File size: 18,493 Bytes
a484e22 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 | """Three tool-call-memory arms.
An arm exposes two operations:
- draft(query, functions, user_id) -> canonical draft string
- observe(query, functions, user_id, target_name, target_args) -> None
record the genuine target so future drafts can reuse it.
Arms:
1. NoMemory -- schema-only draft (best you can do with zero history).
2. StaticGlobal -- ToolSpec-style: one global datastore, built during warmup
and then FROZEN (no growth, no eviction, no per-user view).
3. PersonalMemory (ours) -- per-user store that grows across sessions, with an
eviction policy (LRU or LFU) capping per-user size, and
personalized retrieval (query only the user's own store,
backing off to a small shared store when empty).
Retrieval is top-1 cosine similarity over sentence-transformer embeddings of the
query text. The drafted call is the canonicalized (name, arguments) of the most
similar past observation.
"""
from __future__ import annotations
import re
from collections import OrderedDict
from dataclasses import dataclass, field
from typing import Any
import numpy as np
from .metrics import canonical_call_str
# --------------------------------------------------------------------------- #
# Embedding backend
# --------------------------------------------------------------------------- #
class Embedder:
def __init__(self, name: str = "sentence-transformers/all-MiniLM-L6-v2"):
from sentence_transformers import SentenceTransformer
self.model = SentenceTransformer(name, device="cpu")
self._cache: dict[str, np.ndarray] = {}
def embed(self, text: str) -> np.ndarray:
v = self._cache.get(text)
if v is None:
v = self.model.encode(text, normalize_embeddings=True)
v = np.asarray(v, dtype=np.float32)
self._cache[text] = v
return v
def schema_draft(functions: list[dict[str, Any]]) -> str:
"""Zero-history draft: the first tool's signature with placeholder args."""
if not functions:
return ""
f = functions[0]
props = (f.get("parameters", {}) or {}).get("properties", {}) or {}
required = (f.get("parameters", {}) or {}).get("required", []) or list(props)
args = {k: None for k in required}
return canonical_call_str(f["name"], args)
# --------------------------------------------------------------------------- #
# Datastore entry
# --------------------------------------------------------------------------- #
@dataclass
class Entry:
emb: np.ndarray
call: str # canonical target call string
freq: int = 1 # times this (query-region) has been reinforced
def _best_match(emb: np.ndarray, entries: list[Entry]) -> tuple[int, float]:
if not entries:
return -1, -1.0
mat = np.stack([e.emb for e in entries]) # (N, d), rows unit-norm
sims = mat @ emb # cosine (emb unit-norm)
i = int(np.argmax(sims))
return i, float(sims[i])
# --------------------------------------------------------------------------- #
# Arms
# --------------------------------------------------------------------------- #
class NoMemory:
name = "no_memory"
def draft(self, query, functions, user_id, embedder) -> str:
return schema_draft(functions)
def observe(self, *a, **k):
return None
class StaticGlobal:
"""ToolSpec-style frozen global datastore."""
name = "static_global"
def __init__(self):
self.entries: list[Entry] = []
self.frozen = False
def freeze(self):
self.frozen = True
def draft(self, query, functions, user_id, embedder) -> str:
emb = embedder.embed(query)
i, _ = _best_match(emb, self.entries)
if i < 0:
return schema_draft(functions)
return self.entries[i].call
def observe(self, query, functions, user_id, name, args, embedder) -> None:
if self.frozen:
return
self.entries.append(Entry(embedder.embed(query),
canonical_call_str(name, args)))
class PersonalMemory:
"""Ours: per-user growing store + eviction + personalization."""
name = "personal_memory"
def __init__(self, capacity: int = 32, eviction: str = "lru",
sim_threshold: float = 0.35):
self.capacity = capacity
self.eviction = eviction # "lru" | "lfu"
self.sim_threshold = sim_threshold
# per-user ordered store (insertion/most-recent-use order for LRU)
self.stores: dict[str, "OrderedDict[int, Entry]"] = {}
self.shared: list[Entry] = [] # cold-start backoff
self._next_id = 0
def _store(self, user_id) -> "OrderedDict[int, Entry]":
return self.stores.setdefault(user_id, OrderedDict())
def draft(self, query, functions, user_id, embedder) -> str:
emb = embedder.embed(query)
store = self._store(user_id)
entries = list(store.values())
i, sim = _best_match(emb, entries)
if i >= 0 and sim >= self.sim_threshold:
key = list(store.keys())[i]
entry = store[key]
if self.eviction == "lru": # mark most-recently-used
store.move_to_end(key)
entry.freq += 1
return entry.call
# personal store empty / too dissimilar -> shared backoff
j, sj = _best_match(emb, self.shared)
if j >= 0 and sj >= self.sim_threshold:
return self.shared[j].call
return schema_draft(functions)
def observe(self, query, functions, user_id, name, args, embedder) -> None:
emb = embedder.embed(query)
call = canonical_call_str(name, args)
store = self._store(user_id)
eid = self._next_id
self._next_id += 1
store[eid] = Entry(emb, call)
store.move_to_end(eid)
self._evict(store)
def seed_shared(self, query, name, args, embedder) -> None:
self.shared.append(Entry(embedder.embed(query),
canonical_call_str(name, args)))
def _evict(self, store: "OrderedDict[int, Entry]") -> None:
while len(store) > self.capacity:
if self.eviction == "lru":
store.popitem(last=False) # drop least-recently-used
elif self.eviction == "lfu":
k = min(store, key=lambda x: store[x].freq)
del store[k]
else:
store.popitem(last=False)
def total_entries(self) -> int:
return sum(len(s) for s in self.stores.values())
# --------------------------------------------------------------------------- #
# Ablation arms (2x2: personalized? x evicting?) -- see run_ablation.py.
# These decompose PersonalMemory's gain over StaticGlobal into the contribution
# of per-user partitioning vs. the contribution of online-growth+eviction.
# --------------------------------------------------------------------------- #
class PersonalNoEvict(PersonalMemory):
"""[+personalization, -eviction]: per-user store that grows online but is
never bounded/evicted. Isolates how much of ours' gain is eviction."""
name = "personal_noevict"
def __init__(self, sim_threshold: float = 0.35):
# capacity = +inf so _evict() never fires.
super().__init__(capacity=10**9, eviction="lru",
sim_threshold=sim_threshold)
class GlobalEvict:
"""[-personalization, +online-growth+eviction]: a single GLOBAL store (not
per-user) that keeps ingesting after warmup and LRU-evicts at a total
capacity. Isolates how much of ours' gain is per-user partitioning: it
differs from PersonalMemory only in that retrieval ignores user id."""
name = "global_evict"
def __init__(self, capacity: int = 1920, sim_threshold: float = 0.35):
self.capacity = capacity
self.sim_threshold = sim_threshold
self.store: "OrderedDict[int, Entry]" = OrderedDict()
self._next_id = 0
def draft(self, query, functions, user_id, embedder) -> str:
emb = embedder.embed(query)
entries = list(self.store.values())
i, sim = _best_match(emb, entries)
if i >= 0 and sim >= self.sim_threshold:
key = list(self.store.keys())[i]
self.store.move_to_end(key) # LRU touch
self.store[key].freq += 1
return self.store[key].call
return schema_draft(functions)
def observe(self, query, functions, user_id, name, args, embedder) -> None:
emb = embedder.embed(query)
eid = self._next_id
self._next_id += 1
self.store[eid] = Entry(emb, canonical_call_str(name, args))
self.store.move_to_end(eid)
while len(self.store) > self.capacity:
self.store.popitem(last=False) # drop least-recently-used
def seed_shared(self, *a, **k):
return None
# --------------------------------------------------------------------------- #
# Phase 4.4 — faithful stronger ToolSpec-style baseline
# --------------------------------------------------------------------------- #
def _schema_scaffold(functions, name) -> str:
"""Schema-aware structural draft for a NAMED function: the function's
required argument keys in canonical (sorted) order with placeholder values.
This is the structurally-valid fallback a schema-aware FSM emits when
retrieval is not confident enough to commit a concrete prior call."""
fn = next((f for f in functions if f.get("name") == name), None)
if fn is None:
return schema_draft(functions)
params = fn.get("parameters", {}) or {}
props = params.get("properties", {}) or {}
required = params.get("required", []) or list(props)
return canonical_call_str(name, {k: None for k in required})
class ToolSpecBaseline:
"""Faithful ToolSpec reproduction (Xia et al., 2026): a *frozen global*
retrieval store (no eviction, no personalization — the ToolSpec regime)
with the two ToolSpec mechanisms the simple ``StaticGlobal`` proxy omits:
1. **Confidence-gated retrieval.** Return the nearest stored call verbatim
only while its similarity clears ``sim_lo``; ``StaticGlobal`` instead
returns its single nearest neighbour unconditionally, so on a cold /
far query it drafts a wholly unrelated call.
2. **Schema-aware fallback (FSM surrogate).** On a cold miss, rather than
emitting a random far neighbour we emit a *structurally valid* draft
for the nearest neighbour's function (its required-arg scaffold in
canonical order) — the acceptance a schema-constrained decoder
guarantees on the call's structural tokens even without a value hit.
This makes the arm **strictly at least as strong as ``StaticGlobal``**:
identical on confident hits, better on cold misses.
ToolSpec has no public code, so the FSM is approximated by this schema-aware
scaffold. We also tested a
``k``-NN summed-similarity vote on the target *function* (retrieval-augmented
denoising); it *degraded* MAT on these traces because the highly skewed
telecom workload (one diagnostic call dominates) lets the majority function
override correct top-1 picks — reported honestly in the write-up, and NOT
used here. Everything else (frozen, global, eviction-free) matches ToolSpec
and is deliberately NOT personalized — the property under test.
"""
name = "toolspec"
def __init__(self, sim_lo: float = 0.30):
self.entries: list[Entry] = []
self.frozen = False
self.sim_lo = sim_lo
self._names: list[str] = [] # parallel function name per entry
def freeze(self):
self.frozen = True
def draft(self, query, functions, user_id, embedder) -> str:
if not self.entries:
return schema_draft(functions)
emb = embedder.embed(query)
i, sim = _best_match(emb, self.entries)
if sim >= self.sim_lo:
return self.entries[i].call # confident retrieval hit
return _schema_scaffold(functions, self._names[i]) # schema-aware miss
def observe(self, query, functions, user_id, name, args, embedder) -> None:
if self.frozen:
return
self.entries.append(Entry(embedder.embed(query),
canonical_call_str(name, args)))
self._names.append(name)
def seed_shared(self, *a, **k):
return None
# --------------------------------------------------------------------------- #
# SuffixDecoding baseline (Oliaro et al., NeurIPS 2025; arXiv 2411.04975)
# --------------------------------------------------------------------------- #
_TOK_RE = re.compile(r"\w+|[^\w\s]")
_SEP = " "
def _tokenize(text: str) -> tuple[str, ...]:
"""Word/punctuation-level tokens (lowercased). A deliberate approximation of
SuffixDecoding's model-BPE tokens: the mechanistic contrast under test is
*exact token matching vs. embedding similarity*, which this preserves; the
exact subword vocabulary is not what distinguishes the two arms."""
return tuple(_TOK_RE.findall(text.lower()))
def _suffix_key(tokens: tuple[str, ...]) -> str:
"""Separator-delimited form so whole-token substring tests never match
across partial tokens (every boundary is a _SEP)."""
return _SEP + _SEP.join(tokens) + _SEP
def _longest_suffix_match(q: tuple[str, ...], stored_key: str,
floor: int) -> int:
"""Longest k>floor such that the k-token *suffix* of the current query q is
a contiguous whole-token substring of a stored sequence (its suffix key).
Returns 0 if no suffix longer than `floor` matches. This is SuffixDecoding's
'walk the tree to the node matching the context suffix' step, adapted to our
per-request query context.
Uses binary search: the predicate ``q[-k:] is a substring of stored`` is
monotonic in k (if the k-token suffix matches, every shorter suffix does),
so we find the largest matching k in O(log|q|) containment tests rather than
O(|q|) — essential because dialogue-context queries run to hundreds of
tokens."""
best = 0
lo, hi = floor + 1, len(q)
while lo <= hi:
mid = (lo + hi) // 2
cand = _SEP + _SEP.join(q[-mid:]) + _SEP
if cand in stored_key:
best = mid
lo = mid + 1
else:
hi = mid - 1
return best
class SuffixDecodingBaseline:
"""Faithful adaptation of **SuffixDecoding** (Oliaro et al., *SuffixDecoding:
Extreme Speculative Decoding for Emerging AI Applications*, NeurIPS 2025
Spotlight; arXiv 2411.04975) as a retrieval arm.
SuffixDecoding keeps a **global suffix tree accumulated from previous
requests' token streams** (live/growing across the deployment, so request N
benefits from request N-1), matches the **suffix of the current context**
against the tree at each step, and speculates the highest-frequency
continuation with adaptive length. It is **token-level (exact match),
global, and NOT personalized**; the tree is **size-capped** (~10.75 B/token,
~31 days on a 144 GB host), not unbounded.
We reproduce that regime and swap **exactly one mechanism** vs. our own
``GlobalEvict`` arm: retrieval is **longest-token-suffix match** on the query
instead of embedding cosine similarity. Everything else — global live
write-back, canonicalization, a size cap, no per-user partitioning — is held
identical, so the comparison isolates *retrieval mechanism* (exact token
match vs. semantic embedding), not confounds like different data pools or
personalization. Frequency and recency break ties in match length, mirroring
SuffixDecoding's frequency-ranked tree paths.
Faithful vs. adapted:
- Faithful: global, live-growing, size-capped, non-personalized store;
exact token-suffix matching; frequency-ranked selection.
- Adapted: we match the *query* token context (which selects a tool call
in our one-call-per-request setting) rather than a running generation,
and our token-LCP acceptance already truncates the speculated call at the
first mismatch, subsuming SuffixDecoding's adaptive speculation length.
Tokenization is word/punct level, not the model BPE (approximation).
"""
name = "suffixdecoding"
def __init__(self, capacity: int = 1920, min_match: int = 1):
# capacity matches GlobalEvict's total footprint (U*C = 40*48) for a
# same-size comparison; min_match=1 lets any shared suffix token yield a
# (frequency-ranked) speculation, as SuffixDecoding's tree always does.
self.capacity = capacity
self.min_match = min_match
# eid -> [q_tokens, suffix_key, call, freq]
self.store: "OrderedDict[int, list]" = OrderedDict()
self._next_id = 0
def draft(self, query, functions, user_id, embedder) -> str:
q = _tokenize(query)
if not q:
return schema_draft(functions)
floor = self.min_match - 1
best_key, best_eid = None, None # rank by (match_len, freq), then recency
for eid, rec in self.store.items():
m = _longest_suffix_match(q, rec[1], floor) # true length (fixed floor)
if m < self.min_match:
continue
key = (m, rec[3])
if best_key is None or key >= best_key: # >= => later (more recent) wins ties
best_key, best_eid = key, eid
if best_eid is None:
return schema_draft(functions)
rec = self.store[best_eid]
rec[3] += 1 # frequency reinforcement
self.store.move_to_end(best_eid) # recency touch
return rec[2]
def observe(self, query, functions, user_id, name, args, embedder) -> None:
q = _tokenize(query)
eid = self._next_id
self._next_id += 1
self.store[eid] = [q, _suffix_key(q), canonical_call_str(name, args), 1]
self.store.move_to_end(eid)
while len(self.store) > self.capacity:
self.store.popitem(last=False) # drop oldest (size cap)
def seed_shared(self, *a, **k):
return None
def total_entries(self) -> int:
return len(self.store)
|