File size: 8,120 Bytes
9c940cd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""
Load CVE-KGRAG chunks into Qdrant with dense (transformer) + sparse (BM25) vectors.

Standalone — does not require the upstream CVE-KGRAG codebase. Creates four
collections: cve_chunks, mitre_techniques, capec_patterns, cwe_entries.

Each point carries a named "dense" vector (640-dim by default, microsoft/harrier-oss-v1-270m)
and a named "sparse" vector encoded from bm25/vocab.json.

Resumes from checkpoint at <kb-dir>/bm25/qdrant_checkpoint.json so you can re-run
after an interruption without re-encoding everything.

Usage:
    python load_qdrant.py --kb-dir ./data/knowledge_base \
        --qdrant-url http://localhost:6333 \
        --model microsoft/harrier-oss-v1-270m
"""
from __future__ import annotations

import argparse
import json
import logging
import math
import re
import sys
import uuid
from collections import Counter
from pathlib import Path
from typing import Any, Dict, Iterable, List, Tuple

from qdrant_client import QdrantClient
from qdrant_client.models import (
    Distance, PointStruct, SparseIndexParams, SparseVector,
    SparseVectorParams, VectorParams,
)
from sentence_transformers import SentenceTransformer
from tqdm import tqdm

logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
log = logging.getLogger("load_qdrant")

COLLECTIONS = {
    "cve":   "cve_chunks",
    "mitre": "mitre_techniques",
    "capec": "capec_patterns",
    "cwe":   "cwe_entries",
}
CHUNK_FILES = {
    "cve":   "rag_exports/cve_chunks.json",
    "mitre": "rag_exports/mitre_chunks.json",
    "capec": "rag_exports/capec_chunks.json",
    "cwe":   "rag_exports/cwe_chunks.json",
}
DENSE = "dense"
SPARSE = "sparse"
TOKEN_RE = re.compile(r"[a-z0-9][a-z0-9_-]*", re.IGNORECASE)
K1 = 1.5
B = 0.75


def tokenize(text: str) -> List[str]:
    return [t.lower() for t in TOKEN_RE.findall(text or "")]


def point_uuid(chunk_id: str) -> str:
    return str(uuid.uuid5(uuid.NAMESPACE_URL, chunk_id))


class SparseEncoder:
    """BM25-weighted sparse vectors using the prebuilt vocab.json."""

    def __init__(self, vocab_path: Path, avgdl: float = 200.0) -> None:
        with vocab_path.open("r", encoding="utf-8") as f:
            raw = json.load(f)
        self.token_to_id = {t: int(v["id"]) for t, v in raw.items()}
        self.idf = {int(v["id"]): float(v["idf"]) for v in raw.values()}
        self.avgdl = avgdl

    def encode(self, text: str) -> SparseVector:
        tokens = tokenize(text)
        dl = len(tokens) or 1
        tf = Counter(tokens)
        indices: List[int] = []
        values: List[float] = []
        for tok, count in tf.items():
            tid = self.token_to_id.get(tok)
            if tid is None:
                continue
            num = count * (K1 + 1)
            den = count + K1 * (1 - B + B * dl / self.avgdl)
            indices.append(tid)
            values.append(self.idf[tid] * num / den)
        return SparseVector(indices=indices, values=values)


def ensure_collections(client: QdrantClient, dense_dim: int) -> None:
    existing = {c.name for c in client.get_collections().collections}
    for name in COLLECTIONS.values():
        if name in existing:
            log.info(f"  {name} already exists")
            continue
        client.create_collection(
            collection_name=name,
            vectors_config={DENSE: VectorParams(size=dense_dim, distance=Distance.COSINE)},
            sparse_vectors_config={SPARSE: SparseVectorParams(index=SparseIndexParams())},
        )
        log.info(f"  created {name}")


def sanitize_payload(p: Dict[str, Any]) -> Dict[str, Any]:
    """Qdrant indexes scalars and flat lists fine; nested dicts become JSON strings."""
    out: Dict[str, Any] = {}
    for k, v in p.items():
        if isinstance(v, (str, int, float, bool)) or v is None:
            out[k] = v
        elif isinstance(v, list) and all(isinstance(x, (str, int, float, bool)) for x in v):
            out[k] = v
        else:
            out[k] = json.dumps(v, ensure_ascii=False)
    return out


def upsert_collection(
    client: QdrantClient,
    collection: str,
    chunks: List[Dict[str, Any]],
    dense_model: SentenceTransformer,
    sparse: SparseEncoder,
    *,
    batch: int = 2000,
    resume_from: int = 0,
) -> int:
    total = len(chunks)
    log.info(f"  → {collection}: {total - resume_from:,} / {total:,} chunks to encode")
    for start in tqdm(range(resume_from, total, batch), desc=collection):
        end = min(start + batch, total)
        slab = chunks[start:end]
        texts = [c["text"] for c in slab]
        dense = dense_model.encode(texts, batch_size=128, show_progress_bar=False,
                                   convert_to_numpy=True, normalize_embeddings=True)
        points = []
        for c, d in zip(slab, dense):
            points.append(PointStruct(
                id=point_uuid(c["id"]),
                vector={
                    DENSE: d.tolist(),
                    SPARSE: sparse.encode(c["text"]),
                },
                payload={**sanitize_payload(c.get("payload") or {}), "chunk_id": c["id"]},
            ))
        client.upsert(collection_name=collection, points=points)
        yield end


def load_chunks(path: Path) -> List[Dict[str, Any]]:
    with path.open("r", encoding="utf-8") as f:
        return json.load(f)


def main() -> int:
    p = argparse.ArgumentParser(description=__doc__)
    p.add_argument("--kb-dir", required=True, type=Path,
                   help="Knowledge-base root (containing rag_exports/ and bm25/)")
    p.add_argument("--qdrant-url", default="http://localhost:6333")
    p.add_argument("--qdrant-api-key", default=None)
    p.add_argument("--model", default="microsoft/harrier-oss-v1-270m",
                   help="HF sentence-transformers model for dense vectors")
    p.add_argument("--device", default=None, help="cuda|cpu (auto if omitted)")
    p.add_argument("--batch", type=int, default=2000)
    p.add_argument("--recreate", action="store_true",
                   help="Drop and recreate collections before loading")
    args = p.parse_args()

    if not (args.kb_dir / "rag_exports").is_dir():
        log.error(f"--kb-dir missing rag_exports/: {args.kb_dir}")
        return 2

    log.info(f"Loading dense model: {args.model}")
    dense_model = SentenceTransformer(args.model, device=args.device)
    dense_dim = dense_model.get_sentence_embedding_dimension()
    log.info(f"  dim={dense_dim}")

    log.info("Loading sparse encoder")
    sparse = SparseEncoder(args.kb_dir / "bm25" / "vocab.json")

    client = QdrantClient(url=args.qdrant_url, api_key=args.qdrant_api_key, timeout=60)

    if args.recreate:
        log.info("Recreating collections")
        for name in COLLECTIONS.values():
            try:
                client.delete_collection(name)
            except Exception:
                pass
    ensure_collections(client, dense_dim)

    ckpt_path = args.kb_dir / "bm25" / "qdrant_checkpoint.json"
    ckpt: Dict[str, int] = {}
    if ckpt_path.exists():
        ckpt = json.loads(ckpt_path.read_text())
        log.info(f"Resuming from checkpoint: {ckpt}")

    for key, rel in CHUNK_FILES.items():
        path = args.kb_dir / rel
        if not path.exists():
            log.warning(f"  {path} missing, skipping {key}")
            continue
        chunks = load_chunks(path)
        col = COLLECTIONS[key]
        resume = ckpt.get(col, 0) if not args.recreate else 0
        if resume >= len(chunks):
            log.info(f"  {col}: already complete ({resume:,}/{len(chunks):,})")
            continue
        for done in upsert_collection(
            client, col, chunks, dense_model, sparse,
            batch=args.batch, resume_from=resume,
        ):
            ckpt[col] = done
            ckpt_path.write_text(json.dumps(ckpt))
        log.info(f"  {col}: complete")

    log.info("Done. Counts:")
    for name in COLLECTIONS.values():
        info = client.get_collection(name)
        log.info(f"  {name}: {info.points_count:,} points")
    return 0


if __name__ == "__main__":
    sys.exit(main())