superlocalmemory 2.8.5 → 3.0.0

package/src/superlocalmemory/encoding/scene_builder.py

@@ -0,0 +1,183 @@
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under the MIT License - see LICENSE file
+# Part of SuperLocalMemory V3 | https://qualixar.com | https://varunpratap.com
+
+"""SuperLocalMemory V3 — Scene Builder (Memory Clustering).
+
+Groups related facts into thematic scenes (EverMemOS MemScene pattern).
+Scenes provide contextual retrieval — related facts come together.
+
+V1 had this module but NEVER CALLED it. Now wired into the encoding pipeline.
+
+Part of Qualixar | Author: Varun Pratap Bhardwaj
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from datetime import UTC, datetime
+
+from superlocalmemory.storage.models import AtomicFact, MemoryScene
+
+logger = logging.getLogger(__name__)
+
+# Similarity threshold for assigning fact to existing scene
+_ASSIGN_THRESHOLD = 0.6
+
+
+class SceneBuilder:
+    """Cluster related facts into thematic scenes.
+
+    When a new fact arrives:
+    1. Compute similarity to existing scenes (via scene theme embedding)
+    2. If above threshold: assign to nearest scene, update scene
+    3. If below threshold: create new scene
+    """
+
+    def __init__(self, db, embedder=None) -> None:
+        self._db = db
+        self._embedder = embedder
+        self._scene_embeddings_cache: dict[str, list[float]] = {}
+
+    def assign_to_scene(
+        self,
+        new_fact: AtomicFact,
+        profile_id: str,
+    ) -> MemoryScene:
+        """Assign a fact to an existing scene or create a new one.
+
+        Always embeds the incoming fact content (when embedder is available)
+        so that the embedding is ready for comparison against existing scenes.
+        """
+        if self._embedder is None:
+            return self._create_scene(new_fact, profile_id)
+
+        # Always compute fact embedding first — needed for comparisons
+        fact_emb = self._embedder.embed(new_fact.content)
+
+        scenes = self._get_scenes(profile_id)
+        if not scenes:
+            return self._create_scene(new_fact, profile_id)
+
+        # Find best matching scene
+        best_scene: MemoryScene | None = None
+        best_sim = -1.0
+
+        for scene in scenes:
+            # Use cached embedding if available, otherwise compute fresh
+            if scene.theme in self._scene_embeddings_cache:
+                theme_emb = self._scene_embeddings_cache[scene.theme]
+            else:
+                theme_emb = self._embedder.embed(scene.theme)
+                self._scene_embeddings_cache[scene.theme] = theme_emb
+            sim = _cosine(fact_emb, theme_emb)
+            if sim > best_sim:
+                best_sim = sim
+                best_scene = scene
+
+        if best_scene is not None and best_sim >= _ASSIGN_THRESHOLD:
+            return self._add_to_scene(best_scene, new_fact, profile_id)
+
+        return self._create_scene(new_fact, profile_id)
+
+    def get_scene_for_fact(self, fact_id: str, profile_id: str) -> MemoryScene | None:
+        """Get the scene containing a specific fact."""
+        rows = self._db.execute(
+            "SELECT * FROM memory_scenes WHERE profile_id = ?", (profile_id,)
+        )
+        for row in rows:
+            d = dict(row)
+            fids = json.loads(d.get("fact_ids_json", "[]"))
+            if fact_id in fids:
+                return self._row_to_scene(d)
+        return None
+
+    def get_all_scenes(self, profile_id: str) -> list[MemoryScene]:
+        """Get all scenes for a profile."""
+        return self._get_scenes(profile_id)
+
+    # -- Internal ----------------------------------------------------------
+
+    def _create_scene(self, fact: AtomicFact, profile_id: str) -> MemoryScene:
+        """Create a new scene from a single fact.
+
+        Pre-computes and caches the theme embedding for efficient later
+        comparisons in assign_to_scene.
+        """
+        theme = fact.content[:200]
+        # Pre-compute theme embedding for future comparisons
+        if self._embedder is not None:
+            self._scene_embeddings_cache[theme] = self._embedder.embed(theme)
+
+        scene = MemoryScene(
+            profile_id=profile_id,
+            theme=theme,
+            fact_ids=[fact.fact_id],
+            entity_ids=list(fact.canonical_entities),
+            created_at=datetime.now(UTC).isoformat(),
+            last_updated=datetime.now(UTC).isoformat(),
+        )
+        self._save_scene(scene)
+        return scene
+
+    def _add_to_scene(
+        self, scene: MemoryScene, fact: AtomicFact, profile_id: str
+    ) -> MemoryScene:
+        """Add a fact to an existing scene."""
+        new_fact_ids = [*scene.fact_ids, fact.fact_id]
+        new_entity_ids = list(set(scene.entity_ids) | set(fact.canonical_entities))
+        updated = MemoryScene(
+            scene_id=scene.scene_id,
+            profile_id=profile_id,
+            theme=scene.theme,
+            fact_ids=new_fact_ids,
+            entity_ids=new_entity_ids,
+            created_at=scene.created_at,
+            last_updated=datetime.now(UTC).isoformat(),
+        )
+        self._save_scene(updated)
+        return updated
+
+    def _get_scenes(self, profile_id: str) -> list[MemoryScene]:
+        """Load all scenes from DB."""
+        rows = self._db.execute(
+            "SELECT * FROM memory_scenes WHERE profile_id = ? ORDER BY last_updated DESC",
+            (profile_id,),
+        )
+        return [self._row_to_scene(dict(r)) for r in rows]
+
+    def _save_scene(self, scene: MemoryScene) -> None:
+        """Upsert scene to DB."""
+        self._db.execute(
+            """INSERT OR REPLACE INTO memory_scenes
+               (scene_id, profile_id, theme, fact_ids_json, entity_ids_json,
+                created_at, last_updated)
+               VALUES (?, ?, ?, ?, ?, ?, ?)""",
+            (
+                scene.scene_id, scene.profile_id, scene.theme,
+                json.dumps(scene.fact_ids), json.dumps(scene.entity_ids),
+                scene.created_at, scene.last_updated,
+            ),
+        )
+
+    @staticmethod
+    def _row_to_scene(d: dict) -> MemoryScene:
+        return MemoryScene(
+            scene_id=d["scene_id"],
+            profile_id=d["profile_id"],
+            theme=d.get("theme", ""),
+            fact_ids=json.loads(d.get("fact_ids_json", "[]")),
+            entity_ids=json.loads(d.get("entity_ids_json", "[]")),
+            created_at=d.get("created_at", ""),
+            last_updated=d.get("last_updated", ""),
+        )
+
+
+def _cosine(a: list[float], b: list[float]) -> float:
+    dot = sum(x * y for x, y in zip(a, b))
+    na = sum(x * x for x in a) ** 0.5
+    nb = sum(x * x for x in b) ** 0.5
+    if na == 0 or nb == 0:
+        return 0.0
+    return dot / (na * nb)

package/src/superlocalmemory/encoding/signal_inference.py

@@ -0,0 +1,90 @@
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under the MIT License - see LICENSE file
+# Part of SuperLocalMemory V3 | https://qualixar.com | https://varunpratap.com
+
+"""SuperLocalMemory V3 — Signal Inference (6 Types).
+
+Infers the signal type of a memory/fact from its content.
+Ported from V2.8 — used to adjust retrieval channel weights.
+
+Part of Qualixar | Author: Varun Pratap Bhardwaj
+"""
+
+from __future__ import annotations
+
+import re
+
+from superlocalmemory.storage.models import SignalType
+
+# Compiled patterns for each signal type
+_PATTERNS: dict[SignalType, re.Pattern] = {
+    SignalType.EMOTIONAL: re.compile(
+        r"\b(happy|sad|angry|frustrated|excited|worried|anxious|"
+        r"love|hate|afraid|grateful|disappointed|thrilled|upset|"
+        r"feeling|emotion|mood)\b",
+        re.IGNORECASE,
+    ),
+    SignalType.TEMPORAL: re.compile(
+        r"\b(when|date|time|schedule|deadline|tomorrow|yesterday|"
+        r"last week|next month|ago|soon|later|earlier|"
+        r"january|february|march|april|may|june|july|"
+        r"august|september|october|november|december|"
+        r"monday|tuesday|wednesday|thursday|friday|saturday|sunday)\b",
+        re.IGNORECASE,
+    ),
+    SignalType.OPINION: re.compile(
+        r"\b(think|believe|prefer|opinion|recommend|suggest|"
+        r"should|better|worse|best|worst|favorite|"
+        r"in my view|personally|i feel|i guess)\b",
+        re.IGNORECASE,
+    ),
+    SignalType.REQUEST: re.compile(
+        r"\b(please|could you|would you|can you|help|need|"
+        r"want|looking for|searching|find|tell me|show me|"
+        r"remind me|remember)\b",
+        re.IGNORECASE,
+    ),
+    SignalType.SOCIAL: re.compile(
+        r"\b(friend|family|colleague|partner|boss|team|"
+        r"relationship|together|married|dating|met with|"
+        r"call|message|email|chat)\b",
+        re.IGNORECASE,
+    ),
+}
+
+# Priority order: more specific types checked first.
+# SOCIAL before TEMPORAL because social signals often co-occur with
+# temporal markers ("met yesterday") and the social context is primary.
+_PRIORITY = [
+    SignalType.REQUEST,
+    SignalType.SOCIAL,
+    SignalType.TEMPORAL,
+    SignalType.EMOTIONAL,
+    SignalType.OPINION,
+]
+
+
+def infer_signal(text: str) -> SignalType:
+    """Infer the signal type of a text.
+
+    Returns the most specific matching type, or FACTUAL as default.
+    """
+    for stype in _PRIORITY:
+        pattern = _PATTERNS.get(stype)
+        if pattern and pattern.search(text):
+            return stype
+    return SignalType.FACTUAL
+
+
+def infer_signal_scores(text: str) -> dict[SignalType, float]:
+    """Compute a signal score for each type (0.0-1.0).
+
+    Useful for multi-signal content that mixes types.
+    Score = number of pattern matches / 5 (capped at 1.0).
+    """
+    scores: dict[SignalType, float] = {}
+    for stype, pattern in _PATTERNS.items():
+        matches = pattern.findall(text)
+        scores[stype] = min(1.0, len(matches) / 5.0)
+    scores[SignalType.FACTUAL] = 1.0 - max(scores.values()) if scores else 1.0
+    return scores

package/src/superlocalmemory/encoding/temporal_parser.py

@@ -0,0 +1,426 @@
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under the MIT License - see LICENSE file
+# Part of SuperLocalMemory V3 | https://qualixar.com | https://varunpratap.com
+
+"""SuperLocalMemory V3 — Temporal Parser (3-Date Model).
+
+Parses and enriches temporal information for every memory:
+  1. observation_date  — when the conversation happened
+  2. referenced_date   — date mentioned in content ("last Tuesday")
+  3. temporal_interval  — [start, end] for duration events
+
+Mastra achieved 95.5% temporal reasoning with this pattern.
+V1 stored session_date as raw strings and ignored content dates entirely.
+
+Part of Qualixar | Author: Varun Pratap Bhardwaj
+"""
+
+from __future__ import annotations
+
+import logging
+import re
+from datetime import UTC, datetime, timedelta
+from typing import Any
+
+from dateutil.parser import parse as dateutil_parse, ParserError
+from dateutil.relativedelta import relativedelta
+
+from superlocalmemory.storage.models import AtomicFact, TemporalEvent
+
+logger = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Compiled regex patterns (compiled once, reused)
+# ---------------------------------------------------------------------------
+
+_ISO_DATE = re.compile(r"\b\d{4}-\d{2}-\d{2}\b")
+
+_US_DATE = re.compile(r"\b\d{1,2}/\d{1,2}/\d{2,4}\b")
+
+_WRITTEN_DATE = re.compile(
+    r"\b(?:Jan(?:uary)?|Feb(?:ruary)?|Mar(?:ch)?|Apr(?:il)?|May|Jun(?:e)?|"
+    r"Jul(?:y)?|Aug(?:ust)?|Sep(?:tember)?|Oct(?:ober)?|Nov(?:ember)?|Dec(?:ember)?)"
+    r"\s+\d{1,2},?\s*\d{4}\b",
+    re.IGNORECASE,
+)
+
+_WRITTEN_DATE_DMY = re.compile(
+    r"\b\d{1,2}\s+(?:Jan(?:uary)?|Feb(?:ruary)?|Mar(?:ch)?|Apr(?:il)?|May|Jun(?:e)?|"
+    r"Jul(?:y)?|Aug(?:ust)?|Sep(?:tember)?|Oct(?:ober)?|Nov(?:ember)?|Dec(?:ember)?)"
+    r",?\s*\d{4}\b",
+    re.IGNORECASE,
+)
+
+# "January 2023", "March 2024" — month + year without day
+_MONTH_YEAR = re.compile(
+    r"\b(?:Jan(?:uary)?|Feb(?:ruary)?|Mar(?:ch)?|Apr(?:il)?|May|Jun(?:e)?|"
+    r"Jul(?:y)?|Aug(?:ust)?|Sep(?:tember)?|Oct(?:ober)?|Nov(?:ember)?|Dec(?:ember)?)"
+    r"\s+\d{4}\b",
+    re.IGNORECASE,
+)
+
+_WEEKDAYS = r"(?:Monday|Tuesday|Wednesday|Thursday|Friday|Saturday|Sunday)"
+_TIME_UNITS = r"(?:days?|weeks?|months?|years?)"
+
+_RELATIVE = re.compile(
+    rf"\b(?:last|next|this)\s+(?:{_WEEKDAYS}|week|month|year|spring|summer|autumn|fall|winter)\b",
+    re.IGNORECASE,
+)
+
+_AGO = re.compile(
+    rf"\b(\d+)\s+{_TIME_UNITS}\s+ago\b",
+    re.IGNORECASE,
+)
+
+_IN_FUTURE = re.compile(
+    rf"\bin\s+(\d+)\s+{_TIME_UNITS}\b",
+    re.IGNORECASE,
+)
+
+_DURATION = re.compile(
+    r"\bfrom\s+(.+?)\s+to\s+(.+?)(?:\.|,|;|$)",
+    re.IGNORECASE,
+)
+
+_FOR_DURATION = re.compile(
+    rf"\bfor\s+(\d+)\s+{_TIME_UNITS}\b",
+    re.IGNORECASE,
+)
+
+_VAGUE_TERMS: dict[str, int] = {
+    "yesterday": -1,
+    "today": 0,
+    "tomorrow": 1,
+    "recently": -7,
+    "a while ago": -30,
+    "soon": 7,
+    "the other day": -3,
+    "last night": -1,
+}
+
+_SEASON_MONTHS: dict[str, tuple[int, int]] = {
+    "spring": (3, 5),
+    "summer": (6, 8),
+    "autumn": (9, 11),
+    "fall": (9, 11),
+    "winter": (12, 2),
+}
+
+_WEEKDAY_MAP: dict[str, int] = {
+    "monday": 0, "tuesday": 1, "wednesday": 2, "thursday": 3,
+    "friday": 4, "saturday": 5, "sunday": 6,
+}
+
+
+def _safe_iso(dt: datetime | None) -> str | None:
+    """Convert datetime to ISO-8601 string, or None."""
+    if dt is None:
+        return None
+    return dt.strftime("%Y-%m-%dT%H:%M:%S")
+
+
+def _extract_unit(text: str) -> str:
+    """Extract time unit from a matched span (days, weeks, months, years)."""
+    lower = text.lower()
+    for unit in ("year", "month", "week", "day"):
+        if unit in lower:
+            return unit
+    return "day"
+
+
+class TemporalParser:
+    """3-date temporal parser for memory enrichment.
+
+    Extracts observation_date, referenced_date, and temporal intervals
+    from session dates and fact content. Handles absolute, relative,
+    duration, and vague temporal expressions.
+    """
+
+    def __init__(self, reference_date: str | None = None) -> None:
+        """Initialize with an optional reference date for relative computation.
+
+        Args:
+            reference_date: ISO-8601 string used as "today" for relative dates.
+                           Defaults to current UTC time if None.
+        """
+        if reference_date is not None:
+            try:
+                parsed = dateutil_parse(reference_date)
+                self._ref = parsed.replace(tzinfo=UTC) if parsed.tzinfo is None else parsed
+            except (ParserError, ValueError):
+                logger.warning("Unparseable reference_date %r, using UTC now", reference_date)
+                self._ref = datetime.now(UTC)
+        else:
+            self._ref = datetime.now(UTC)
+
+    # ------------------------------------------------------------------
+    # Public API
+    # ------------------------------------------------------------------
+
+    def parse_session_date(self, raw_date: str) -> str | None:
+        """Parse LoCoMo-style session dates into ISO-8601.
+
+        Handles formats like:
+          - "1:56 pm on 8 May, 2023"
+          - "May 8, 2023"
+          - "2023-05-08"
+          - "8:56 pm on 20 July, 2023"
+
+        Returns:
+            ISO-8601 string or None if unparseable.
+        """
+        if not raw_date or not raw_date.strip():
+            return None
+        try:
+            dt = dateutil_parse(raw_date, fuzzy=True)
+            return _safe_iso(dt)
+        except (ParserError, ValueError, OverflowError):
+            logger.debug("Could not parse session_date: %r", raw_date)
+            return None
+
+    def extract_dates_from_text(self, text: str) -> dict[str, str | None]:
+        """Extract dates mentioned in fact content.
+
+        Strategy:
+          1. Regex pass for structured date patterns
+          2. Relative date resolution against reference_date
+          3. Vague term mapping to approximate offsets
+          4. Duration extraction for intervals
+
+        Returns:
+            dict with keys: referenced_date, interval_start, interval_end
+            All values are ISO-8601 strings or None.
+        """
+        if not text:
+            return {"referenced_date": None, "interval_start": None, "interval_end": None}
+
+        dates_found: list[datetime] = []
+        interval_start: datetime | None = None
+        interval_end: datetime | None = None
+
+        # --- Pass 1: Duration expressions ("from X to Y") ---
+        dur_match = _DURATION.search(text)
+        if dur_match:
+            start_dt = self._try_parse(dur_match.group(1).strip())
+            end_dt = self._try_parse(dur_match.group(2).strip())
+            if start_dt is not None and end_dt is not None:
+                interval_start = start_dt
+                interval_end = end_dt
+
+        # --- Pass 2: "for N units" duration ---
+        if interval_start is None:
+            for_match = _FOR_DURATION.search(text)
+            if for_match:
+                count = int(for_match.group(1))
+                unit = _extract_unit(for_match.group(0))
+                delta = self._unit_delta(count, unit)
+                interval_start = self._ref
+                interval_end = self._ref + delta
+
+        # --- Pass 3: Absolute date patterns ---
+        for pattern in (_ISO_DATE, _WRITTEN_DATE, _WRITTEN_DATE_DMY, _MONTH_YEAR, _US_DATE):
+            for match in pattern.finditer(text):
+                dt = self._try_parse(match.group(0))
+                if dt is not None:
+                    dates_found.append(dt)
+
+        # --- Pass 4: Relative expressions ---
+        for match in _RELATIVE.finditer(text):
+            dt = self._resolve_relative(match.group(0))
+            if dt is not None:
+                dates_found.append(dt)
+
+        # --- Pass 5: "N units ago" ---
+        for match in _AGO.finditer(text):
+            count = int(match.group(1))
+            unit = _extract_unit(match.group(0))
+            dt = self._ref - self._unit_delta(count, unit)
+            dates_found.append(dt)
+
+        # --- Pass 6: "in N units" ---
+        for match in _IN_FUTURE.finditer(text):
+            count = int(match.group(1))
+            unit = _extract_unit(match.group(0))
+            dt = self._ref + self._unit_delta(count, unit)
+            dates_found.append(dt)
+
+        # --- Pass 7: Vague terms ---
+        text_lower = text.lower()
+        for term, offset_days in _VAGUE_TERMS.items():
+            if term in text_lower:
+                dates_found.append(self._ref + timedelta(days=offset_days))
+                break  # Take first vague match only
+
+        # --- Assemble result ---
+        referenced: datetime | None = None
+        if dates_found:
+            referenced = dates_found[0]
+            # Two distinct dates without explicit "from...to" -> interval
+            if len(dates_found) >= 2 and interval_start is None:
+                # Normalize tz-awareness before comparing
+                _normed = [d.replace(tzinfo=UTC) if d.tzinfo is None else d for d in dates_found[:2]]
+                sorted_dates = sorted(_normed)
+                interval_start = sorted_dates[0]
+                interval_end = sorted_dates[1]
+
+        return {
+            "referenced_date": _safe_iso(referenced),
+            "interval_start": _safe_iso(interval_start),
+            "interval_end": _safe_iso(interval_end),
+        }
+
+    def build_temporal_event(
+        self,
+        fact: AtomicFact,
+        session_date: str | None,
+        entity_id: str,
+    ) -> TemporalEvent | None:
+        """Create a TemporalEvent for a fact-entity pair.
+
+        Combines the parsed session_date (observation) with dates
+        extracted from fact content (referenced + interval).
+
+        Returns:
+            TemporalEvent if any temporal info exists, else None.
+        """
+        obs_date = self.parse_session_date(session_date) if session_date else None
+        content_dates = self.extract_dates_from_text(fact.content)
+
+        ref_date = content_dates["referenced_date"]
+        int_start = content_dates["interval_start"]
+        int_end = content_dates["interval_end"]
+
+        # Use fact-level fields if already populated (from upstream)
+        if fact.observation_date and not obs_date:
+            obs_date = fact.observation_date
+        if fact.referenced_date and not ref_date:
+            ref_date = fact.referenced_date
+        if fact.interval_start and not int_start:
+            int_start = fact.interval_start
+        if fact.interval_end and not int_end:
+            int_end = fact.interval_end
+
+        # Only create event if we have at least one temporal anchor
+        if not any([obs_date, ref_date, int_start, int_end]):
+            return None
+
+        return TemporalEvent(
+            profile_id=fact.profile_id,
+            entity_id=entity_id,
+            fact_id=fact.fact_id,
+            observation_date=obs_date,
+            referenced_date=ref_date,
+            interval_start=int_start,
+            interval_end=int_end,
+            description=fact.content[:200],
+        )
+
+    def build_entity_timeline(
+        self,
+        entity_id: str,
+        facts: list[AtomicFact],
+        session_date: str | None = None,
+    ) -> list[TemporalEvent]:
+        """Build a chronological timeline for an entity from its facts.
+
+        Returns:
+            List of TemporalEvents sorted by earliest available date.
+        """
+        events: list[TemporalEvent] = []
+        for fact in facts:
+            s_date = session_date or fact.observation_date
+            event = self.build_temporal_event(fact, s_date, entity_id)
+            if event is not None:
+                events.append(event)
+
+        # Sort by the earliest available date
+        def _sort_key(ev: TemporalEvent) -> str:
+            for field in (ev.referenced_date, ev.observation_date, ev.interval_start):
+                if field:
+                    return field
+            return "9999-12-31T23:59:59"
+
+        return sorted(events, key=_sort_key)
+
+    # ------------------------------------------------------------------
+    # Internal helpers
+    # ------------------------------------------------------------------
+
+    def _try_parse(self, text: str) -> datetime | None:
+        """Attempt to parse a text fragment as a date."""
+        if not text or len(text.strip()) < 3:
+            return None
+        try:
+            return dateutil_parse(text, fuzzy=True)
+        except (ParserError, ValueError, OverflowError):
+            return None
+
+    def _resolve_relative(self, expr: str) -> datetime | None:
+        """Resolve a relative expression like 'last Tuesday' or 'next month'."""
+        lower = expr.lower().strip()
+        parts = lower.split()
+        if len(parts) < 2:
+            return None
+
+        modifier = parts[0]  # last / next / this
+        target = parts[1]
+
+        # Weekday resolution
+        if target in _WEEKDAY_MAP:
+            target_day = _WEEKDAY_MAP[target]
+            current_day = self._ref.weekday()
+            if modifier == "last":
+                diff = (current_day - target_day) % 7
+                diff = diff if diff > 0 else 7
+                return self._ref - timedelta(days=diff)
+            elif modifier == "next":
+                diff = (target_day - current_day) % 7
+                diff = diff if diff > 0 else 7
+                return self._ref + timedelta(days=diff)
+            else:  # this
+                diff = (target_day - current_day) % 7
+                return self._ref + timedelta(days=diff)
+
+        # Week / month / year
+        if target == "week":
+            offset = {"last": -7, "next": 7, "this": 0}
+            return self._ref + timedelta(days=offset.get(modifier, 0))
+        if target == "month":
+            offset = {"last": -1, "next": 1, "this": 0}
+            return self._ref + relativedelta(months=offset.get(modifier, 0))
+        if target == "year":
+            offset = {"last": -1, "next": 1, "this": 0}
+            return self._ref + relativedelta(years=offset.get(modifier, 0))
+
+        # Seasons
+        if target in _SEASON_MONTHS:
+            start_m, end_m = _SEASON_MONTHS[target]
+            year = self._ref.year
+            if modifier == "last":
+                year -= 1
+            elif modifier == "next":
+                year += 1
+            # Return midpoint of season
+            if start_m <= end_m:
+                mid_m = (start_m + end_m) // 2
+            else:
+                mid_m = 1  # winter spans Dec-Feb, midpoint ~ Jan
+            try:
+                return datetime(year, mid_m, 15)
+            except ValueError:
+                return None
+
+        return None
+
+    @staticmethod
+    def _unit_delta(count: int, unit: str) -> timedelta | relativedelta:
+        """Build a timedelta/relativedelta from count + unit string."""
+        if unit == "day":
+            return timedelta(days=count)
+        if unit == "week":
+            return timedelta(weeks=count)
+        if unit == "month":
+            return relativedelta(months=count)
+        if unit == "year":
+            return relativedelta(years=count)
+        return timedelta(days=count)