qdrant-advanced（高级 qdrant）

Qdrant Advanced

Production-ready Qdrant vector database operations for AI agents. Complete toolkit for semantic search, document ingestion, collection management, backups, and migrations.

Quick Start

# Set environment variables
export QDRANT_HOST="localhost"
export QDRANT_PORT="6333"
export OPENAI_API_KEY="sk-..."

# List collections
bash manage.sh list

# Create a collection
bash manage.sh create my_collection 1536 cosine

# Ingest a document
bash ingest.sh /path/to/document.txt my_collection paragraph

# Search
bash search.sh "my search query" my_collection 5

Scripts Overview

Environment Variables

*Required for ingest and search operations

Detailed Usage

Semantic Search

bash search.sh <query> <collection> [limit] [filter_json]

Examples:

# Basic search
bash search.sh "machine learning tutorials" my_docs 10

# With metadata filter
bash search.sh "deployment guide" my_docs 5 '{"must": [{"key": "category", "match": {"value": "devops"}}]}'

# Score threshold
bash search.sh "error handling" my_docs 10 "" 0.8

Output:

{
  "results": [
    {
      "id": "doc-001",
      "score": 0.92,
      "text": "When handling errors in production...",
      "metadata": {"source": "docs/error-handling.md"}
    }
  ]
}

Document Ingestion

bash ingest.sh <file_path> <collection> [chunk_strategy] [metadata_json]

Chunk Strategies:

\ ) | Articles, docs | | sentence | Split by sentences | Short content | | fixed | Fixed 1000 char chunks | Code, logs | | semantic | Semantic boundaries | Long documents |

Examples:

# Ingest with paragraph chunking
bash ingest.sh article.md my_collection paragraph

# With custom metadata
bash ingest.sh api.md my_collection paragraph '{"category": "api", "version": "2.0"}'

# Ingest multiple files
for f in docs/*.md; do
    bash ingest.sh "$f" my_collection paragraph
done

Collection Management

bash manage.sh <command> [args...]

Commands:

Examples:

bash manage.sh list
bash manage.sh create my_vectors 1536 cosine
bash manage.sh create my_vectors 768 euclid
bash manage.sh info my_vectors
bash manage.sh optimize my_vectors
bash manage.sh delete my_vectors

Backup & Restore

bash backup.sh <command> [args...]

Commands:

Examples:

# Create snapshot
bash backup.sh snapshot my_collection
bash backup.sh snapshot my_collection backup_2026_02_10

# List snapshots
bash backup.sh list my_collection

# Restore
bash backup.sh restore my_collection backup_2026_02_10

# Delete old snapshot
bash backup.sh delete my_collection old_backup

Migration

bash migrate.sh <source_collection> <target_collection> [options]

Migration Types:

1. Copy Collection: Same embedding model, different name
2. Model Upgrade: Upgrade to new embedding model (re-embeds)
3. Filter Migration: Migrate subset with filter

Examples:

# Simple copy
bash migrate.sh old_collection new_collection

# With model upgrade (re-embeds all content)
bash migrate.sh old_collection new_collection --upgrade-model

# Filtered migration
bash migrate.sh old_collection new_collection --filter '{"category": "public"}'

# Batch size for large collections
bash migrate.sh old_collection new_collection --batch-size 50

Chunking Deep Dive

The ingest script provides intelligent chunking to preserve context:

Paragraph Chunking

• Splits on double newlines
• Preserves paragraph structure
• Adds overlap of 2 sentences between chunks
• Best for: Articles, documentation, blogs

Sentence Chunking

• Splits on sentence boundaries
• Minimal overlap
• Best for: Short content, tweets, quotes

Fixed Chunking

• Fixed 1000 character chunks
• 200 character overlap
• Best for: Code files, logs, unstructured text

Semantic Chunking

• Uses paragraph + header detection
• Preserves document structure
• Best for: Long documents with headers

API Reference

All scripts use Qdrant REST API:

GET    /collections              # List collections
PUT    /collections/{name}       # Create collection
DELETE /collections/{name}       # Delete collection
GET    /collections/{name}       # Collection info
POST   /collections/{name}/points/search     # Search
PUT    /collections/{name}/points           # Upsert points
POST   /snapshots                # Create snapshot
GET    /collections/{name}/snapshots         # List snapshots

Full docs: https://qdrant.tech/documentation/

Performance Tips

1. Batch uploads: ingest.sh automatically batches uploads (default 100)
2. Optimize after bulk insert: bash manage.sh optimize my_collection
3. Use filters: Narrow search scope with metadata filters
4. Set score thresholds: Filter low-quality matches
5. Index metadata: Add payload indexes for faster filtering

Troubleshooting

"Connection refused"

• Check Qdrant is running: curl http://$QDRANT_HOST:$QDRANT_PORT/healthz
• Verify host/port environment variables

"Collection not found"

• List collections: bash manage.sh list
• Check collection name spelling

"No search results"

• Verify documents were ingested: bash manage.sh info my_collection
• Check vector dimensions match (e.g., 1536 for text-embedding-3-small)
• Try lowering score threshold

Embedding errors

• Verify OPENAI_API_KEY is set
• Check API key has quota available
• Verify network access to OpenAI API

Snapshot fails

• Check disk space available
• Verify Qdrant has snapshot permissions
• For large collections, try during low-traffic periods

Requirements

• Qdrant server v1.0+
• curl, python3, bash
• OpenAI API key (for embeddings)
• Network access to Qdrant and OpenAI

See Also

• Qdrant Docs: https://qdrant.tech/documentation/
• OpenAI Embeddings: https://platform.openai.com/docs/guides/embeddings
• Vector Search Guide: https://qdrant.tech/documentation/concepts/search/