-            from transformers import AutoModel, AutoTokenizer
-            LocalEmbedder._tok = AutoTokenizer.from_pretrained(self.model_name)
-            LocalEmbedder._model = AutoModel.from_pretrained(self.model_name).eval()
-            LocalEmbedder._torch = torch
+            import onnxruntime
+            from tokenizers import Tokenizer
+            tokenizer_file, model_file = model_paths(self.model_name)
+            tokenizer = Tokenizer.from_file(tokenizer_file)
+            tokenizer.enable_truncation(256)          # as transformers did: max_length=256
+            tokenizer.enable_padding()
+            options = onnxruntime.SessionOptions()
+            options.log_severity_level = 3            # warnings about unused initializers: noise
+            LocalEmbedder._tok = tokenizer
+            LocalEmbedder._model = onnxruntime.InferenceSession(
+                model_file, options, providers=["CPUExecutionProvider"])
+            LocalEmbedder._loaded.set()
+
+    def ready(self, timeout) -> bool:
+        """Whether the model is loaded, waiting at most `timeout` seconds. The first call starts
+        loading it in the background, so a later question finds it ready."""
+        if LocalEmbedder._model is not None:
+            return True
+        if LocalEmbedder._loading.acquire(blocking=False):
+            def load():
+                try:
+                    with LocalEmbedder._lock:
+                        self._load()
+                except Exception:
+                    pass                     # no model on disk: word search keeps answering
+                finally:
+                    LocalEmbedder._loading.release()
+            threading.Thread(target=load, name="graph-mind-model", daemon=True).start()
+        return LocalEmbedder._loaded.wait(timeout)
 
     def _key(self, text):
         return hashlib.sha256((self.model_name + "\0" + text).encode("utf-8")).hexdigest()
 
     def _encode(self, texts):
-        import torch
-        torch = LocalEmbedder._torch
-        batch = LocalEmbedder._tok(texts, padding=True, truncation=True,
-                                   max_length=256, return_tensors="pt")
-        with torch.no_grad():
-            out = LocalEmbedder._model(**batch)
-        mask = batch["attention_mask"].unsqueeze(-1).float()
-        emb = (out.last_hidden_state * mask).sum(1) / mask.sum(1).clamp(min=1e-9)
-        emb = torch.nn.functional.normalize(emb, dim=1)
-        return emb.tolist()
+        import numpy as np
+        encoded = LocalEmbedder._tok.encode_batch(texts)
+        ids = np.array([e.ids for e in encoded], dtype=np.int64)
+        mask = np.array([e.attention_mask for e in encoded], dtype=np.int64)
+        feed = {"input_ids": ids, "attention_mask": mask}
+        if "token_type_ids" in {i.name for i in LocalEmbedder._model.get_inputs()}:
+            feed["token_type_ids"] = np.zeros_like(ids)
+        hidden = LocalEmbedder._model.run(None, feed)[0]
+        weights = mask[..., None].astype(np.float32)
+        pooled = (hidden * weights).sum(1) / np.clip(weights.sum(1), 1e-9, None)
+        pooled /= np.clip(np.linalg.norm(pooled, axis=1, keepdims=True), 1e-12, None)
+        return pooled.tolist()
 
     def embed(self, texts):
         texts = [t.strip()[:2000] for t in texts]

{graph_mind_memory-0.3.3 → graph_mind_memory-0.4.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "graph-mind-memory"
-version = "0.3.3"
+version = "0.4.0"
 description = "Local-first memory for AI coding assistants: verbatim, on your own PC, no LLM calls at write time."
 readme = "README.md"
 requires-python = ">=3.10"
@@ -18,11 +18,11 @@ classifiers = [
     "Operating System :: OS Independent",
 ]
 dependencies = [
-    "torch>=2.2,<3",
     "mcp>=2,<3",
     "numpy",
-    "transformers",
-    "tiktoken",
+    "onnxruntime>=1.17",
+    "tokenizers>=0.15",
+    "huggingface_hub>=0.20",
     # hosts the shared brain's Postgres; built for Python 3.9-3.12 only. Joining another PC's
     # brain (psycopg) and everything else work on any supported Python.
     "pgserver; python_version < '3.13'",