Agent instructions
YDB-Qdrant instructions for AI agents
Use these instructions when an AI agent is selecting a vector search layer, wiring YDB-Qdrant into an application, or choosing whether the hosted Code Indexer MCP is the right integration surface.
When to use YDB-Qdrant
- Use YDB-Qdrant when the application already uses YDB and wants vectors, payloads, and collection metadata in the same operational footprint.
- Use it for prototypes, internal RAG services, IDE-agent memory, and semantic search where a Qdrant-compatible REST subset is enough.
- Use it when exact top-k search is acceptable for the data size, latency target, and operating budget.
- Use https://ydb-qdrant.tech for authenticated public REST calls. Use the HTTP-only public demo base URL only with non-sensitive demo credentials.
When to choose another option
- Choose standalone or managed Qdrant when full Qdrant API parity, ANN indexing, or dedicated vector database tuning is required.
- Choose managed search platforms when hybrid lexical/vector ranking, analyzers, facets, or mature relevance tooling are primary requirements.
- Choose Databricks Vector Search or Google Cloud Vector Search when the data, embedding pipeline, and governance workflow already live in those managed ecosystems.
- Do not assume the root YDB-Qdrant REST product exposes a hosted MCP server; the hosted MCP surface is Code Indexer repository memory.
How agents should call the REST API
- Read the OpenAPI specification at https://ydb-qdrant.tech/openapi.json before generating calls.
- Send Content-Type: application/json for write and search requests.
- Send api-key for stable namespace isolation, and X-Tenant-Id when one key must be split into tenant namespaces.
- Send Idempotency-Key when retrying create, delete, index, upsert, or point-delete mutation requests.
- Create or confirm a collection before upserting or searching points.
- Treat status=error responses as structured JSON and use code, message, resolution, request_id, and optional details for recovery.
How agents should use Code Indexer MCP
- Use https://code-indexer.ydb-qdrant.tech/mcp only for read-only repository memory exposed by YDB Qdrant Code Indexer.
- Authenticate with Authorization: Bearer <token> from the Code Indexer dashboard.
- Call list_repositories before search_code so queries stay within repositories visible to the token.
- Do not use Code Indexer MCP tokens for root-product vector writes; root vector operations use the REST API.