92 lines
2.5 KiB
Python
92 lines
2.5 KiB
Python
|
"""
|
||
|
|
Lightweight embedding utilities for local-only deployments.
|
||
|
|
|
||
|
|
We intentionally avoid heavyweight ML deps (torch/sentence-transformers) and external services.
|
||
|
|
This module provides a deterministic "hashed embedding" (similar to feature hashing):
|
||
|
|
- Tokenize text
|
||
|
|
- Hash tokens into a fixed-size dense vector
|
||
|
|
- L2 normalize
|
||
|
|
|
||
|
|
It is not as semantically strong as modern transformer embeddings, but it enables:
|
||
|
|
- Vector storage in SQLite
|
||
|
|
- Cosine similarity retrieval
|
||
|
|
- Recency/return weighted ranking
|
||
|
|
"""
|
||
|
|
|
||
|
|
from __future__ import annotations
|
||
|
|
|
||
|
|
import math
|
||
|
|
import os
|
||
|
|
import re
|
||
|
|
import struct
|
||
|
|
import hashlib
|
||
|
|
from typing import List, Optional
|
||
|
|
|
||
|
|
|
||
|
|
_TOKEN_RE = re.compile(r"[A-Za-z0-9_]+", re.UNICODE)
|
||
|
|
|
||
|
|
|
||
|
|
def _tokenize(text: str) -> List[str]:
|
||
|
|
t = (text or "").lower()
|
||
|
|
return _TOKEN_RE.findall(t)
|
||
|
|
|
||
|
|
|
||
|
|
class EmbeddingService:
|
||
|
|
"""Deterministic local embedding service."""
|
||
|
|
|
||
|
|
def __init__(self, dim: Optional[int] = None):
|
||
|
|
self.dim = int(dim or os.getenv("AGENT_MEMORY_EMBEDDING_DIM", "256") or 256)
|
||
|
|
if self.dim <= 0:
|
||
|
|
self.dim = 256
|
||
|
|
|
||
|
|
def embed(self, text: str) -> List[float]:
|
||
|
|
"""
|
||
|
|
Return a dense, L2-normalized embedding vector.
|
||
|
|
"""
|
||
|
|
vec = [0.0] * self.dim
|
||
|
|
tokens = _tokenize(text)
|
||
|
|
if not tokens:
|
||
|
|
return vec
|
||
|
|
|
||
|
|
# Feature hashing with signed counts
|
||
|
|
for tok in tokens:
|
||
|
|
h = hashlib.blake2b(tok.encode("utf-8"), digest_size=8).digest()
|
||
|
|
# Use first 8 bytes as unsigned int
|
||
|
|
v = int.from_bytes(h, "little", signed=False)
|
||
|
|
idx = v % self.dim
|
||
|
|
sign = -1.0 if ((v >> 63) & 1) else 1.0
|
||
|
|
vec[idx] += sign
|
||
|
|
|
||
|
|
# L2 normalize
|
||
|
|
norm = math.sqrt(sum(x * x for x in vec)) or 1.0
|
||
|
|
return [x / norm for x in vec]
|
||
|
|
|
||
|
|
def to_bytes(self, vec: List[float]) -> bytes:
|
||
|
|
"""
|
||
|
|
Pack float vector into little-endian float32 bytes for SQLite BLOB storage.
|
||
|
|
"""
|
||
|
|
if not vec:
|
||
|
|
return b""
|
||
|
|
return struct.pack("<" + "f" * len(vec), *[float(x) for x in vec])
|
||
|
|
|
||
|
|
def from_bytes(self, blob: bytes) -> List[float]:
|
||
|
|
if not blob:
|
||
|
|
return []
|
||
|
|
n = len(blob) // 4
|
||
|
|
if n <= 0:
|
||
|
|
return []
|
||
|
|
return list(struct.unpack("<" + "f" * n, blob[: n * 4]))
|
||
|
|
|
||
|
|
|
||
|
|
def cosine_sim(a: List[float], b: List[float]) -> float:
|
||
|
|
"""
|
||
|
|
Cosine similarity for L2-normalized vectors.
|
||
|
|
If vectors are not normalized, this becomes a scaled dot product.
|
||
|
|
"""
|
||
|
|
if not a or not b:
|
||
|
|
return 0.0
|
||
|
|
n = min(len(a), len(b))
|
||
|
|
return float(sum(a[i] * b[i] for i in range(n)))
|
||
|
|
|
||
|
|
|