Qdrant

Qdrant in OpenRAG

Qdrant is the high-performance vector database used for semantic search in OpenRAG.

Configuration

Image: qdrant/qdrant:latest Ports:

6333: HTTP API
6334: gRPC API

Volume: qdrant_data:/qdrant/storage Dashboard: http://localhost:6333/dashboard

Vector Configuration

Dimension: 384 (matches sentence-transformers all-MiniLM-L6-v2) Distance Metric: Cosine similarity Index Type: HNSW (Hierarchical Navigable Small World)

Collections

OpenRAG uses Qdrant collections to organize vectors: Default Collection: default Collection Parameters:

{
    "vectors": {
        "size": 384,
        "distance": "Cosine"
    },
    "optimizers_config": {
        "indexing_threshold": 20000
    },
    "hnsw_config": {
        "m": 16,
        "ef_construct": 100
    }
}

REST API Usage

List Collections:

curl http://localhost:6333/collections | jq

Response:

{
  "result": {
    "collections": [
      {
        "name": "default"
      }
    ]
  },
  "status": "ok",
  "time": 0.000123
}

Get Collection Info:

curl http://localhost:6333/collections/default | jq

Response:

{
  "result": {
    "status": "green",
    "points_count": 928,
    "indexed_vectors_count": 928,
    "vectors_count": 928,
    "config": {
      "params": {
        "vectors": {
          "size": 384,
          "distance": "Cosine"
        }
      }
    }
  }
}

Search Vectors:

curl -X POST http://localhost:6333/collections/default/points/search \
  -H "Content-Type: application/json" \
  -d '{
    "vector": [0.1, 0.2, ...],  # 384 dimensions
    "limit": 5,
    "with_payload": true
  }'

Insert Point:

curl -X PUT http://localhost:6333/collections/default/points \
  -H "Content-Type: application/json" \
  -d '{
    "points": [
      {
        "id": "550e8400-e29b-41d4-a716-446655440000",
        "vector": [0.1, 0.2, ...],
        "payload": {
          "document_id": "abc123",
          "chunk_index": 0,
          "text": "Sample text..."
        }
      }
    ]
  }'

Python Client Usage

OpenRAG uses the qdrant-client Python library:

from qdrant_client import QdrantClient
from qdrant_client.models import Distance, VectorParams, PointStruct

client = QdrantClient(host="qdrant", port=6333)

# Create collection
client.create_collection(
    collection_name="documents_embeddings",
    vectors_config=VectorParams(
        size=384,
        distance=Distance.COSINE
    )
)

# Insert vectors
client.upsert(
    collection_name="documents_embeddings",
    points=[
        PointStruct(
            id=str(uuid.uuid4()),
            vector=embedding,
            payload={
                "document_id": doc_id,
                "chunk_index": idx,
                "content": chunk_text
            }
        )
    ]
)

# Search
results = client.search(
    collection_name="documents_embeddings",
    query_vector=query_embedding,
    limit=5,
    score_threshold=0.3
)

Current Data in Qdrant

Test Results (after uploading 31 WTE documents):

curl http://localhost:6333/collections/default | jq '.result | {points: .points_count, status: .status}'

Output:

{
  "points": 928,
"status": "green"
}

Breakdown:

Documents processed: 28
Total chunks: 928
Average chunks per document: 33
Vector dimension: 384
Storage status: green (healthy)

Search Performance

Search Latency: 50-150ms (typical) Throughput: Hundreds of searches per second Accuracy: Cosine similarity scores range from 0.0 (orthogonal) to 1.0 (identical) Typical Score Thresholds:

0.7+: Highly relevant
0.5-0.7: Moderately relevant
0.3-0.5: Potentially relevant
<0.3: Not relevant (filtered out in OpenRAG)

Dashboard Features

Access at http://localhost:6333/dashboard Features:

Collection browser
Point inspector
Search testing
Performance metrics
Configuration viewer

Collection Management

Create New Collection:

curl -X PUT http://localhost:6333/collections/my_collection \
  -H "Content-Type: application/json" \
  -d '{
    "vectors": {
      "size": 384,
      "distance": "Cosine"
    }
  }'

Delete Collection:

curl -X DELETE http://localhost:6333/collections/my_collection

Collection Aliases:

curl -X POST http://localhost:6333/collections/aliases \
  -H "Content-Type: application/json" \
  -d '{
    "actions": [
      {
        "create_alias": {
          "collection_name": "default",
          "alias_name": "production"
        }
      }
    ]
  }'

Filtering

Qdrant supports payload filtering:

results = client.search(
    collection_name="default",
    query_vector=embedding,
    query_filter={
        "must": [
            {
                "key": "category",
                "match": {
                    "value": "cisco_phones"
                }
            }
        ]
    },
    limit=10
)

Optimization

HNSW Parameters:

m: Number of connections per layer (16 recommended)
ef_construct: Construction time/accuracy tradeoff (100-200 recommended)
ef: Search time/accuracy tradeoff (dynamic, typically 128)

Indexing Threshold:

Points indexed after reaching threshold
Default: 20,000 points
Lower for better search accuracy, higher for faster insertions

Backup and Restore

Create Snapshot:

curl -X POST http://localhost:6333/collections/default/snapshots

List Snapshots:

curl http://localhost:6333/collections/default/snapshots | jq

Download Snapshot:

curl -o snapshot.tar http://localhost:6333/collections/default/snapshots/snapshot_name

Restore (via volume mount):

# Stop Qdrant
sudo docker-compose stop qdrant

# Copy snapshot to volume
sudo docker cp snapshot.tar openrag-qdrant:/qdrant/storage/

# Restart
sudo docker-compose start qdrant

Monitoring

Collection Stats:

curl http://localhost:6333/collections/default | jq '.result | {
  points: .points_count,
  indexed: .indexed_vectors_count,
  segments: .segments_count,
  status: .status
}'

Cluster Info (if using cluster mode):

curl http://localhost:6333/cluster | jq

Troubleshooting

No Results Returned:

Check score_threshold (try lowering to 0.2)
Verify vector dimensions match (384)
Ensure collection has points

Slow Search:

Increase ef parameter
Optimize HNSW configuration
Check system resources

Storage Issues:

Monitor disk space
Create snapshots and cleanup old data
Consider collection partitioning

View Logs:

sudo docker logs openrag-qdrant --tail=100

API Reference

Full Qdrant API documentation: https://qdrant.tech/documentation/ Common endpoints used in OpenRAG:

GET /collections: List all collections
GET /collections/{name}: Collection info
POST /collections/{name}/points/search: Vector search
PUT /collections/{name}/points: Insert/update points
DELETE /collections/{name}/points: Delete points

Getting Started

API Reference

Components

Experiments & Research

Tests

Qdrant in OpenRAG

Configuration

Vector Configuration

Collections

REST API Usage

Python Client Usage

Current Data in Qdrant

Search Performance

Dashboard Features

Collection Management

Filtering

Optimization

Backup and Restore

Monitoring

Troubleshooting

API Reference

​Qdrant in OpenRAG

​Configuration

​Vector Configuration

​Collections

​REST API Usage

​Python Client Usage

​Current Data in Qdrant

​Search Performance

​Dashboard Features

​Collection Management

​Filtering

​Optimization

​Backup and Restore

​Monitoring

​Troubleshooting

​API Reference

Qdrant in OpenRAG

Configuration

Vector Configuration

Collections

REST API Usage

Python Client Usage

Current Data in Qdrant

Search Performance

Dashboard Features

Collection Management

Filtering

Optimization

Backup and Restore

Monitoring

Troubleshooting

API Reference