Spaces:
Running on Zero
Running on Zero
File size: 16,650 Bytes
c8fbdf1 | 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 | #!/usr/bin/env python3
"""Constraint tracker for cross-turn memory and constraint application.
Detects user-defined constraints (word limits, formatting rules, anchors/phrases)
in turn 1 and enforces them across subsequent turns using LoRA-backed learning.
Example:
Turn 1: "For this session, keep answers under 15 words and remember the phrase cobalt anchor."
Turn 2: "What should you remember?"
Expected response: Should include "cobalt anchor" and be ≤15 words.
"""
from __future__ import annotations
import re
from dataclasses import dataclass, field
from typing import Optional, Dict, Any, List
@dataclass
class DetectedConstraint:
"""A parsed constraint from user input."""
kind: str # "word_limit", "sentence_limit", "anchor_phrase", "format_rule", etc.
value: Any # numeric (word/sentence count) or string (anchor phrase)
raw_text: str # original text where constraint was found
confidence: float = 0.95
@dataclass
class SessionConstraints:
"""Container for all constraints detected in a session."""
constraints: List[DetectedConstraint] = field(default_factory=list)
anchor_phrases: List[str] = field(default_factory=list)
word_limit: Optional[int] = None
sentence_limit: Optional[int] = None
format_rules: List[str] = field(default_factory=list)
detected_at_turn: int = 0
def to_dict(self) -> Dict[str, Any]:
"""Serialize for session storage."""
return {
"anchor_phrases": self.anchor_phrases,
"word_limit": self.word_limit,
"sentence_limit": self.sentence_limit,
"format_rules": self.format_rules,
"detected_at_turn": self.detected_at_turn,
"raw_constraints": [
{
"kind": c.kind,
"value": c.value,
"raw_text": c.raw_text,
"confidence": c.confidence
}
for c in self.constraints
]
}
@classmethod
def from_dict(cls, data: Dict[str, Any]) -> SessionConstraints:
"""Deserialize from session storage."""
sc = cls()
sc.anchor_phrases = data.get("anchor_phrases", [])
sc.word_limit = data.get("word_limit")
sc.sentence_limit = data.get("sentence_limit")
sc.format_rules = data.get("format_rules", [])
sc.detected_at_turn = data.get("detected_at_turn", 0)
# Reconstruct constraints
for c in data.get("raw_constraints", []):
sc.constraints.append(DetectedConstraint(
kind=c.get("kind"),
value=c.get("value"),
raw_text=c.get("raw_text"),
confidence=c.get("confidence", 0.95)
))
return sc
class ConstraintDetector:
"""Detect constraints from user input."""
# Patterns for detecting various constraint types
WORD_LIMIT_PATTERNS = [
r"keep\s+answers?\s+(?:under|below|within|to)\s+(\d+)\s+words?",
r"(?:answer|respond)\s+in\s+(?:under|fewer than)\s+(\d+)\s+words?",
r"(\d+)\s+words?\s+(?:max|maximum|or\s+less)",
r"limit\s+(?:your\s+)?answers?\s+to\s+(\d+)\s+words?",
]
SENTENCE_LIMIT_PATTERNS = [
r"keep\s+(?:answers?|responses?)\s+to\s+(\d+)\s+sentences?",
r"(?:answer|respond)\s+in\s+(\d+)\s+sentences?\s+(?:or\s+less)?",
r"(\d+)\s+sentences?\s+(?:max|maximum)",
]
ANCHOR_PHRASE_PATTERNS = [
# Quoted phrases: remember "phrase" or remember the phrase "phrase"
r"remember\s+(?:the\s+phrase\s+)?['\"]([^'\"]+)['\"]",
# Unquoted phrase: remember the phrase X (where X doesn't start a new sentence)
r"remember\s+the\s+phrase\s+([a-z][a-z\s]+?)(?:\s+and\s+|\s+or\s+|\.|\s*$)",
# Generic remember without phrase keyword
r"remember\s+['\"]?([a-z][a-z\s]*?)['\"]?(?:\s+(?:and|or)|\.)",
# use/include/mention with optional quotes
r"remember\s+(?:to\s+)?(?:use|include|mention)\s+['\"]?([^'\"\.]+?)['\"]?(?:\s|\.)",
# anchor/key phrase with colon (matches multi-word phrases)
r"anchor\s*(?:phrase|word|term)?\s*:\s*([a-z][a-z\s]*?)(?:\s*\.|\s*$)",
r"(?:key\s+phrase):\s+([a-z][a-z\s]*?)(?:\s*\.|\s*$)",
# ── Informal phrasings ──────────────────────────────────────────
# "don't forget X" / "don't forget the phrase X"
r"don'?t\s+forget\s+(?:the\s+(?:phrase|word|term)\s+)?['\"]?([a-z][a-z\s]+?)['\"]?(?:[.,;]|\s+and\s+|\s*$)",
# "keep in mind X" / "keep in mind the phrase X"
r"keep\s+in\s+mind\s+(?:the\s+(?:phrase|word|term)\s+)?['\"]?([a-z][a-z\s]+?)['\"]?(?:[.,;]|\s+and\s+|\s*$)",
# "call it/this X" / "refer to it/this as X"
r"(?:call\s+(?:it|this)\s+|refer\s+to\s+(?:it|this)\s+as\s+)['\"]?([a-z][a-z\s]+?)['\"]?(?:[.,;]|\s+and\s+|\s*$)",
]
FORMAT_RULE_PATTERNS = [
r"(use\s+(?:bullet\s+)?points?)",
r"(format\s+as\s+(?:json|markdown|yaml))",
# Negated formatting rules — restricted to real formatting targets so
# ordinary negations ("no word constraint", "no constraints needed",
# "no problem", "no idea") are NOT captured as constraints.
r"((?:no|avoid|without|don'?t\s+use|do\s+not\s+use)\s+"
r"(?:bullet\s*points?|bullets?|numbered\s+lists?|lists?|markdown|json|"
r"yaml|xml|code\s*blocks?|headers?|headings?|emojis?|emoji|jargon|"
r"tables?|formatting|prose|paragraphs?))",
]
# Phrases that explicitly DECLINE constraints — when present, the query is
# asking for NO restrictions, so we must not derive constraints from it.
CONSTRAINT_NEGATION_PATTERNS = [
r"\bno\s+(?:word|sentence|length|format(?:ting)?|character)?\s*constraints?\b",
r"\bno\s+constraints?\s+(?:needed|required|please)\b",
r"\bno\s+(?:word|character|length)\s+limit\b",
r"\bwithout\s+(?:any\s+)?constraints?\b",
r"\bignore\s+(?:the\s+|any\s+|previous\s+)?constraints?\b",
r"\bno\s+restrictions?\b",
]
def detect(self, query: str, turn_num: int = 1) -> SessionConstraints:
"""Detect all constraints in a query.
Args:
query: User input text
turn_num: Turn number (used to track when constraints were set)
Returns:
SessionConstraints with detected constraints
"""
sc = SessionConstraints(detected_at_turn=turn_num)
# If the user explicitly declines constraints, derive none from this turn.
for neg in self.CONSTRAINT_NEGATION_PATTERNS:
if re.search(neg, query, re.IGNORECASE):
return sc
# Detect word limits
for pattern in self.WORD_LIMIT_PATTERNS:
match = re.search(pattern, query, re.IGNORECASE)
if match:
try:
limit = int(match.group(1))
sc.word_limit = limit
sc.constraints.append(DetectedConstraint(
kind="word_limit",
value=limit,
raw_text=match.group(0),
confidence=0.95
))
break
except (ValueError, IndexError):
pass
# Detect sentence limits
for pattern in self.SENTENCE_LIMIT_PATTERNS:
match = re.search(pattern, query, re.IGNORECASE)
if match:
try:
limit = int(match.group(1))
sc.sentence_limit = limit
sc.constraints.append(DetectedConstraint(
kind="sentence_limit",
value=limit,
raw_text=match.group(0),
confidence=0.95
))
break
except (ValueError, IndexError):
pass
# Detect anchor phrases
for pattern in self.ANCHOR_PHRASE_PATTERNS:
matches = re.finditer(pattern, query, re.IGNORECASE)
for match in matches:
try:
phrase = match.group(1).strip()
if phrase and len(phrase) > 2: # At least 3 chars
sc.anchor_phrases.append(phrase)
sc.constraints.append(DetectedConstraint(
kind="anchor_phrase",
value=phrase,
raw_text=match.group(0),
confidence=0.90
))
except IndexError:
pass
# Detect format rules
for pattern in self.FORMAT_RULE_PATTERNS:
matches = re.finditer(pattern, query, re.IGNORECASE)
for match in matches:
try:
rule = match.group(1).lower().strip()
if rule not in sc.format_rules:
sc.format_rules.append(rule)
sc.constraints.append(DetectedConstraint(
kind="format_rule",
value=rule,
raw_text=match.group(0),
confidence=0.85
))
except IndexError:
pass
return sc
class ConstraintEnforcer:
"""Enforce detected constraints on responses."""
@staticmethod
def word_count(text: str) -> int:
"""Count words in text (roughly)."""
return len([w for w in text.split() if w.strip()])
@staticmethod
def sentence_count(text: str) -> int:
"""Count sentences (roughly)."""
sentences = re.split(r'[.!?]+', text.strip())
return len([s for s in sentences if s.strip()])
@staticmethod
def has_anchor_phrases(text: str, phrases: List[str]) -> bool:
"""Check if all anchor phrases are present."""
text_lower = text.lower()
return all(phrase.lower() in text_lower for phrase in phrases)
@staticmethod
def build_constraint_reminder(constraints: SessionConstraints) -> str:
"""Build a constraint reminder string for the system prompt."""
if not constraints.constraints:
return ""
lines = ["[SESSION CONSTRAINTS]"]
if constraints.word_limit:
lines.append(f"- Keep your response to {constraints.word_limit} words or fewer")
if constraints.sentence_limit:
lines.append(f"- Keep your response to {constraints.sentence_limit} sentences or fewer")
if constraints.anchor_phrases:
phrases_str = ", ".join(f'"{p}"' for p in constraints.anchor_phrases)
lines.append(f"- IMPORTANT: Include these anchor phrases in your response: {phrases_str}")
if constraints.format_rules:
for rule in constraints.format_rules:
lines.append(f"- Format: {rule}")
lines.append("")
return "\n".join(lines)
class ConstraintTracker:
"""Main tracker for managing constraints across a session."""
def __init__(self):
self.detector = ConstraintDetector()
self.enforcer = ConstraintEnforcer()
self.session_constraints: Optional[SessionConstraints] = None
self.turn_count = 0
def process_turn(self, query: str, is_first_turn: bool = False) -> SessionConstraints:
"""Process a turn and detect/retrieve constraints.
Always scans the current query for new constraints. On the first turn the
session constraints are replaced; on subsequent turns newly-found anchors,
limits, and format rules are merged in without clobbering what was already set.
Args:
query: User input
is_first_turn: Whether this is the first turn (resets constraints)
Returns:
SessionConstraints for this turn
"""
self.turn_count += 1
if is_first_turn:
# First turn: full reset — detect fresh from this query
self.session_constraints = self.detector.detect(query, turn_num=1)
else:
# Fast-path: skip regex work entirely when the query has no constraint
# keywords. "What is the weather?" never contains anchors or limits —
# the keyword scan is O(n) and avoids 20+ regex compilations per turn.
_CONSTRAINT_SIGNALS = (
'remember', 'anchor', 'phrase', 'keyword', 'keep', 'limit',
'word', 'sentence', 'format', 'avoid', 'under', 'within', 'maximum',
'forget', 'note', 'call', 'refer',
)
q_lower = query.lower()
if not any(kw in q_lower for kw in _CONSTRAINT_SIGNALS):
return self.session_constraints or SessionConstraints()
# Mid-session: detect new constraints and merge (never clobber existing)
new_sc = self.detector.detect(query, turn_num=self.turn_count)
if new_sc.constraints:
if not self.session_constraints:
self.session_constraints = new_sc
else:
self._merge_into(new_sc)
return self.session_constraints or SessionConstraints()
def _merge_into(self, new_sc: SessionConstraints) -> None:
"""Merge new_sc into self.session_constraints without overwriting set values."""
sc = self.session_constraints
for c in new_sc.constraints:
if c.kind == "anchor_phrase" and c.value not in sc.anchor_phrases:
sc.anchor_phrases.append(c.value)
sc.constraints.append(c)
elif c.kind == "word_limit" and sc.word_limit is None:
sc.word_limit = c.value
sc.constraints.append(c)
elif c.kind == "sentence_limit" and sc.sentence_limit is None:
sc.sentence_limit = c.value
sc.constraints.append(c)
elif c.kind == "format_rule" and c.value not in sc.format_rules:
sc.format_rules.append(c.value)
sc.constraints.append(c)
def get_constraint_reminder(self) -> str:
"""Get the constraint reminder to inject into system prompt."""
if not self.session_constraints or not self.session_constraints.constraints:
return ""
return self.enforcer.build_constraint_reminder(self.session_constraints)
def check_constraint_compliance(self, response: str) -> Dict[str, Any]:
"""Check if response meets constraints.
Returns:
Dict with compliance status and violations.
"""
if not self.session_constraints or not self.session_constraints.constraints:
return {"compliant": True, "violations": []}
violations = []
if self.session_constraints.word_limit:
wc = self.enforcer.word_count(response)
if wc > self.session_constraints.word_limit:
violations.append({
"kind": "word_limit",
"expected": self.session_constraints.word_limit,
"actual": wc
})
if self.session_constraints.sentence_limit:
sc = self.enforcer.sentence_count(response)
if sc > self.session_constraints.sentence_limit:
violations.append({
"kind": "sentence_limit",
"expected": self.session_constraints.sentence_limit,
"actual": sc
})
if self.session_constraints.anchor_phrases:
if not self.enforcer.has_anchor_phrases(response, self.session_constraints.anchor_phrases):
violations.append({
"kind": "missing_anchor_phrases",
"expected": self.session_constraints.anchor_phrases
})
return {
"compliant": len(violations) == 0,
"violations": violations
}
def reset(self):
"""Reset tracker for new session."""
self.session_constraints = None
self.turn_count = 0
|