File size: 11,677 Bytes
3fbbaab | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 | #!/usr/bin/env python3
"""
corpus_cache.py -- On-disk cache of embedding vectors per skill/agent.
The intake similarity check needs an embedding per existing subject
to rank a candidate against. Re-embedding the whole corpus on every
``skill_add`` is wasteful; this module caches each embedding keyed
by (subject_id, content_sha256). Stale entries (body changed) miss
silently and are re-computed on next access.
Layout::
<root>/
<backend_key>/
_manifest.json metadata only
<subject_id>__<sha16>.npy raw float32 vector
``backend_key`` is derived from the embedder's ``name``. Switching
embedders therefore lands in a separate directory and never mixes
dimensions (ST=384 vs Ollama=768).
Race-freeness: the content hash is encoded in the filename, not
just the manifest. A concurrent ``put`` cannot replace a file under
a reader — the new vector lands at a different path.
Security:
- ``subject_id`` is validated against a strict regex (prevents
path traversal via crafted IDs like ``../evil``).
- Vector files are written atomically (``os.replace`` from a
sibling tempfile).
- Manifest writes are serialised via ``_file_lock.file_lock`` so
concurrent processes don't clobber each other's entries.
"""
from __future__ import annotations
import hashlib
import json
import os
import re
import shutil
import tempfile
from dataclasses import dataclass
from pathlib import Path
from typing import Iterator
import numpy as np
from ctx.utils._file_lock import file_lock
DEFAULT_CACHE_ROOT = Path(os.path.expanduser("~/.claude/skills/_embeddings"))
# Same name policy as skill_telemetry: alnum start, alnum/_/-/. inside,
# bounded length. Prevents path-traversal and whitespace injection
# via crafted skill/agent IDs.
_SUBJECT_ID_RE = re.compile(r"^[a-zA-Z0-9][a-zA-Z0-9_\-\.]{0,127}$")
# Backend names like "sentence-transformers:all-MiniLM-L6-v2" contain
# characters (":", "/") that are illegal on Windows, so we slugify
# before using the name as a directory component.
_BACKEND_UNSAFE_RE = re.compile(r"[^a-zA-Z0-9_\-\.]")
# 64 bits of content hash in the filename. Within a single subject_id
# the inputs are always distinct strings (we only write on content
# change), so collision probability is effectively zero.
_SHA_FILENAME_LEN = 16
def _slug_backend_key(backend_name: str) -> str:
"""Turn an ``Embedder.name`` into a filesystem-safe directory name."""
if not isinstance(backend_name, str) or not backend_name:
raise ValueError("backend_name must be a non-empty string")
slug = _BACKEND_UNSAFE_RE.sub("_", backend_name).strip("_")
if not slug or not slug[0].isalnum():
raise ValueError(f"backend name yields unsafe slug: {backend_name!r}")
return slug[:64]
def _sha256_text(text: str) -> str:
return hashlib.sha256(text.encode("utf-8")).hexdigest()
@dataclass(frozen=True)
class CacheEntry:
"""Manifest row for one cached vector."""
content_sha256: str
dim: int
class CorpusCache:
"""Per-subject embedding cache.
Instances are cheap — they only touch disk on ``get``/``put``/
``invalidate``. Safe under concurrent writers within a process
and across processes via ``file_lock`` on the manifest.
"""
def __init__(self, backend_name: str, *, root: Path | None = None) -> None:
self._root = (root or DEFAULT_CACHE_ROOT).expanduser()
self._backend_dir = self._root / _slug_backend_key(backend_name)
self._manifest_path = self._backend_dir / "_manifest.json"
@property
def backend_dir(self) -> Path:
return self._backend_dir
@property
def manifest_path(self) -> Path:
return self._manifest_path
def _ensure_dir(self) -> None:
self._backend_dir.mkdir(parents=True, exist_ok=True)
@staticmethod
def _validate_subject_id(subject_id: str) -> None:
if not isinstance(subject_id, str) or not _SUBJECT_ID_RE.match(subject_id):
raise ValueError(f"invalid subject_id: {subject_id!r}")
def _vector_path(self, subject_id: str, content_sha: str) -> Path:
# Content hash is part of the filename so readers and writers
# never contend on the same path. Old files become orphans
# and are swept on the next ``put`` for that subject.
return self._backend_dir / f"{subject_id}__{content_sha[:_SHA_FILENAME_LEN]}.npy"
def _read_manifest(self) -> dict[str, dict[str, object]]:
try:
with open(self._manifest_path, encoding="utf-8") as fh:
data = json.load(fh)
except FileNotFoundError:
return {}
except json.JSONDecodeError:
# Manifest corruption drops the entire cache metadata —
# we re-embed rather than serving wrong vectors. Vector
# files on disk become unreachable and will be cleaned
# up by the next ``put`` for each subject.
return {}
if not isinstance(data, dict):
return {}
return data
def _write_manifest(self, data: dict[str, dict[str, object]]) -> None:
self._ensure_dir()
# Atomic write via tempfile + os.replace so a crash mid-write
# never leaves a partial manifest.
fd, tmp = tempfile.mkstemp(
prefix="_manifest.", suffix=".json.tmp", dir=self._backend_dir
)
try:
with os.fdopen(fd, "w", encoding="utf-8") as fh:
json.dump(data, fh, ensure_ascii=False, sort_keys=True, indent=2)
os.replace(tmp, self._manifest_path)
except Exception:
try:
Path(tmp).unlink()
except FileNotFoundError:
pass
raise
def get(self, subject_id: str, content: str) -> np.ndarray | None:
"""Return the cached vector iff the content hash matches.
Reads the vector file directly by content-hash-encoded path.
No manifest round-trip — a concurrent ``put`` for the same
subject lands at a different filename and cannot poison the
read.
"""
self._validate_subject_id(subject_id)
path = self._vector_path(subject_id, _sha256_text(content))
try:
vec = np.load(path, allow_pickle=False)
except (FileNotFoundError, ValueError, OSError):
return None
if vec.ndim != 1:
return None
if vec.dtype != np.float32:
vec = vec.astype(np.float32, copy=False)
return vec
def put(self, subject_id: str, content: str, vector: np.ndarray) -> None:
"""Store ``vector`` for ``subject_id`` keyed by content hash.
Writes the vector to a content-addressed filename, then
updates the manifest inside a file lock. Prior vector files
for the same subject are removed (orphan sweep).
"""
self._validate_subject_id(subject_id)
if not isinstance(vector, np.ndarray) or vector.ndim != 1:
raise ValueError(f"vector must be a 1-D numpy array, got {vector!r}")
vec = np.ascontiguousarray(vector, dtype=np.float32)
sha = _sha256_text(content)
self._ensure_dir()
vec_path = self._vector_path(subject_id, sha)
# Atomic write of the vector: mkstemp in the target directory,
# np.save, os.replace. Suffix is ``.npy`` so np.save does not
# append a second extension.
fd, tmp = tempfile.mkstemp(prefix=".", suffix=".npy", dir=self._backend_dir)
try:
os.close(fd)
np.save(tmp, vec, allow_pickle=False)
os.replace(tmp, vec_path)
except Exception:
try:
Path(tmp).unlink()
except FileNotFoundError:
pass
raise
# Update manifest + sweep any previous file for this subject.
with file_lock(self._manifest_path):
manifest = self._read_manifest()
prev = manifest.get(subject_id)
prev_sha = prev.get("content_sha256") if isinstance(prev, dict) else None
manifest[subject_id] = {
"content_sha256": sha,
"dim": int(vec.shape[0]),
}
self._write_manifest(manifest)
if isinstance(prev_sha, str) and prev_sha != sha:
old = self._vector_path(subject_id, prev_sha)
try:
old.unlink()
except FileNotFoundError:
pass
def invalidate(self, subject_id: str) -> bool:
"""Remove an entry. Returns True iff something was removed."""
self._validate_subject_id(subject_id)
removed = False
with file_lock(self._manifest_path):
manifest = self._read_manifest()
entry = manifest.pop(subject_id, None)
if entry is not None:
self._write_manifest(manifest)
removed = True
sha = entry.get("content_sha256") if isinstance(entry, dict) else None
if isinstance(sha, str):
try:
self._vector_path(subject_id, sha).unlink()
except FileNotFoundError:
pass
return removed
def entry(self, subject_id: str) -> CacheEntry | None:
"""Return the manifest entry for ``subject_id`` if present."""
self._validate_subject_id(subject_id)
data = self._read_manifest().get(subject_id)
if not isinstance(data, dict):
return None
sha = data.get("content_sha256")
dim = data.get("dim")
if not isinstance(sha, str) or not isinstance(dim, int):
return None
return CacheEntry(content_sha256=sha, dim=dim)
def subjects(self) -> Iterator[str]:
"""Yield subject IDs currently in the manifest."""
return iter(sorted(self._read_manifest().keys()))
def load_all(self) -> dict[str, np.ndarray]:
"""Load every cached vector. Skips any whose ``.npy`` is missing.
Returns a plain ``dict`` — callers stacking into a matrix
should order keys explicitly before ``np.vstack`` to keep
row order stable.
"""
out: dict[str, np.ndarray] = {}
manifest = self._read_manifest()
for sid, meta in manifest.items():
if not isinstance(meta, dict):
continue
sha = meta.get("content_sha256")
if not isinstance(sha, str):
continue
try:
vec = np.load(
self._vector_path(sid, sha), allow_pickle=False
)
except (FileNotFoundError, ValueError, OSError):
continue
if vec.ndim != 1:
continue
if vec.dtype != np.float32:
vec = vec.astype(np.float32, copy=False)
out[sid] = vec
return out
def size(self) -> int:
return len(self._read_manifest())
def clear(self) -> None:
"""Wipe this backend's cache directory. Destructive by design.
Exposed because users who switch embedding models may want to
reclaim disk; the caller must be explicit.
"""
if self._backend_dir.exists():
shutil.rmtree(self._backend_dir)
|