File size: 12,668 Bytes
dbb04e4 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 | """
Contextual Query Masking via XOR Attention (Phase 4.0)
======================================================
Implements an XOR-based soft attention mechanism over Binary HDV space.
How it works:
1. A "context key" is constructed by bundling recent HOT-tier vectors.
2. A XOR attention mask is generated: mask = query XOR context_key
This creates a residual vector that is ORTHOGONAL to the context,
effectively suppressing already-known dimensions and amplifying novel ones.
3. Query results are re-ranked by a composite score:
composite = alpha * raw_similarity + beta * novelty_boost(mask, mem_hdv)
4. The mask is also available for downstream gap-detection.
Motivation (VSA theory):
- XOR in binary HDV space is the self-inverse binding operator.
- query.xor(context) ≈ "what about this query is NOT already represented in context?"
- Hamming similarity(mask, candidate) ≈ novelty of candidate relative to context.
Phase 4.1: XOR-based Project Isolation
======================================
XORIsolationMask provides deterministic project-based memory isolation:
- Each project_id derives a unique binary mask via SHA256(project_id) -> seed -> RNG
- store(): masked_hdv = original_hdv XOR project_mask
- query(): unmasked_query = query_hdv XOR project_mask (then search in masked space)
- Memories from different projects are effectively orthogonal (~50% similarity)
"""
from __future__ import annotations
import hashlib
from dataclasses import dataclass, field
from typing import Dict, List, Optional, Tuple
import numpy as np
from loguru import logger
from .binary_hdv import BinaryHDV, majority_bundle
@dataclass
class AttentionConfig:
"""Tunable hyperparameters for XOR attention."""
alpha: float = 0.6 # Weight for raw similarity
beta: float = 0.4 # Weight for novelty score from XOR mask
context_sample_n: int = 50 # How many HOT nodes to include in context key
min_novelty_boost: float = 0.0 # Floor for novelty contribution
enabled: bool = True
def validate(self) -> None:
assert 0.0 <= self.alpha <= 1.0, "alpha must be in [0, 1]"
assert 0.0 <= self.beta <= 1.0, "beta must be in [0, 1]"
assert abs((self.alpha + self.beta) - 1.0) < 1e-6, "alpha + beta must equal 1.0"
@dataclass
class AttentionResult:
"""Enriched result from contextual reranking."""
node_id: str
raw_score: float
novelty_score: float
composite_score: float
attention_mask: Optional[BinaryHDV] = field(default=None, repr=False)
class XORAttentionMasker:
"""
Contextual query masking using XOR binding in binary HDV space.
Usage:
masker = XORAttentionMasker(config)
mask = masker.build_attention_mask(query_vec, context_vecs)
reranked = masker.rerank(raw_scores, memory_vectors, mask)
"""
def __init__(self, config: Optional[AttentionConfig] = None):
self.config = config or AttentionConfig()
def build_context_key(self, context_nodes_hdv: List[BinaryHDV]) -> BinaryHDV:
"""
Bundle HOT-tier vectors into a single context summary key.
Uses majority vote bundling (sum > threshold → 1, else → 0).
Falls back to zero-vector if no context is available.
"""
if not context_nodes_hdv:
return BinaryHDV.zeros(context_nodes_hdv[0].dimension if context_nodes_hdv else 16384)
return majority_bundle(context_nodes_hdv)
def build_attention_mask(
self,
query_vec: BinaryHDV,
context_key: BinaryHDV,
) -> BinaryHDV:
"""
Compute XOR attention mask: mask = query XOR context_key.
The mask represents "query minus context" — bits that are unique
to the query compared to what the system already holds in working memory.
High Hamming similarity between mask and a candidate → that candidate
is novel / peripheral relative to the current context.
"""
mask = query_vec.xor_bind(context_key)
logger.debug(
"Built XOR attention mask — "
f"query/context Hamming dist = {query_vec.normalized_distance(context_key):.4f}"
)
return mask
def novelty_score(self, mask: BinaryHDV, candidate_hdv: BinaryHDV) -> float:
"""
Calculate novelty of a candidate relative to the context.
Defined as: Hamming similarity(mask, candidate) in [0, 1].
Higher value → candidate is more "attention-worthy" given the query context.
"""
return mask.similarity(candidate_hdv)
def rerank(
self,
raw_scores: Dict[str, float],
memory_vectors: Dict[str, BinaryHDV],
mask: BinaryHDV,
) -> List[AttentionResult]:
"""
Re-rank retrieved memories using the composite XOR attention score.
Args:
raw_scores: {node_id: raw_similarity} from initial retrieval.
memory_vectors: {node_id: BinaryHDV} for novelty calculation.
mask: XOR attention mask built from query and context.
Returns:
Sorted list of AttentionResult (highest composite first).
"""
cfg = self.config
results: List[AttentionResult] = []
for node_id, raw in raw_scores.items():
hdv = memory_vectors.get(node_id)
if hdv is None:
novelty = cfg.min_novelty_boost
else:
novelty = max(self.novelty_score(mask, hdv), cfg.min_novelty_boost)
composite = cfg.alpha * raw + cfg.beta * novelty
results.append(
AttentionResult(
node_id=node_id,
raw_score=raw,
novelty_score=novelty,
composite_score=composite,
attention_mask=mask,
)
)
results.sort(key=lambda r: r.composite_score, reverse=True)
return results
def extract_scores(
self, results: List[AttentionResult]
) -> List[Tuple[str, float]]:
"""Convert AttentionResult list to the standard (node_id, score) tuple format."""
return [(r.node_id, r.composite_score) for r in results]
# ==============================================================================
# Phase 4.1: XOR-based Project Isolation
# ==============================================================================
@dataclass
class IsolationConfig:
"""Configuration for XOR-based project isolation."""
enabled: bool = True
dimension: int = 16384
def validate(self) -> None:
assert self.dimension > 0, "dimension must be positive"
assert self.dimension % 8 == 0, "dimension must be multiple of 8"
class XORIsolationMask:
"""
Deterministic XOR-based isolation mask for multi-tenant memory isolation.
Design:
-------
Each project_id derives a unique binary mask through:
SHA256(project_id) -> 256-bit digest -> seed -> np.random.Generator -> binary mask
The mask is applied via XOR binding:
- store(content, project_id="A"): masked_hdv = original_hdv XOR mask_A
- query(query_text, project_id="A"): unmasked = query_hdv XOR mask_A
Properties:
-----------
- Self-inverse: XOR twice with the same mask recovers the original vector
- Deterministic: Same project_id always produces the same mask
- Orthogonal isolation: Different projects' masks are ~50% different (random)
- No key management: project_id IS the key (no external secrets needed)
Security Model:
---------------
This provides cryptographic isolation via the one-time pad principle:
- A masked vector reveals NO information about the original without the mask
- Cross-project queries will match random noise (~50% similarity baseline)
- The isolation strength depends on the secrecy of project_ids
Usage:
------
masker = XORIsolationMask(config)
mask = masker.get_mask("project-alpha") # Deterministic mask
# Store
masked_hdv = masker.apply_mask(original_hdv, "project-alpha")
# Query (apply same mask to query to search in masked space)
masked_query = masker.apply_mask(query_hdv, "project-alpha")
# Remove mask (if needed for inspection)
original = masker.remove_mask(masked_hdv, "project-alpha")
"""
def __init__(self, config: Optional[IsolationConfig] = None):
self.config = config or IsolationConfig()
self._mask_cache: Dict[str, BinaryHDV] = {}
def _derive_seed(self, project_id: str) -> int:
"""
Derive a deterministic 64-bit seed from project_id using SHA256.
Args:
project_id: Unique project identifier string.
Returns:
64-bit integer seed for numpy's Generator.
"""
digest = hashlib.sha256(f"mnemo_isolation_v1:{project_id}".encode()).digest()
return int.from_bytes(digest[:8], byteorder="big", signed=False)
def get_mask(self, project_id: str) -> BinaryHDV:
"""
Get or create the deterministic isolation mask for a project.
The mask is cached for efficiency. Same project_id always returns
the same BinaryHDV mask.
Args:
project_id: Unique project identifier.
Returns:
BinaryHDV mask of dimension self.config.dimension.
"""
if project_id in self._mask_cache:
return self._mask_cache[project_id]
seed = self._derive_seed(project_id)
rng = np.random.default_rng(seed)
# Generate random binary mask
n_bytes = self.config.dimension // 8
mask_bytes = rng.integers(0, 256, size=n_bytes, dtype=np.uint8)
mask = BinaryHDV(data=mask_bytes, dimension=self.config.dimension)
self._mask_cache[project_id] = mask
logger.debug(f"Generated isolation mask for project '{project_id}' (seed={seed})")
return mask
def apply_mask(self, hdv: BinaryHDV, project_id: str) -> BinaryHDV:
"""
Apply project isolation mask to a vector (XOR binding).
Args:
hdv: The BinaryHDV to mask.
project_id: Project identifier for mask derivation.
Returns:
Masked BinaryHDV (original XOR project_mask).
"""
if not self.config.enabled:
return hdv
mask = self.get_mask(project_id)
return hdv.xor_bind(mask)
def remove_mask(self, masked_hdv: BinaryHDV, project_id: str) -> BinaryHDV:
"""
Remove project isolation mask from a vector (XOR is self-inverse).
Note: This is identical to apply_mask() due to XOR's self-inverse property.
Kept as a separate method for semantic clarity.
Args:
masked_hdv: The masked BinaryHDV.
project_id: Project identifier used for masking.
Returns:
Original unmasked BinaryHDV.
"""
return self.apply_mask(masked_hdv, project_id)
def clear_cache(self) -> None:
"""Clear the mask cache (useful for testing)."""
self._mask_cache.clear()
def is_isolated(
self,
hdv_a: BinaryHDV,
project_id_a: str,
hdv_b: BinaryHDV,
project_id_b: str,
threshold: float = 0.55,
) -> bool:
"""
Check if two vectors are properly isolated (different projects).
After masking, vectors from different projects should have ~50% similarity.
This method checks if the cross-project similarity is within expected bounds.
Args:
hdv_a: First (unmasked) vector.
project_id_a: First vector's project.
hdv_b: Second (unmasked) vector.
project_id_b: Second vector's project.
threshold: Maximum similarity for "isolated" (default 0.55).
Returns:
True if vectors are isolated (different projects), False otherwise.
"""
if project_id_a == project_id_b:
return False # Same project = not isolated
masked_a = self.apply_mask(hdv_a, project_id_a)
masked_b = self.apply_mask(hdv_b, project_id_b)
similarity = masked_a.similarity(masked_b)
return similarity < threshold
|