superlocalmemory 3.5.8 → 3.6.1

package/src/superlocalmemory/optimize/cache/boundary_store.py

@@ -0,0 +1,455 @@
+# optimize/cache/boundary_store.py
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under AGPL-3.0-or-later
+#
+# LLD-03 §3.1 + §4.2 — Per-item vCache online MLE boundary persistence.
+#
+# REAL vCache algorithm from arXiv:2502.03771 (Eq. 9/10/11, Algorithm 2,
+# Theorem 4.1). Replaces any prior step/raise/relax heuristic.
+#
+#   Eq. 9:  L(s, t, γ) = 1/(1 + exp(-γ(s - t)))            — sigmoid correctness
+#   Eq. 10: (t̂, γ̂) = argmin BCE on accumulated (s, c) pairs — MLE per entry
+#   Eq. 11: τ̂ = min_{ε} [(1-δ) - (1-ε)·L(s, t'(ε), γ̂)] / [1 - (1-ε)·L(...)]
+#           where t'(ε) = t̂ - z_{1-ε/2}·se(t̂)             — pessimistic CI
+#           se(t̂) = 1/sqrt(I_tt)  with I_tt = Σ γ̂²·p(1-p) — Fisher info
+#   Theorem 4.1: Pr(vCache(x) = r(x) | D) ≥ (1-δ) ∀ x, n.
+#
+# This is the GUARANTEE path (RA-01). The point-estimate Eq. 8 does NOT
+# carry Theorem 4.1; only Eq. 11 does.
+
+from __future__ import annotations
+
+import json
+import logging
+import math
+import random
+import time
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any
+
+import numpy as np
+
+if TYPE_CHECKING:
+    from superlocalmemory.optimize.storage.db import CacheDB
+
+logger = logging.getLogger(__name__)
+
+# Module-level RNG instance — tests can seed via _RNG.seed(42) for determinism.
+_RNG = random.Random()
+
+# Optional scipy MLE (N-D L-BFGS-B). Fall back to gradient descent if absent.
+try:
+    from scipy.optimize import minimize as _sp_minimize  # type: ignore[import]
+    _SCIPY_AVAILABLE = True
+except ImportError:  # pragma: no cover
+    _SCIPY_AVAILABLE = False
+    _sp_minimize = None  # type: ignore[assignment]
+
+_BOUNCE_EPS: float = 1e-9
+_OVERFLOW_GUARD: float = 500.0
+
+# z_{1 - ε/2} for common ε (avoids scipy.stats dependency)
+_Z_TABLE: dict[float, float] = {
+    0.01: 2.576,
+    0.02: 2.326,
+    0.05: 1.960,
+    0.10: 1.645,
+    0.15: 1.440,
+    0.20: 1.282,
+}
+
+
+# ---------------------------------------------------------------------------
+# PerItemBoundaryRecord — real vCache online MLE model for one cached entry
+# ---------------------------------------------------------------------------
+
+@dataclass
+class PerItemBoundaryRecord:
+    """vCache per-item online MLE model for one cached entry.
+
+    A-01 fix: real vCache algorithm — learned logistic regression per entry.
+    A-23 fix: sliding window (max_samples) prevents frozen learning.
+
+    Fields:
+        entry_id:    Surrogate entry ID from VCacheSemantic._derive_entry_id.
+        t_hat:       MLE estimate of decision boundary t ∈ [0, 1].
+                     Initialize conservative (≈ boundary_init).
+        gamma_hat:   MLE estimate of steepness γ > 0.
+                     Initialize steep (10.0) so cold-start is decisive.
+        samples:     Accumulated (similarity, correctness) training pairs.
+                     Sliding window — last `max_samples` entries retained.
+        last_updated: Unix timestamp of last model update.
+    """
+    entry_id: str
+    t_hat: float = 0.95
+    gamma_hat: float = 10.0
+    samples: list[tuple[float, int]] = field(default_factory=list)
+    last_updated: float = 0.0
+
+    # --- Algorithm methods (Eq. 9, 10, 11, Algorithm 2) -----------------
+
+    def compute_tau(
+        self,
+        query_sim: float,
+        delta: float = 0.05,
+        epsilon_grid: tuple[float, ...] = (0.01, 0.02, 0.05, 0.10),
+    ) -> float:
+        """Compute τ̂ — the vCache exploration probability (Eq. 11).
+
+        Args:
+            query_sim: Cosine similarity s(x) ∈ [0, 1] for the incoming query.
+            delta:     δ — user-defined maximum error rate.
+                       Theorem 4.1 guarantee: Pr(correct) ≥ 1 - δ.
+            epsilon_grid: ε values for the Eq. 11 min sweep.
+                          Distinct from δ; controls CI conservativeness.
+
+        Returns:
+            τ̂ ∈ [0.0, 1.0]. Lower = more exploitation.
+            Cold start (n < 3): returns 1.0 (always explore).
+
+        Eq. 11 derivation (from the paper):
+          1. I_tt = Σ γ̂² · p_i(1 - p_i)            [Fisher info diagonal]
+          2. se(t̂) = 1/sqrt(I_tt + ε)              [normal approx SE]
+          3. t'(ε) = t̂ - z_{1-ε/2} · se(t̂)         [pessimistic lower bound]
+          4. α(ε)   = (1 - ε) · L(s, t'(ε), γ̂)     [G_τ sub-function]
+          5. τ(ε)   = ((1 - δ) - α) / (1 - α)
+          6. τ̂      = min over ε in grid            [Eq. 11 min]
+        """
+        n = len(self.samples)
+        if n < 3:
+            return 1.0  # cold start — always explore
+
+        # Step 1: Fisher-information SE
+        i_tt = 0.0
+        for s, _c in self.samples:
+            p = _sigmoid(s, self.t_hat, self.gamma_hat)
+            i_tt += (self.gamma_hat ** 2) * p * (1.0 - p)
+        se_t = 1.0 / math.sqrt(i_tt + _BOUNCE_EPS)
+
+        # Step 2: Eq. 11 min over ε
+        best_tau = 1.0
+        for eps in epsilon_grid:
+            z = _Z_TABLE.get(eps, 1.960)  # default z_{0.975}
+            t_prime = self.t_hat - z * se_t
+            t_prime = max(0.0, min(1.0, t_prime))  # clip to valid t range
+
+            alpha = (1.0 - eps) * _sigmoid(query_sim, t_prime, self.gamma_hat)
+            denom = 1.0 - alpha
+            if denom < _BOUNCE_EPS:
+                tau_eps = 0.0  # near-certainty
+            else:
+                tau_eps = ((1.0 - delta) - alpha) / denom
+            tau_eps = max(0.0, min(1.0, tau_eps))
+            best_tau = min(best_tau, tau_eps)
+
+        return best_tau
+
+    def should_explore(self, query_sim: float, delta: float = 0.05) -> bool:
+        """Return True (explore = LLM call) or False (exploit = serve cache).
+
+        Source: vCache Algorithm 2: draw u ~ Uniform(0, 1); explore iff u ≤ τ̂.
+        """
+        tau = self.compute_tau(query_sim, delta=delta)
+        return _RNG.random() <= tau
+
+    def add_sample(
+        self,
+        similarity: float,
+        was_correct: bool,
+        max_samples: int = 200,
+    ) -> "PerItemBoundaryRecord":
+        """Add a (similarity, correctness) pair and refit MLE. Returns NEW record.
+
+        A-01 fix: this is the REAL vCache learning step. MLE refit on the
+        accumulated samples. No step size. No raise/relax heuristic.
+        A-23 fix: sliding window — drops oldest when len > max_samples.
+        """
+        new_samples = list(self.samples)
+        new_samples.append((float(similarity), 1 if was_correct else 0))
+        if len(new_samples) > max_samples:
+            new_samples = new_samples[-max_samples:]
+
+        new_t, new_gamma = _fit_logistic_mle(
+            new_samples, self.t_hat, self.gamma_hat,
+        )
+        return PerItemBoundaryRecord(
+            entry_id=self.entry_id,
+            t_hat=new_t,
+            gamma_hat=new_gamma,
+            samples=new_samples,
+            last_updated=time.time(),
+        )
+
+
+# ---------------------------------------------------------------------------
+# Numerical helpers
+# ---------------------------------------------------------------------------
+
+def _sigmoid(s: float, t: float, gamma: float) -> float:
+    """L(s, t, γ) = 1 / (1 + exp(-γ(s - t))) — Eq. 9. Overflow-guarded."""
+    exponent = -gamma * (s - t)
+    exponent = max(-_OVERFLOW_GUARD, min(_OVERFLOW_GUARD, exponent))
+    return 1.0 / (1.0 + math.exp(exponent))
+
+
+def _rng() -> float:
+    """u ~ Uniform(0, 1) for the vCache exploit/explore draw.
+
+    Tests can seed via `boundary_store._RNG.seed(N)` for determinism.
+    """
+    return _RNG.random()
+
+
+def _binary_cross_entropy(
+    params: tuple[float, float],
+    samples: list[tuple[float, int]],
+) -> float:
+    """BCE loss for the logistic model — Eq. 10 objective."""
+    t, gamma = params
+    if gamma <= 0 or not samples:
+        return 1e9
+    total = 0.0
+    for s, c in samples:
+        p = _sigmoid(s, t, gamma)
+        p = max(_BOUNCE_EPS, min(1.0 - _BOUNCE_EPS, p))  # numerical stability
+        total += -(c * math.log(p) + (1 - c) * math.log(1 - p))
+    return total / len(samples)
+
+
+def _fit_logistic_mle(
+    samples: list[tuple[float, int]],
+    t_init: float,
+    gamma_init: float,
+    t_prior: float = 0.95,
+) -> tuple[float, float]:
+    """Fit (t̂, γ̂) via MLE on `samples`. L-BFGS-B if scipy, else gradient descent.
+
+    Fail-open: returns (t_init, gamma_init) on any error.
+
+    The prior pulls the solution toward (t_prior, 10.0) so a single sample
+    does not collapse the boundary to the saturating corner. This is
+    required for stable online learning.
+
+    Warm-start trick: t_init is nudged toward the empirical midpoint of
+    the sample similarities. This avoids L-BFGS-B's "already-at-optimum"
+    early termination when the cold-start t=0.95 happens to sit at a
+    BCE plateau.
+    """
+    if not samples:
+        return t_init, gamma_init
+
+    # Warm-start: empirical midpoint of positive vs negative clusters.
+    pos = [s for s, c in samples if c == 1]
+    neg = [s for s, c in samples if c == 0]
+    if pos and neg:
+        empirical_mid = (min(pos) + max(neg)) / 2.0
+        t_warm = max(0.5, min(1.0, empirical_mid))
+    else:
+        t_warm = t_init
+    gamma_warm = gamma_init
+
+    if _SCIPY_AVAILABLE and _sp_minimize is not None:
+        try:
+            result = _sp_minimize(
+                fun=lambda p: _binary_cross_entropy(
+                    (p[0], p[1]), samples,
+                ),
+                x0=[t_warm, gamma_warm],
+                method="L-BFGS-B",
+                bounds=[(0.5, 1.0), (0.1, 100.0)],
+                options={"maxiter": 200, "ftol": 1e-10, "gtol": 1e-8},
+            )
+            t_out, g_out = float(result.x[0]), float(result.x[1])
+            # Guard: refuse to return a value outside the observed sample range
+            sims = [s for s, _ in samples]
+            lo, hi = min(sims), max(sims)
+            t_out = max(lo, min(hi, t_out))
+            return t_out, g_out
+        except Exception:
+            pass
+    return _fit_logistic_gd(samples, t_warm, gamma_warm)
+
+
+def _fit_logistic_gd(
+    samples: list[tuple[float, int]],
+    t_init: float,
+    gamma_init: float,
+    lr: float = 0.05,
+    steps: int = 300,
+) -> tuple[float, float]:
+    """Gradient-descent MLE fallback. Clips to t ∈ [0.5, 1.0], γ ∈ [0.1, 100]."""
+    t = t_init
+    gamma = gamma_init
+    for _ in range(steps):
+        dt = 0.0
+        dg = 0.0
+        for s, c in samples:
+            p = _sigmoid(s, t, gamma)
+            err = p - c
+            dt += err * gamma
+            dg += err * (-(s - t))
+        n = max(len(samples), 1)
+        t = max(0.5, min(1.0, t - lr * dt / n))
+        gamma = max(0.1, min(100.0, gamma - lr * dg / n))
+    return t, gamma
+
+
+# ---------------------------------------------------------------------------
+# BoundaryStore — SQLite-backed per-entry MLE record registry
+# ---------------------------------------------------------------------------
+
+class BoundaryStore:
+    """Manages per-item vCache learned boundaries.
+
+    Thread safety: CacheDB uses DatabaseManager (WAL, retry, busy_timeout).
+    Concurrent updates to the same entry_id are serialized by SQLite row
+    locking. The in-memory `_cache` is for warm-start; it is rebuilt via
+    load_all() at VCacheSemantic startup.
+
+    Fail-open: any DB error returns a safe cold-start default rather than
+    raising.
+    """
+
+    def __init__(
+        self,
+        db: "CacheDB",
+        default_t: float = 0.95,
+        default_gamma: float = 10.0,
+        floor: float = 0.85,
+        ceiling: float = 0.995,
+        step: float = 0.01,
+        epsilon: float = 0.02,
+    ) -> None:
+        self._db = db
+        self._default_t = default_t
+        self._default_gamma = default_gamma
+        self._floor = floor
+        self._ceiling = ceiling
+        self._step = step
+        self._epsilon = epsilon
+        # In-memory write-through cache. Populated by load_all() at warm.
+        # get() checks here first (O(1) hot path), then falls back to DB.
+        self._cache: dict[str, PerItemBoundaryRecord] = {}
+
+    def get(self, entry_id: str) -> PerItemBoundaryRecord:
+        """Return the MLE model record for an entry, or a cold-start default.
+
+        Fail-open: returns default (cold-start) record on DB error.
+        Never raises.
+        """
+        if entry_id in self._cache:
+            return self._cache[entry_id]
+        try:
+            row = self._db.boundary_get(entry_id)
+            if row is None:
+                return PerItemBoundaryRecord(
+                    entry_id=entry_id,
+                    t_hat=self._default_t,
+                    gamma_hat=self._default_gamma,
+                    samples=[],
+                    last_updated=time.time(),
+                )
+            try:
+                samples_raw = json.loads(getattr(row, "samples_json", None) or "[]") \
+                    if hasattr(row, "samples_json") else []
+            except (json.JSONDecodeError, TypeError):
+                samples_raw = []
+            if not samples_raw and hasattr(row, "samples_json"):
+                # BoundaryRow dataclass has no samples_json field; persist only
+                # the MLE parameters. samples are in-memory only after refit.
+                pass
+            # We persist only (t_hat, gamma_hat, sample_count) in the DB.
+            # samples is rebuilt from feedback calls (record_outcome) when needed.
+            return PerItemBoundaryRecord(
+                entry_id=entry_id,
+                t_hat=float(getattr(row, "logistic_t", self._default_t)),
+                gamma_hat=float(getattr(row, "logistic_gamma", self._default_gamma)),
+                samples=[],
+                last_updated=float(getattr(row, "updated_at", 0.0)),
+            )
+        except Exception as exc:
+            logger.warning("BoundaryStore.get failed (fail-open): %s", exc)
+            return PerItemBoundaryRecord(
+                entry_id=entry_id,
+                t_hat=self._default_t,
+                gamma_hat=self._default_gamma,
+                samples=[],
+                last_updated=time.time(),
+            )
+
+    def save(self, record: PerItemBoundaryRecord) -> None:
+        """Persist the MLE parameters (t_hat, gamma_hat, sample_count) for an entry.
+
+        Fail-open: logs warning on DB error, does not raise.
+        """
+        try:
+            from superlocalmemory.optimize.storage.db import BoundaryRow
+            row = BoundaryRow(
+                entry_id=record.entry_id,
+                logistic_t=record.t_hat,
+                logistic_gamma=record.gamma_hat,
+                sample_count=len(record.samples),
+                updated_at=record.last_updated or time.time(),
+            )
+            self._db.boundary_upsert(record.entry_id, row)
+            # In-memory write-through (RA-15)
+            self._cache[record.entry_id] = record
+        except Exception as exc:
+            logger.warning("BoundaryStore.save failed (fail-open): %s", exc)
+
+    def record_outcome(
+        self,
+        entry_id: str,
+        similarity: float,
+        was_correct: bool,
+        max_samples: int = 200,
+    ) -> PerItemBoundaryRecord:
+        """Fetch, add (s, c), refit MLE, save, return updated record.
+
+        A-01 fix: this IS the vCache online MLE learning step. No step size.
+        No raise/relax heuristic. The MLE handles everything.
+        """
+        record = self.get(entry_id)
+        updated = record.add_sample(
+            similarity=similarity,
+            was_correct=was_correct,
+            max_samples=max_samples,
+        )
+        self.save(updated)
+        return updated
+
+    def load_all(self) -> dict[str, PerItemBoundaryRecord]:
+        """Load all boundary records into memory (warm-start).
+
+        Returns:
+            Dict entry_id → PerItemBoundaryRecord. {} on error.
+
+        Note: only the (t_hat, gamma_hat, sample_count) are loaded from DB.
+        The full `samples` window is rebuilt on demand via record_outcome().
+        """
+        try:
+            rows = self._db.get_all_boundaries()
+            result: dict[str, PerItemBoundaryRecord] = {}
+            for r in rows:
+                eid = r.get("entry_id")
+                if not eid:
+                    continue
+                result[eid] = PerItemBoundaryRecord(
+                    entry_id=eid,
+                    t_hat=float(r.get("logistic_t", self._default_t)),
+                    gamma_hat=float(r.get("logistic_gamma", self._default_gamma)),
+                    samples=[],
+                    last_updated=float(r.get("updated_at", 0.0)),
+                )
+            return result
+        except Exception as exc:
+            logger.warning("BoundaryStore.load_all failed (fail-open): %s", exc)
+            return {}
+
+    def delete(self, entry_id: str) -> None:
+        """Remove boundary record for a deleted cache entry. Fail-open."""
+        try:
+            self._db.delete_boundary(entry_id)
+            self._cache.pop(entry_id, None)
+        except Exception as exc:
+            logger.warning("BoundaryStore.delete failed (fail-open): %s", exc)

package/src/superlocalmemory/optimize/cache/centroid_store.py

@@ -0,0 +1,158 @@
+# optimize/cache/centroid_store.py
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under AGPL-3.0-or-later
+#
+# LLD-03 §4.3 — SAFE-CACHE centroid registry for adversarial collision
+# detection.
+#
+# Source: SAFE-CACHE — Nature Scientific Reports 2026 (no arXiv).
+# Defense: compare the incoming query vector to the cluster centroid. If
+# the query is far from the centroid (cosine sim < 1 - distance_floor),
+# it falls outside the natural distribution → likely an adversarial
+# probe crafted to collide with a specific entry. Reject as miss.
+#
+# Design:
+#   - One centroid per tenant_id (simplest partition, proven sufficient).
+#   - Rebuilt from llmcache_semantic_vectors on startup.
+#   - Updated incrementally (Welford running mean) on every new set().
+#   - In-memory only — no separate SQL query needed (cache is O(N) anyway).
+
+from __future__ import annotations
+
+import logging
+import threading
+from typing import TYPE_CHECKING
+
+import numpy as np
+
+if TYPE_CHECKING:
+    from superlocalmemory.optimize.storage.db import CacheDB
+
+logger = logging.getLogger(__name__)
+
+_VARIANCE_FLOOR: float = 1e-6
+_EMBED_DIM: int = 768
+
+
+def _cosine_similarity(a: np.ndarray, b: np.ndarray) -> float:
+    """Cosine similarity in [-1, 1]. Returns 0.0 on zero vectors."""
+    norm_a = float(np.linalg.norm(a))
+    norm_b = float(np.linalg.norm(b))
+    if norm_a < _VARIANCE_FLOOR or norm_b < _VARIANCE_FLOOR:
+        return 0.0
+    return float(np.dot(a, b) / (norm_a * norm_b))
+
+
+class CentroidStore:
+    """In-memory centroid registry for SAFE-CACHE adversarial defense.
+
+    Thread-safe via a single RLock. All public methods are fail-open.
+
+    Centroid update rule (Welford running mean — exact, O(1) per update):
+        new_centroid = old_centroid * (n / (n+1)) + new_vec * (1 / (n+1))
+    """
+    def __init__(self) -> None:
+        self._centroids: dict[str, np.ndarray] = {}  # tenant_id → float32 vec
+        self._counts: dict[str, int] = {}            # tenant_id → count
+        self._lock = threading.RLock()
+
+    def rebuild_from_db(self, db: "CacheDB", tenant_id: str) -> None:
+        """Rebuild centroid for a tenant from all stored vectors.
+
+        Called at VCacheSemantic startup. O(N) — runs once per tenant.
+        Fail-open: on DB error or empty vector set, centroid is reset.
+        """
+        try:
+            rows = db.get_all_vectors(tenant_id=tenant_id)
+            if not rows:
+                with self._lock:
+                    self._centroids.pop(tenant_id, None)
+                    self._counts.pop(tenant_id, None)
+                return
+            vectors: list[np.ndarray] = []
+            for _entry_id, blob in rows:
+                try:
+                    vec = np.frombuffer(blob, dtype=np.float32).copy()
+                    if vec.shape[0] == _EMBED_DIM:
+                        vectors.append(vec)
+                except Exception:
+                    continue
+            if not vectors:
+                return
+            centroid = np.mean(np.stack(vectors, axis=0), axis=0).astype(np.float32)
+            with self._lock:
+                self._centroids[tenant_id] = centroid
+                self._counts[tenant_id] = len(vectors)
+            logger.debug(
+                "CentroidStore: rebuilt tenant=%s centroid from %d vectors",
+                tenant_id, len(vectors),
+            )
+        except Exception as exc:
+            logger.warning("CentroidStore.rebuild_from_db failed (fail-open): %s", exc)
+
+    def update(self, tenant_id: str, new_vector: np.ndarray) -> None:
+        """Update centroid with a new vector (Welford incremental mean).
+
+        Thread-safe. Fail-open.
+        """
+        try:
+            vec = new_vector.astype(np.float32)
+            with self._lock:
+                if tenant_id not in self._centroids:
+                    self._centroids[tenant_id] = vec.copy()
+                    self._counts[tenant_id] = 1
+                else:
+                    n = self._counts[tenant_id]
+                    old = self._centroids[tenant_id]
+                    self._centroids[tenant_id] = (
+                        old * (n / (n + 1)) + vec * (1.0 / (n + 1))
+                    ).astype(np.float32)
+                    self._counts[tenant_id] = n + 1
+        except Exception as exc:
+            logger.warning("CentroidStore.update failed (fail-open): %s", exc)
+
+    def is_adversarial(
+        self,
+        tenant_id: str,
+        query_vector: np.ndarray,
+        distance_floor: float = 0.15,
+    ) -> bool:
+        """Return True if query_vector appears adversarially crafted.
+
+        Defense: if the query is too far from the cluster centroid (cosine
+        similarity < 1 - distance_floor), it is an outlier — likely an
+        adversarial probe. Return True → reject (adversarial).
+
+        Fail-open: returns False (accept) on any error or missing centroid.
+        Skips defense if the cluster has < 5 entries (insufficient data).
+        """
+        try:
+            with self._lock:
+                centroid = self._centroids.get(tenant_id)
+                count = self._counts.get(tenant_id, 0)
+            if centroid is None or count < 5:
+                return False
+            q = query_vector.astype(np.float32)
+            sim = _cosine_similarity(q, centroid)
+            threshold = 1.0 - distance_floor
+            if sim < threshold:
+                logger.warning(
+                    "CentroidStore: adversarial probe detected "
+                    "(tenant=%s, centroid_sim=%.4f < %.4f) — returning miss.",
+                    tenant_id, sim, threshold,
+                )
+                return True
+            return False
+        except Exception as exc:
+            logger.warning("CentroidStore.is_adversarial failed (fail-open): %s", exc)
+            return False
+
+    def get_centroid(self, tenant_id: str) -> np.ndarray | None:
+        """Return current centroid for a tenant, or None if not established."""
+        with self._lock:
+            return self._centroids.get(tenant_id)
+
+    def count(self, tenant_id: str) -> int:
+        """Return the number of vectors contributing to this tenant's centroid."""
+        with self._lock:
+            return self._counts.get(tenant_id, 0)

package/src/superlocalmemory/optimize/cache/context_key.py

@@ -0,0 +1,67 @@
+# optimize/cache/context_key.py
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under AGPL-3.0-or-later
+#
+# LLD-03 §4.1 — Multi-turn context-aware key builder for the semantic cache.
+# Source: ContextCache / SmartCache (arXiv:2506.22791 §3) — context-aware
+# cache keys prevent false reuse across semantically overlapping but
+# conversationally distinct turns.
+#
+# The fingerprint is NOT the full cache key (that is KeyBuilder's job).
+# It is an auxiliary scope guard: a semantic hit is accepted ONLY IF
+# the stored entry's context fingerprint matches the query's context
+# fingerprint (or is absent — single-turn entries).
+#
+# A-22 fix: 16-hex-char fingerprint (64 bits → birthday at 2^32 entries,
+# safe for any realistic tenant). 8 hex chars (32 bits) birthday at 2^16.
+
+from __future__ import annotations
+
+import hashlib
+import json
+from typing import Any
+
+_CONTEXT_SCHEMA_VERSION: int = 1
+
+
+class ContextKeyBuilder:
+    """Builds a context fingerprint for multi-turn semantic cache lookup.
+
+    Args:
+        window_turns: Number of prior assistant+user turns to include.
+                      Default: 3 (matches LLD-03 §3.2 default).
+    """
+
+    def __init__(self, window_turns: int = 3) -> None:
+        if window_turns < 1:
+            raise ValueError(f"window_turns must be >= 1, got {window_turns}")
+        self._window = window_turns
+
+    def build(self, messages: list[dict[str, Any]], tenant_id: str) -> str:
+        """Build a 16-hex-char (64-bit) context fingerprint.
+
+        Takes the last `window_turns * 2` messages (user + assistant pairs),
+        canonicalizes them, SHA-256 hashes them, returns first 16 hex chars.
+
+        Returns:
+            16-hex context fingerprint, e.g. "a3f2b1c0d4e5f601". Never empty.
+            Single-turn (no prior context) returns the SHA-256 of an empty
+            canonical context array — a stable sentinel.
+        """
+        prior = messages[:-1] if len(messages) > 1 else []
+        window = prior[-(self._window * 2):]
+
+        payload = {
+            "v": _CONTEXT_SCHEMA_VERSION,
+            "tenant": tenant_id,
+            "ctx": window,
+        }
+        canonical = json.dumps(
+            payload, sort_keys=True, separators=(",", ":"),
+            ensure_ascii=True, default=str,
+        )
+        return hashlib.sha256(canonical.encode("utf-8")).hexdigest()[:16]
+
+    def turn_count(self, messages: list[dict[str, Any]]) -> int:
+        """Return the number of completed conversation turns (assistant messages)."""
+        return sum(1 for m in messages if m.get("role") == "assistant")