Qdrant MCP Server: Connect an AI Assistant to Vector Search

The Qdrant MCP server lets an AI assistant search a Qdrant vector database directly during a session. Instead of you retrieving documents and pasting them into the chat, the assistant calls the search tool, gets the most relevant results, and continues with the retrieved content in context. Tools like fenbs keep the project rules and decisions that apply to how those results are used, separate from the retrieved documents themselves — so each session has both the right source content and the agreed constraints.

Qdrant is an open-source vector database that stores data points alongside vector embeddings (numeric representations of meaning) and retrieves the most similar ones to a given query. It is used to build retrieval-augmented generation (RAG) systems, semantic search, and recommendation engines. The official Qdrant MCP server is maintained by the Qdrant team.

What the Qdrant MCP server does

The official Qdrant MCP server exposes tools to search collections, retrieve specific points by ID, and list collection information. The assistant sends a natural-language query, the server converts it to an embedding using a configured model, searches the Qdrant collection for the closest matches, and returns the matching documents with their scores and metadata. The assistant then works with those documents in the conversation.

Set up the Qdrant MCP server

The Qdrant MCP server is available as a published package. You need a running Qdrant instance — either local on Docker or on Qdrant Cloud — and an API key if the instance requires authentication. The server also needs access to an embedding model to convert queries to vectors; this is typically an OpenAI embeddings endpoint or a compatible alternative. Pass the Qdrant URL, API key, and embedding configuration as environment variables.

Once configured, add the server to your MCP client configuration. Test by asking the assistant to search a collection with a simple query and confirm it returns relevant results. Check the returned scores: very low similarity scores usually mean either the query is off or the collection was indexed with a different embedding model than the one the server is using.

Understand what vector search returns

Vector search returns the documents most similar to the query embedding, not the documents that literally contain the query terms. This means a search for "contract renewal" might return a document about "subscription extension" if they are semantically close. The ranking is a similarity score, not a confidence score. A score of 0.85 means the document is similar to the query, not that it is 85% correct. Always read the returned content rather than trusting the score alone.

For assistants using Qdrant results to answer factual questions, the retrieved documents are the source material, not the answer. The assistant should cite which documents it used and let the person verify the claims against the original source. Treat the search results as evidence, not as facts.

Handling collection quality issues

Vector search results depend heavily on collection quality. If documents were indexed with a different embedding model than the one used at query time, or if the collection contains outdated documents that were never removed, results will be misleading even when similarity scores look acceptable. Before using a Qdrant collection for important work, verify the embedding model in use, check whether documents are versioned or deduplication exists, and confirm the collection was populated from a trusted source.

For team setups where multiple people populate the same collection, decide in advance what makes a document suitable for inclusion: minimum length, required metadata fields, source restrictions. A collection that anyone can add to without review is hard to trust. Document the inclusion criteria so the team and the assistant know what the collection is supposed to contain.

Keep project rules separate from the document collection

A Qdrant collection holds documents and metadata — it is not a good place for project rules, current task state, or decisions about how to use the retrieved content. Those belong in a system designed for them. If your team's rule is that retrieved content must be verified against a primary source before use in a report, that rule is not in the Qdrant collection. It is in fenbs, where every connected assistant reads it before starting work.

How to use Qdrant with fenbs for project context

Connect both the Qdrant MCP server and fenbs to your assistant. Use Qdrant for source retrieval and fenbs for the rules that govern how those sources are used.

1. Create a free account and start a project for the work that uses the Qdrant collection. Describe what the collection contains and what questions it answers.
2. Write Decisions and rules: what collections the assistant may search, how results should be cited, and when a result must be verified against a primary source before use.
3. After each session, record any collection quality issues as Lessons learned so future sessions know which kinds of queries return unreliable results.
4. Connect your assistant via Integrations so it reads the project rules before searching the collection.
5. Ask the assistant to save what it found and what it verified before ending a session, so the next session continues from the current state.