File size: 13,020 Bytes
effde1c |
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 |
"""
Simple pluggable VectorStore with a FAISS adapter and a numpy brute-force fallback.
This file provides:
- EmbeddingAdapter: deterministic text->vector adapter for development.
- VectorStore: in-memory store that uses FAISS when available.
- get_global_vector_store(): convenience singleton for the app to reuse.
Designed to be lightweight and safe to import when FAISS is not installed.
"""
from typing import Optional, Dict, Any, List, Tuple
import hashlib
import numpy as np
try:
import faiss
except Exception:
faiss = None
class EmbeddingAdapter:
"""Deterministic embedding adapter for development.
It hashes the input text and produces a fixed-size float vector. Not
a production-quality embedder but useful for development and tests.
"""
def __init__(self, dim: int = 128):
self.dim = dim
def embed(self, text: str) -> np.ndarray:
h = hashlib.sha256(text.encode("utf-8")).digest()
# Expand digest material to required dim by repeating digest
needed = self.dim
data = (h * ((needed * 32) // len(h) + 1))[:needed]
arr = np.frombuffer(data, dtype=np.uint8).astype(np.float32)
# normalize to unit vector
if arr.sum() == 0:
return np.zeros(self.dim, dtype=np.float32)
vec = arr / np.linalg.norm(arr)
return vec
class VectorStore:
def __init__(self, dim: int = 128):
self.dim = dim
self._emb = EmbeddingAdapter(dim=dim)
self._meta: Dict[str, Dict[str, Any]] = {}
self._vectors: Dict[str, np.ndarray] = {}
self._faiss_index = None
self._use_faiss = False
self._build_index()
def _build_index(self):
if faiss is None:
self._use_faiss = False
self._faiss_index = None
return
try:
index = faiss.IndexFlatL2(self.dim)
self._faiss_index = index
self._use_faiss = True
except Exception:
self._use_faiss = False
self._faiss_index = None
def add_vector(self, id: str, vector: np.ndarray, metadata: Optional[Dict[str, Any]] = None):
v = np.asarray(vector, dtype=np.float32)
if v.shape != (self.dim,):
raise ValueError(f"vector must have shape ({self.dim},), got {v.shape}")
self._vectors[id] = v
self._meta[id] = metadata or {}
if self._use_faiss and self._faiss_index is not None:
try:
# faiss expects a 2D array
self._faiss_index.add(np.expand_dims(v, axis=0))
except Exception:
# fallback: rebuild index
self._rebuild_faiss_index()
def add_text(self, id: str, text: str, metadata: Optional[Dict[str, Any]] = None):
vec = self._emb.embed(text)
self.add_vector(id, vec, metadata)
def _rebuild_faiss_index(self):
if faiss is None:
return
try:
index = faiss.IndexFlatL2(self.dim)
if len(self._vectors) > 0:
mats = np.stack(list(self._vectors.values(), axis=0).astype(np.float32))
index.add(mats)
self._faiss_index = index
self._use_faiss = True
except Exception:
self._faiss_index = None
self._use_faiss = False
def get(self, id: str) -> Optional[Dict[str, Any]]:
if id not in self._vectors:
return None
return {"id": id, "vector": self._vectors[id], "metadata": self._meta.get(id, {})}
def query_vector(self, vector: np.ndarray, k: int = 5) -> List[Tuple[str, float, Dict[str, Any]]]:
v = np.asarray(vector, dtype=np.float32)
if self._use_faiss and self._faiss_index is not None:
D, I = self._faiss_index.search(np.expand_dims(v, axis=0), k)
# faiss returns distances and indices
results: List[Tuple[str, float, Dict[str, Any]]] = []
ids = list(self._vectors.keys())
for dist, idx in zip(D[0], I[0]):
if idx < 0 or idx >= len(ids):
continue
rid = ids[idx]
results.append((rid, float(dist), self._meta.get(rid, {})))
return results
# fallback brute-force
results = []
for rid, rv in self._vectors.items():
dist = float(np.linalg.norm(rv - v))
results.append((rid, dist, self._meta.get(rid, {})))
results.sort(key=lambda x: x[1])
return results[:k]
def query_text(self, text: str, k: int = 5) -> List[Tuple[str, float, Dict[str, Any]]]:
vec = self._emb.embed(text)
return self.query_vector(vec, k=k)
# simple global store for the app
_GLOBAL_STORE: Optional[VectorStore] = None
def get_global_vector_store() -> VectorStore:
global _GLOBAL_STORE
if _GLOBAL_STORE is None:
_GLOBAL_STORE = VectorStore()
return _GLOBAL_STORE
"""Simple pluggable vector store with FAISS backend and numpy fallback.
This file provides a minimal VectorStore interface used by the REST API.
It intentionally keeps dependencies optional: if `faiss` isn't installed the
implementation falls back to an in-memory numpy-based nearest-neighbour search
for development and testing.
"""
from typing import List, Optional, Dict, Any
try:
import faiss
FAISS_AVAILABLE = True
except Exception:
faiss = None # type: ignore
FAISS_AVAILABLE = False
import numpy as np
class VectorStore:
"""A tiny vector store abstraction.
- `add(ids, vectors, metas)` stores vectors and optional metadata.
- `search(query_vector, top_k)` returns nearest neighbours with scores.
"""
def __init__(self, dim: int = 128):
self.dim = dim
if FAISS_AVAILABLE:
# Use IndexFlatL2 for simplicity (no IDs support so we map manually)
self.index = faiss.IndexFlatL2(dim)
self._id_map: Dict[int, Any] = {}
self._next_index = 0
else:
self.vectors = np.zeros((0, dim), dtype=np.float32)
self.ids: List[str] = []
self.metas: Dict[str, Any] = {}
def add(self, ids: List[str], vectors: np.ndarray, metas: Optional[List[Any]] = None) -> int:
"""Add vectors to the store.
Args:
ids: list of string IDs (one per vector).
vectors: numpy array of shape (N, dim).
metas: optional list of metadata objects parallel to ids.
Returns:
number of indexed vectors after insertion.
"""
vecs = np.asarray(vectors, dtype=np.float32)
if vecs.ndim != 2 or vecs.shape[1] != self.dim:
raise ValueError(f"vectors must be shape (N, {self.dim})")
if FAISS_AVAILABLE:
self.index.add(vecs)
for i, id_ in enumerate(ids):
self._id_map[self._next_index] = {"id": id_, "meta": metas[i] if metas else None}
self._next_index += 1
return int(self.index.ntotal)
else:
if self.vectors.size == 0:
self.vectors = vecs
else:
self.vectors = np.vstack([self.vectors, vecs])
self.ids.extend(ids)
if metas:
for i, id_ in enumerate(ids):
self.metas[id_] = metas[i]
return len(self.ids)
def search(self, query_vector: np.ndarray, top_k: int = 5) -> List[Dict[str, Any]]:
"""Return nearest neighbours as a list of {id, score, meta}.
Score is L2 distance (lower is better).
"""
q = np.asarray(query_vector, dtype=np.float32)
if q.ndim == 1:
q = q.reshape(1, -1)
if q.shape[1] != self.dim:
raise ValueError(f"query_vector must have dimension {self.dim}")
if FAISS_AVAILABLE:
D, I = self.index.search(q, top_k)
results = []
for dist, idx in zip(D[0], I[0]):
if idx < 0:
continue
info = self._id_map.get(int(idx), {"id": str(idx), "meta": None})
results.append({"id": info["id"], "score": float(dist), "meta": info.get("meta")})
return results
else:
if self.vectors.shape[0] == 0:
return []
# compute L2 distances
diffs = self.vectors - q
dists = np.linalg.norm(diffs, axis=1)
idxs = np.argsort(dists)[:top_k]
out = []
for i in idxs:
out.append({"id": self.ids[int(i)], "score": float(dists[int(i)]), "meta": self.metas.get(self.ids[int(i)])})
return out
_default_store: Optional[VectorStore] = None
def get_default_store(dim: int = 128) -> VectorStore:
global _default_store
if _default_store is None:
_default_store = VectorStore(dim=dim)
return _default_store
"""Simple pluggable vector store with FAISS backend (optional) and numpy brute-force fallback.
This module provides a lightweight interface used by the API for indexing and nearest-neighbor
search. FAISS is optional; if it's not installed the implementation falls back to an in-memory
brute-force search using NumPy (if available) or pure Python.
"""
from typing import List, Dict, Optional, Tuple
try:
import faiss
_has_faiss = True
except Exception:
faiss = None # type: ignore
_has_faiss = False
try:
import numpy as np
_has_numpy = True
except Exception:
np = None # type: ignore
_has_numpy = False
class VectorStore:
"""In-memory vector store with optional FAISS acceleration.
Usage:
store = VectorStore(dim=128)
store.add('id1', vector, metadata={...})
results = store.search(query_vector, k=5)
"""
def __init__(self, dim: int = 128, use_faiss: bool = True):
self.dim = dim
self.ids: List[str] = []
self.vectors: List = [] # numpy arrays if available, else lists
self.metadatas: List[Dict] = []
self._index = None
self._use_faiss = use_faiss and _has_faiss
if self._use_faiss:
# Use inner product (cosine if vectors normalized externally)
self._index = faiss.IndexFlatIP(dim)
def _ensure_numpy(self, vec):
if _has_numpy:
return np.asarray(vec, dtype=np.float32)
return vec
def add(self, id: str, vector, metadata: Optional[Dict] = None):
metadata = metadata or {}
vec = self._ensure_numpy(vector)
self.ids.append(id)
self.vectors.append(vec)
self.metadatas.append(metadata)
if self._use_faiss:
# faiss needs contiguous float32 arrays
arr = np.asarray(vec, dtype=np.float32).reshape(1, -1)
self._index.add(arr)
def search(self, query_vector, k: int = 5) -> List[Tuple[str, float, Dict]]:
"""Return list of (id, score, metadata) ordered by descending score.
Score semantics: if FAISS IndexFlatIP is used, it's inner product. The
fallback uses cosine similarity when numpy is available.
"""
if len(self.ids) == 0:
return []
q = self._ensure_numpy(query_vector)
if self._use_faiss:
q_arr = np.asarray(q, dtype=np.float32).reshape(1, -1)
D, I = self._index.search(q_arr, min(k, len(self.ids)))
results = []
for score, idx in zip(D[0].tolist(), I[0].tolist()):
if idx < 0:
continue
results.append((self.ids[idx], float(score), self.metadatas[idx]))
return results
# numpy brute-force fallback
if _has_numpy:
mats = np.vstack([np.asarray(v, dtype=np.float32).reshape(1, -1) for v in self.vectors])
qv = np.asarray(q, dtype=np.float32).reshape(-1)
# cosine similarity
norms = np.linalg.norm(mats, axis=1) * (np.linalg.norm(qv) + 1e-12)
sims = (mats.dot(qv)) / (norms + 1e-12)
idxs = sims.argsort()[::-1][:k]
return [(self.ids[i], float(sims[i]), self.metadatas[i]) for i in idxs]
# pure-python fallback: dot product over lists
def dot(a, b):
return sum(x * y for x, y in zip(a, b))
scores = [dot(v, q) for v in self.vectors]
ordered = sorted(range(len(scores)), key=lambda i: scores[i], reverse=True)[:k]
return [(self.ids[i], float(scores[i]), self.metadatas[i]) for i in ordered]
# provide a module-level default store for simple usage
default_store = VectorStore(dim=128, use_faiss=True)
__all__ = ["VectorStore", "default_store"]
|