Spaces:
Running
Running
File size: 10,667 Bytes
daafb32 f780124 daafb32 f780124 daafb32 f780124 daafb32 | 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 | """
Qdrant vector database interface for ResearchPilot.
RUNS LOCALLY - no server needed, no Docker, no cloud account.
Qdrant client in local mode stores everything in a directory
on disk, exactly like SQLite does for relational data.
Data lives in: data/qdrant_db/
"""
import json
import uuid
import numpy as np
from pathlib import Path
from typing import Optional
from qdrant_client import QdrantClient
from qdrant_client.models import (
Distance,
VectorParams,
PointStruct,
Filter,
FieldCondition,
MatchValue,
Range,
SearchRequest,
)
from tqdm import tqdm
from src.utils.logger import get_logger
from config.settings import (
QDRANT_COLLECTION_NAME,
QDRANT_PATH,
EMBEDDING_DIMENSION,
TOP_K_RETRIEVAL,
)
logger = get_logger(__name__)
# How many points to upload to Qdrant at once
# Too large = memory spike. Too small = many round trips.
UPSERT_BATCH_SIZE = 256
class QdrantStore:
"""
Manages the Qdrant vector database for chunk storage and retrieval.
UPSERT PATTERN:
We use 'upsert' (update + insert) instead of 'insert'.
If a chunk already exists, upsert updates it.
If it doesn't exist, upsert creates it.
This makes our indexing pipeline idempotent - safe to re-run.
"""
def __init__(self):
# Local mode: pass path= instead of url=
# Qdrant creates/opens a local database at this path
# No server process needed - runs in-process
logger.info(f"Connecting to local Qdrant at: {QDRANT_PATH}")
self.client = QdrantClient(path = QDRANT_PATH)
self.collection_name = QDRANT_COLLECTION_NAME
def collection_exists(self) -> bool:
"""Check if our collection already exists in Qdrant."""
collections = self.client.get_collections().collections
names = [c.name for c in collections]
return self.collection_name in names
def get_collection_size(self) -> int:
"""Return number of points currently in the collections."""
if not self.collection_exists():
return 0
info = self.client.get_collection(self.collection_name)
return info.points_count
def create_collection(self, recreate: bool = False):
"""
Create the Qdrant collection for research paper chunks.
Args:
recreate: If True, delete existing collection and rebuild.
Use this when you want a clean re-index.
COLLECTION CONFIGURATION:
size=768 -> matches BGE-base-en-v1.5 output dimension
distance=COSINE -> similarity metric
WHY COSINE DISTANCE:
Our embeddings are L2-normalized (magnitude = 1.0).
For normalized vectors: cosine_similarity = dot_product
Qdrant's COSINE metric handles this correctly.
Using DOT_PRODUCT would also work but COSINE is more explicit.
"""
if self.collection_exists():
if recreate:
logger.warning(f"Deleting existing collection: {self.collection_name}")
self.client.delete_collection(self.collection_name)
else:
size = self.get_collection_size()
logger.info(
f"Collection: '{self.collection_name}' already exists "
f"with {size:,} points. Skipping creation."
)
return
logger.info(f"Creating collection: {self.collection_name}")
self.client.create_collection(
collection_name = self.collection_name,
vectors_config = VectorParams(
size = EMBEDDING_DIMENSION,
distance = Distance.COSINE,
),
)
logger.info(f"Collection created: {self.collection_name}")
def index_chunks(
self,
embeddings: np.ndarray,
chunk_ids: list[str],
metadata: list[dict],
texts: list[str]
) -> int:
"""
Upload embeddings + metadata into Qdrant.
Args:
embeddings: numpy array (N, 768)
chunk_ids: list of N chunk ID strings
metadata: list of N metadata dicts
texts: list of N chunk text strings
Returns:
Number of points successfully indexed
QDRANT POINT STRUCTURE:
Each point needs:
- id: unique identifier (we use the chunk_id UUID)
- vector: the embedding as a Python list of floats
- payload: dict of any metadata we want to store/filter
WHY INCLUDE TEXT IN PAYLOAD:
When we retrieve a point, we need the text to show to the
user and to send to the LLM. Storing it in the payload
means ONE database query returns everything we need.
Alternative would be a separate text lookup - slower and
more complex.
"""
assert len(embeddings) == len(chunk_ids) == len(metadata) == len(texts), \
"All inputs must have the same length"
total_indexed = 0
# Process in batches to avoid memory spikes
for batch_start in tqdm(
range(0, len(embeddings), UPSERT_BATCH_SIZE),
desc = "Indexing into Qdrant"
):
batch_end = min(batch_start + UPSERT_BATCH_SIZE, len(embeddings))
# Build PointStruct objects for this batch
points = []
for i in range(batch_start, batch_end):
# Qdrant requires UUID format for point IDs
# Our chunk_ids are already UUIDs from Phase 5
point = PointStruct(
id = chunk_ids[i],
vector = embeddings[i].tolist(), # Numpy -> Python List
payload = {
# Store ALL metadata in payload for retrieval
**metadata[i],
"text": texts[i], # Include chunk text
"publication_year": int(metadata[i].get("published_date", "0000")[:4]),
}
)
points.append(point)
# Upsert this batch
self.client.upsert(
collection_name = self.collection_name,
points = points,
)
total_indexed += len(points)
logger.info(
f"Indexing complete. "
f"Total points in collection: {self.get_collection_size():,}"
)
return total_indexed
def search(
self,
query_vector: np.ndarray,
top_k: int = TOP_K_RETRIEVAL,
filter_category: Optional[str] = None,
filter_year_gte: Optional[int] = None,
) -> list[dict]:
"""
Search for most similar chunks to a query vector.
Args:
query_vector: 768-dimensional query embedding
top_k: How many results to return
filter_category: Only return chunks from this ArXiv category
filter_year_gte: Only return chunks from this year or later
Returns:
List of result dicts, each containing:
{
"chunk_id": str,
"score": float (cosine similarity, 0-1),
"text": str,
"paper_id": str,
"title": str,
"authors": list,
"published_date": str,
...all other payload fields
}
FILTERING IN QDRANT:
Qdrant applies metadata filters DURING vector search,
not after. This means it only scores vectors that match
the filter - much faster than post-filtering.
Example: filter_year_gte=2024 means:
"Find the top-20 most similar vectors, but ONLY consider
vectors from papers published in 2024 or later"
"""
# Build optional filter
qdrant_filter = self._build_filter(filter_category, filter_year_gte)
# Execute search
results = self.client.query_points(
collection_name = self.collection_name,
query = query_vector.tolist(),
limit = top_k,
query_filter = qdrant_filter,
with_payload = True, # Return metadata with results
with_vectors = False # Don't return the vectors (saves bandwidth)
).points
# Convert Qdrant ScoredPoint objects to plain dicts
return [
{
"chunk_id": str(r.id),
"score" : round(r.score, 4),
**r.payload, # Unpack all payload fields (text, title, etc.)
}
for r in results
]
def _build_filter(
self,
category: Optional[str],
year_gte: Optional[int],
) -> Optional[Filter]:
"""
Build a Qdrant filter from optional parameters.
Returns None if no filters specified (search everything).
QDRANT FILTER SYNTAX:
Filter(must=[condition1, condition2])
means: results must satisfy condition1 AND condition2
MatchValue -> exact match (equality check)
Range -> numeric range (gte, lte, gt, lt)
"""
conditions = []
if category:
conditions.append(
FieldCondition(
key = "primary_category",
match = MatchValue(value = category)
)
)
if year_gte:
# publication_year is stored as an integer (e.g. 2026)
# Range(gte=year_gte) filters to papers from that year onwards
conditions.append(
FieldCondition(
key = "publication_year",
range = Range(gte = year_gte)
)
)
if not conditions:
return None
return Filter(must = conditions)
def get_collection_info(self) -> dict:
"""Return summary information about the collection."""
if not self.collection_exists():
return {"status": "collection_not_found"}
info = self.client.get_collection(self.collection_name)
return {
"collection_name": self.collection_name,
"points_count" : info.points_count,
"status" : str(info.status),
"vector_size" : info.config.params.vectors.size,
"distance" : str(info.config.params.vectors.distance),
} |