@provos/ironcurtain 0.11.0 → 0.12.0

package/src/workflow/workflows/evolve/scripts/evolve_core/vector_index.py

@@ -0,0 +1,191 @@
+"""Vector index wrapper with FAISS and pickle-backed fallbacks."""
+
+from __future__ import annotations
+
+import pickle
+from pathlib import Path
+from threading import RLock
+from typing import Dict, List, Optional, Tuple
+
+import numpy as np
+
+try:
+    import faiss
+
+    FAISS_AVAILABLE = True
+except Exception:
+    faiss = None
+    FAISS_AVAILABLE = False
+
+
+class FAISSIndex:
+    """Manage semantic retrieval with FAISS when available, else brute force."""
+
+    def __init__(
+        self,
+        dimension: int = 384,
+        index_type: str = "IP",
+        storage_path: Optional[Path] = None,
+    ):
+        self.dimension = dimension
+        self.index_type = index_type
+        self.storage_path = Path(storage_path) if storage_path else None
+        self.lock = RLock()
+        self.use_faiss = FAISS_AVAILABLE
+        self.id_to_idx: Dict[int, int] = {}
+        self.idx_to_id: Dict[int, int] = {}
+        self.next_idx = 0
+        self.vectors: Dict[int, np.ndarray] = {}
+
+        if self.use_faiss:
+            self.index = self._create_faiss_index()
+        else:
+            self.index = None
+
+        if self.storage_path:
+            self._load()
+
+    def add(self, node_id: int, vector: np.ndarray) -> None:
+        with self.lock:
+            if node_id in self.id_to_idx or node_id in self.vectors:
+                return
+
+            normalized = self._normalize(vector.reshape(1, -1).astype(np.float32))[0]
+            if self.use_faiss and self.index is not None:
+                self.index.add(normalized.reshape(1, -1))
+                self.id_to_idx[node_id] = self.next_idx
+                self.idx_to_id[self.next_idx] = node_id
+                self.next_idx += 1
+            else:
+                self.vectors[node_id] = normalized
+
+    def search(
+        self,
+        query_vector: np.ndarray,
+        top_k: int = 5,
+        score_threshold: float = 0.0,
+    ) -> List[Tuple[int, float]]:
+        with self.lock:
+            if self.use_faiss and self.index is not None:
+                if self.index.ntotal == 0:
+                    return []
+                query = self._normalize(query_vector.reshape(1, -1).astype(np.float32))
+                k = min(top_k, self.index.ntotal)
+                scores, indices = self.index.search(query, k)
+                results: List[Tuple[int, float]] = []
+                for score, idx in zip(scores[0], indices[0]):
+                    if idx < 0 or score < score_threshold:
+                        continue
+                    node_id = self.idx_to_id.get(int(idx))
+                    if node_id is not None:
+                        results.append((node_id, float(score)))
+                return results
+
+            if not self.vectors:
+                return []
+
+            query = self._normalize(query_vector.reshape(1, -1).astype(np.float32))[0]
+            scored: List[Tuple[int, float]] = []
+            for node_id, vector in self.vectors.items():
+                if self.index_type == "IP":
+                    score = float(np.dot(query, vector))
+                else:
+                    score = -float(np.linalg.norm(query - vector))
+                if score >= score_threshold:
+                    scored.append((node_id, score))
+            scored.sort(key=lambda item: item[1], reverse=True)
+            return scored[:top_k]
+
+    def remove(self, node_id: int) -> None:
+        with self.lock:
+            if self.use_faiss and node_id in self.id_to_idx:
+                idx = self.id_to_idx.pop(node_id)
+                self.idx_to_id.pop(idx, None)
+            self.vectors.pop(node_id, None)
+
+    def save(self) -> None:
+        if not self.storage_path:
+            return
+
+        with self.lock:
+            self.storage_path.mkdir(parents=True, exist_ok=True)
+            if self.use_faiss and self.index is not None:
+                faiss.write_index(self.index, str(self.storage_path / "faiss.index"))
+                meta = {
+                    "id_to_idx": self.id_to_idx,
+                    "idx_to_id": self.idx_to_id,
+                    "next_idx": self.next_idx,
+                    "use_faiss": True,
+                    "dimension": self.dimension,
+                    "index_type": self.index_type,
+                }
+                with open(self.storage_path / "faiss_meta.pkl", "wb") as handle:
+                    pickle.dump(meta, handle)
+            else:
+                payload = {
+                    "vectors": {node_id: vector.tolist() for node_id, vector in self.vectors.items()},
+                    "use_faiss": False,
+                    "dimension": self.dimension,
+                    "index_type": self.index_type,
+                }
+                with open(self.storage_path / "vector_store.pkl", "wb") as handle:
+                    pickle.dump(payload, handle)
+
+    def reset(self) -> None:
+        with self.lock:
+            if self.use_faiss:
+                self.index = self._create_faiss_index()
+                self.id_to_idx.clear()
+                self.idx_to_id.clear()
+                self.next_idx = 0
+            else:
+                self.vectors.clear()
+
+            if self.storage_path and self.storage_path.exists():
+                for path in self.storage_path.glob("*"):
+                    if path.is_file():
+                        path.unlink()
+
+    def _load(self) -> None:
+        if not self.storage_path:
+            return
+
+        faiss_index = self.storage_path / "faiss.index"
+        faiss_meta = self.storage_path / "faiss_meta.pkl"
+        vector_store = self.storage_path / "vector_store.pkl"
+
+        if self.use_faiss and faiss_index.exists() and faiss_meta.exists():
+            try:
+                self.index = faiss.read_index(str(faiss_index))
+                with open(faiss_meta, "rb") as handle:
+                    meta = pickle.load(handle)
+                self.id_to_idx = meta.get("id_to_idx", {})
+                self.idx_to_id = meta.get("idx_to_id", {})
+                self.next_idx = meta.get("next_idx", 0)
+                return
+            except Exception:
+                self.use_faiss = False
+                self.index = None
+
+        if vector_store.exists():
+            with open(vector_store, "rb") as handle:
+                payload = pickle.load(handle)
+            self.use_faiss = False
+            self.vectors = {
+                int(node_id): np.array(vector, dtype=np.float32)
+                for node_id, vector in payload.get("vectors", {}).items()
+            }
+
+    def _create_faiss_index(self):
+        if not self.use_faiss:
+            return None
+        if self.index_type == "IP":
+            return faiss.IndexFlatIP(self.dimension)
+        return faiss.IndexFlatL2(self.dimension)
+
+    def _normalize(self, vectors: np.ndarray) -> np.ndarray:
+        if self.index_type == "IP":
+            norms = np.linalg.norm(vectors, axis=1, keepdims=True)
+            norms[norms == 0] = 1.0
+            return vectors / norms
+        return vectors