File size: 16,389 Bytes
ad474c9 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 | """Agentic MCP tools held *inside* HDC storage.
A tool is a record, not a row in a database:
record = bundle([
bind(role("name"), hvc("tool/hdc.bind")),
bind(role("cap"), capability_vector("hdc/bind/construct")),
bind(role("arity"), hvc("arity/2")),
])
The record is written to storage as the cell ``toolreg/<name>``. Registration,
discovery and dispatch therefore all run on the same content-addressable
machinery as programs and data: a tool is found by *resembling* the query.
Dispatch has three tiers, each reporting its confidence:
1. exact name — confidence 1.0;
2. character-superposition similarity, so a misspelled or truncated tool name
(``hdc.bundl``) still finds ``hdc.bundle`` rather than failing;
3. capability path similarity, so ``hdc.memory.search`` reaches
``hdc.recall`` because both carry ``hdc`` + ``memory`` atoms.
Tier 2 uses the standard HDC string codec: each character becomes an atom and
its position is bound in by permutation, so two strings that match in most
positions have highly similar vectors. That is what makes fuzzy dispatch a
cleanup operation instead of an edit-distance algorithm.
The registry is what the ``TOOL`` instruction calls, and what
:mod:`hdc.mcp_server` advertises to an outside agent, so a program running
inside HDC and an MCP client reach the same tools by the same path.
"""
from __future__ import annotations
from dataclasses import dataclass, field
from typing import Callable, Sequence
from .hypervec import (
DEFAULT_DIM,
bind,
bundle,
hvc,
hv_to_int,
int_to_hv,
permute,
popcount,
sim,
)
from .memory import HDCStorage
#: Minimum fuzzy similarity accepted for a non-exact tool name.
FUZZY_FLOOR = 0.55
#: Fuzzy matches must beat the runner-up by this much to be trusted.
FUZZY_MARGIN = 0.04
_CAP_CACHE: dict[tuple[str, int], int] = {}
_STR_CACHE: dict[tuple[str, int], int] = {}
def capability_vector(capability: str, d: int = DEFAULT_DIM) -> int:
"""Superpose the path atoms of a capability (``hdc/bind/construct``)."""
key = (capability, d)
hit = _CAP_CACHE.get(key)
if hit is not None:
return hit
atoms = [hvc("cap/" + part, d) for part in capability.replace(".", "/").split("/") if part]
vec = bundle(atoms, d) if atoms else hvc("cap/unknown", d)
_CAP_CACHE[key] = vec
return vec
def string_vector(text: str, d: int = DEFAULT_DIM) -> int:
"""Character-superposition codec: order-sensitive, near-orthogonal, fuzzy.
Each character is bound to its position by permutation and the results are
bundled, so ``sim(string_vector("hdc.bundle"), string_vector("hdc.bundl"))``
is high while unrelated strings stay near the noise floor.
"""
key = (text, d)
hit = _STR_CACHE.get(key)
if hit is not None:
return hit
if not text:
vec = 0
else:
vec = bundle(
[permute(hvc("char/" + ch, d), i, d) for i, ch in enumerate(text)], d
)
_STR_CACHE[key] = vec
return vec
def _json_type(kind: str) -> str:
return "integer" if kind in ("int", "hv") else "string"
@dataclass
class ToolSpec:
"""One agentic tool.
``returns`` matters because a hypervector and a scalar are both Python
``int``s: the machine has to know whether to *encode* a result (scalar) or
store it as a vector in its own right. Without this the two are
indistinguishable at the call site.
"""
name: str
capability: str
params: tuple[tuple[str, str], ...]
doc: str
fn: Callable[..., object]
returns: str = "int"
calls: int = 0
@property
def kinds(self) -> tuple[str, ...]:
return tuple(kind for _, kind in self.params)
def record(self, d: int) -> int:
"""The hypervector record that represents this tool in storage."""
parts = [
bind(self._role("name", d), hvc("tool/" + self.name, d), d),
bind(self._role("cap", d), capability_vector(self.capability, d), d),
bind(self._role("arity", d), hvc("arity/%d" % len(self.params), d), d),
]
return bundle(parts, d)
@staticmethod
def _role(name: str, d: int) -> int:
from .hypervec import role
return role(name, d)
def input_schema(self) -> dict:
"""MCP ``inputSchema`` built from the declared parameter kinds."""
props = {}
for pname, kind in self.params:
props[pname] = {
"type": _json_type(kind),
"description": (
"hypervector: pass the integer, or a hex string like 0x1f3a"
if kind == "hv"
else "scalar integer"
),
}
return {
"type": "object",
"properties": props,
"required": [pname for pname, _ in self.params],
"additionalProperties": False,
}
class ToolRegistry:
"""Tool records in HDC storage, plus the dispatcher that reads them."""
def __init__(self, storage: HDCStorage | None = None, d: int = DEFAULT_DIM):
self.d = d
self.storage = storage if storage is not None else HDCStorage(d=d, label="tools")
self.specs: dict[str, ToolSpec] = {}
self._records: dict[str, int] = {}
self.last_machine = None
self.install_builtins()
# ---------------------------------------------------------- registration
def register(
self,
name: str,
capability: str,
params: Sequence[tuple[str, str]],
doc: str,
fn: Callable[..., object],
returns: str = "int",
store: bool = True,
) -> ToolSpec:
"""Add a tool and write its record into HDC storage."""
spec = ToolSpec(name, capability, tuple(params), doc, fn, returns=returns)
self.specs[name] = spec
rec = spec.record(self.d)
self._records[name] = rec
if store:
self.storage.write("toolreg/" + name, rec, symbol=True)
return spec
def unregister(self, name: str) -> None:
self.specs.pop(name, None)
self._records.pop(name, None)
if "toolreg/" + name in self.storage:
self.storage.cells.pop("toolreg/" + name, None)
self.storage.order.remove("toolreg/" + name)
# -------------------------------------------------------------- lookup
def get(self, name: str) -> ToolSpec | None:
return self.specs.get(name)
def params_for(self, name: str) -> tuple[str, ...]:
spec, _ = self.resolve(name)
if spec is None:
raise ToolError("unknown tool: %r" % (name,))
return spec.kinds
def returns_for(self, name: str) -> str:
spec, _ = self.resolve(name)
if spec is None:
raise ToolError("unknown tool: %r" % (name,))
return spec.returns
def resolve(self, query: str | int) -> tuple[ToolSpec | None, float]:
"""Find a tool by name, by fuzzy name, or by capability resemblance."""
if isinstance(query, int):
best_name, best = "", -2.0
for name, rec in self._records.items():
s = sim(rec, query, self.d)
if s > best:
best_name, best = name, s
return (self.specs.get(best_name), best) if best >= 0.3 else (None, best)
if query in self.specs:
return self.specs[query], 1.0
probe = string_vector(query, self.d)
scored = sorted(
((name, sim(probe, string_vector(name, self.d), self.d)) for name in self.specs),
key=lambda t: t[1],
reverse=True,
)
if scored and scored[0][1] >= FUZZY_FLOOR:
margin = scored[0][1] - (scored[1][1] if len(scored) > 1 else 0.0)
if margin >= FUZZY_MARGIN or len(scored) == 1:
return self.specs[scored[0][0]], scored[0][1]
cap_probe = capability_vector(query, self.d)
scored = sorted(
(
(name, sim(cap_probe, capability_vector(s.capability, self.d), self.d))
for name, s in self.specs.items()
),
key=lambda t: t[1],
reverse=True,
)
if scored and scored[0][1] >= 0.30:
return self.specs[scored[0][0]], scored[0][1]
return None, (scored[0][1] if scored else 0.0)
# --------------------------------------------------------------- calling
def call(self, name: str, args: Sequence[object], machine=None) -> object:
"""Invoke a tool. Raises ``ToolError`` for an unresolvable tool."""
spec, confidence = self.resolve(name)
if spec is None:
raise ToolError(
"no tool resembling %r (best similarity %.3f)" % (name, confidence)
)
if len(args) != len(spec.params):
raise ToolError(
"tool %s takes %d argument(s), got %d"
% (spec.name, len(spec.params), len(args))
)
self.last_machine = machine
spec.calls += 1
return spec.fn(*args)
def call_named(self, name: str, arguments: dict, machine=None) -> object:
"""Invoke a tool with named arguments (the MCP calling convention)."""
spec, confidence = self.resolve(name)
if spec is None:
raise ToolError("unknown tool %r (%.3f)" % (name, confidence))
args = []
for pname, kind in spec.params:
if pname not in arguments:
raise ToolError("missing argument %r for %s" % (pname, spec.name))
value = arguments[pname]
if kind in ("int", "hv") and isinstance(value, str):
value = int(value, 0)
if kind == "int":
value = int(value)
elif kind == "hv":
value = int(value)
args.append(value)
return self.call(spec.name, args, machine=machine)
# ---------------------------------------------------------------- views
def list_tools(self) -> list[dict]:
"""MCP ``tools/list`` payload."""
out = []
for name in sorted(self.specs):
spec = self.specs[name]
out.append(
{
"name": spec.name,
"description": spec.doc,
"inputSchema": spec.input_schema(),
"annotations": {"hdc_capability": spec.capability},
}
)
return out
def specs_text(self) -> str:
rows = ["%-16s %-26s %-5s %s" % ("TOOL", "CAPABILITY", "ARGS", "DESCRIPTION")]
rows.append("-" * 96)
for name in sorted(self.specs):
s = self.specs[name]
rows.append(
"%-16s %-26s %-5d %s"
% (s.name, s.capability, len(s.params), s.doc)
)
return "\n".join(rows)
def memory_report(self) -> str:
"""Show the tool records as they sit in storage."""
rows = ["tool records in %r (d=%d)" % (self.storage.label, self.d)]
for name in sorted(self.specs):
cell = "toolreg/" + name
present = cell in self.storage
hv = self.storage.read(cell, default=0)
rows.append(
" %-16s cell=%-18s pop=%-5d %s"
% (name, cell, popcount(hv), "stored" if present else "MISSING")
)
return "\n".join(rows)
# ------------------------------------------------------------- builtins
def install_builtins(self) -> None:
"""Register the built-in HDC / bit-logic / BitNet tools."""
d = self.d
store = self.storage
def hdc_similarity(a: int, b: int) -> int:
return int(round(sim(a, b, d) * 1000))
def hdc_bind(a: int, b: int) -> int:
return bind(a, b, d)
def hdc_bundle(a: int, b: int) -> int:
return bundle([a, b], d)
def hdc_permute(v: int, k: int) -> int:
return permute(v, k, d)
def hdc_cleanup(v: int) -> int:
hits = store.nearest_symbol(v, k=1)
if not hits:
return -1
syms = store.symbols()
return syms.index(hits[0][0]) if hits[0][0] in syms else -1
def hdc_recall(v: int, k: int) -> int:
hits = store.recall(v, k=max(1, k))
if not hits:
return -1
return store.order.index(hits[0][0]) if hits[0][0] in store.order else -1
def hdc_encode_int(n: int) -> int:
return int_to_hv(n, d)
def hdc_decode_int(v: int) -> int:
return hv_to_int(v, d)
def bit_popcount(v: int) -> int:
return popcount(v)
def bit_hamming(a: int, b: int) -> int:
return (a ^ b).bit_count()
def net_dot1bit(a: int, b: int) -> int:
# A 1-bit dot product is bind-then-count: agree - disagree.
return d - 2 * (a ^ b).bit_count()
def mem_stat() -> int:
return int(store.stat()["cells_used"])
# HDC-native arithmetic, exposed to agents: a ripple-carry circuit over
# bit atoms rather than a decode-and-add.
from .adder import HyperAdder
adder = HyperAdder(d=d)
def hdc_add(a: int, b: int) -> int:
return adder.add_hv(a, b)
def hdc_sub(a: int, b: int) -> int:
return adder.sub_hv(a, b)
def agent_echo(n: int) -> int:
return int(n)
self.register("hdc.similarity", "hdc/bind/similarity", (("a", "hv"), ("b", "hv")),
"Similarity of two hypervectors, scaled to milli-units.", hdc_similarity)
self.register("hdc.bind", "hdc/bind/construct", (("a", "hv"), ("b", "hv")),
"XOR-bind two hypervectors.", hdc_bind, returns="hv")
self.register("hdc.unbind", "hdc/bind/invert", (("a", "hv"), ("b", "hv")),
"Unbind: XOR is self-inverse, so this is bind.", hdc_bind, returns="hv")
self.register("hdc.bundle", "hdc/bundle/construct", (("a", "hv"), ("b", "hv")),
"Majority-bundle two hypervectors.", hdc_bundle, returns="hv")
self.register("hdc.permute", "hdc/order/rotate", (("v", "hv"), ("k", "int")),
"Rotate a hypervector by k bits.", hdc_permute, returns="hv")
self.register("hdc.cleanup", "hdc/memory/cleanup", (("v", "hv"),),
"Index of the stored symbol nearest to a hypervector.", hdc_cleanup)
self.register("hdc.recall", "hdc/memory/search", (("v", "hv"), ("k", "int")),
"Index of the stored cell most similar to a hypervector.", hdc_recall)
self.register("hdc.encode_int", "hdc/codec/encode", (("n", "int"),),
"Encode an integer as a hypervector.", hdc_encode_int, returns="hv")
self.register("hdc.decode_int", "hdc/codec/decode", (("v", "hv"),),
"Decode an integer hypervector.", hdc_decode_int)
self.register("bit.popcount", "bit/logic/count", (("v", "hv"),),
"Population count of a hypervector.", bit_popcount)
self.register("bit.hamming", "bit/logic/distance", (("a", "hv"), ("b", "hv")),
"Hamming distance between two hypervectors.", bit_hamming)
self.register("net.dot1bit", "bitnet/inference/dot", (("a", "hv"), ("b", "hv")),
"1-bit dot product as agreement minus disagreement.", net_dot1bit)
self.register("mem.stat", "storage/memory/stats", (),
"Number of occupied HDC cells.", mem_stat)
self.register("hdc.add", "hdc/arithmetic/add", (("a", "hv"), ("b", "hv")),
"Add two integer hypervectors with a ripple-carry HDC circuit.",
hdc_add, returns="hv")
self.register("hdc.sub", "hdc/arithmetic/subtract", (("a", "hv"), ("b", "hv")),
"Subtract two integer hypervectors (two's complement, HDC circuit).",
hdc_sub, returns="hv")
self.register("agent.echo", "agent/plumbing/echo", (("n", "int"),),
"Return the argument unchanged; useful for wiring tests.", agent_echo)
class ToolError(Exception):
"""A tool could not be resolved or invoked."""
|