qdrant-patterns（定量模式）

Qdrant Patterns

Use the qdrant MCP server tools for persistent vector storage and semantic retrieval.

Available Tools

Collection Configuration

The collection name is configured via environment variable:

• COLLECTION_NAME - Set to ${WORKSPACE_PROFILE:-default}_memories

This provides workspace isolation - each profile gets its own collection.

Storing Documents

Store information with the qdrant-store tool:

Tool: qdrant-store
Information: "GitHub REST API uses OAuth tokens for authentication. Personal access tokens (PATs) provide scoped access to repositories, issues, and other resources. Fine-grained PATs offer more granular permissions than classic tokens."
Metadata:
  source: "https://docs.github.com/rest/authentication"
  type: "documentation"
  harvested_at: "2025-01-04"
  tags: "github,api,authentication"

Metadata Best Practices

Always include:

• source - Original URL or file path
• type - Content type (documentation, code, article, etc.)
• harvested_at - ISO date of collection
• tags - Comma-separated searchable keywords

Optional but useful:

• project - Related project name
• language - Programming language if code
• version - API or library version
• summary - Brief content summary

Querying Documents

Semantic Search

Find related content by meaning:

Tool: qdrant-find
Query: "how to authenticate with OAuth"

The tool returns the most semantically similar stored information.

Search Tips

• Use natural language queries
• Be specific about what you're looking for
• The embedding model (fastembed) handles semantic matching

RAG Workflow

1. Check Existing Knowledge

Before researching, query for existing content:

Tool: qdrant-find
Query: "GitHub Actions workflow syntax"

If results are relevant and recent (check metadata), use them. Otherwise, harvest fresh content.

2. Harvest and Store

When gathering new information:

1. Fetch the content (WebFetch, Read, etc.)
2. Extract key information
3. Store in Qdrant with metadata
4. Reference the stored content

Tool: qdrant-store
Information: "<extracted content here>"
Metadata:
  source: "<url or path>"
  type: "documentation"
  harvested_at: "<today's date>"
  tags: "<relevant,keywords>"

3. Retrieve for Context

When answering questions or implementing features:

1. Query Qdrant for relevant documents
2. Include top results in context
3. Cite sources from metadata

Example: Research Workflow

1. Check existing: Query for topic with qdrant-find
2. Assess freshness: Check harvested_at in results
3. Harvest if needed: Fetch new content
4. Store with metadata: Add via qdrant-store
5. Use for response: Include relevant chunks

Tips

• Keep stored information focused (one topic per entry)
• Use consistent metadata schemas
• Include enough context in each entry to be useful standalone
• Use descriptive tags for easier filtering
• Check existing knowledge before harvesting new content