Qdrant Vector Store for LlamaIndex

0.10.0 · active · verified Thu Apr 16

The `llama-index-vector-stores-qdrant` library provides an integration for using Qdrant as a vector store within the LlamaIndex framework. It enables users to store and retrieve vector embeddings efficiently for building Retrieval-Augmented Generation (RAG) applications. This integration supports various Qdrant features, including hybrid search capabilities, and is part of the LlamaIndex v0.10.0 ecosystem which adopted a modular architecture with separate integration packages.

Common errors

```
AttributeError: 'AsyncQdrantClient' object has no attribute 'search'
```
cause This typically occurs when using an incompatible version of `qdrant-client` with `llama-index-vector-stores-qdrant`. Newer `qdrant-client` versions refactored the `search` method's location or API.

fix Downgrade `qdrant-client` to a compatible version (e.g., `pip install qdrant-client<1.16.0`) or upgrade `llama-index-vector-stores-qdrant` to a version that supports the newer `qdrant-client` APIs.
```
ModuleNotFoundError: No module named 'llama_index.vector_stores.qdrant'
```
cause The `llama-index-vector-stores-qdrant` package is not installed or not correctly installed in the current environment.

fix Install the package: `pip install llama-index-vector-stores-qdrant`.
```
ModuleNotFoundError: No module named 'qdrant_client'
```
cause The `qdrant-client` library, which is a peer dependency, is not installed.

fix Install the `qdrant-client`: `pip install qdrant-client`.
```
PydanticValidationError (or similar validation error) when initializing QdrantClient with local data path.
```
cause This can happen with newer `qdrant-client` versions (e.g., >=1.16.0) that introduced breaking changes to the internal data model for creating collections, leading to schema validation failures.

fix Constrain your `qdrant-client` version to an earlier, compatible one (e.g., `pip install qdrant-client<1.16.0`) or ensure your `llama-index-vector-stores-qdrant` version is updated to support the latest `qdrant-client` schemas.

Warnings

breaking LlamaIndex v0.10.0 introduced a major refactor, splitting core components and integrations into separate packages. The `ServiceContext` abstraction was deprecated. This means imports and configuration patterns have changed significantly.
Fix: Migrate `ServiceContext` usage to direct configuration via `Settings` or by passing components directly to index/query engine constructors. Update imports from `llama_index.<component>` to `from llama_index.core import <Component>` or `from llama_index.<integration_type>.<integration_name> import <Component>`.
breaking Future versions of `qdrant-client` (e.g., 1.16.0 as of a past issue) introduced breaking API changes (e.g., removal of `AsyncQdrantClient.search` method, Pydantic validation issues for local clients) that can cause `AttributeError` or validation failures.
Fix: If encountering errors, try constraining your `qdrant-client` version to one known to be compatible (e.g., `qdrant-client<1.16.0` if using an older `llama-index` setup) or upgrade `llama-index-vector-stores-qdrant` to a version that explicitly supports the newer `qdrant-client` API.
gotcha The very first retrieval query from a LlamaIndex index backed by Qdrant can be significantly slower (e.g., >10 seconds) than subsequent queries. This is due to Qdrant's internal index initialization and connection setup overhead.
Fix: To 'warm up' the system and avoid a slow first user-facing query, perform a dummy query shortly after the `VectorStoreIndex` is initialized but before serving user requests.
gotcha For hybrid search functionality (combining sparse and dense vectors) with Qdrant, additional dependencies like `fastembed` (for local sparse embedding generation) or `transformers` might be required.
Fix: Install necessary packages: `pip install fastembed` (for CPU-friendly local embeddings) or `pip install "transformers[torch]"` if using HuggingFace models for sparse embeddings.

Install

pip install llama-index-vector-stores-qdrant qdrant-client Base Installation
pip install llama-index-vector-stores-qdrant qdrant-client fastembed For Hybrid Search / Local Embeddings

Imports

QdrantVectorStore

from llama_index.vector_stores.qdrant import QdrantVectorStore

QdrantClient
```
from qdrant_client import QdrantClient
```

AsyncQdrantClient

from qdrant_client import AsyncQdrantClient

ServiceContext
wrong:
```
from llama_index.service_context import ServiceContext
```
correct:
```
from llama_index.core import ServiceContext
```
As of LlamaIndex v0.10.0, ServiceContext is deprecated and its components should be configured via `Settings` or passed directly. If used, it's now in `llama_index.core`.

Quickstart

This quickstart demonstrates how to initialize a Qdrant in-memory client, create a `QdrantVectorStore`, configure LlamaIndex settings with an embedding model (defaults to OpenAI), index a few dummy documents, and then query the index. It highlights the modular setup in LlamaIndex v0.10+.

import os
from llama_index.core import VectorStoreIndex, Document
from llama_index.vector_stores.qdrant import QdrantVectorStore
from qdrant_client import QdrantClient
from llama_index.embeddings.openai import OpenAIEmbedding
from llama_index.core import Settings

# Ensure you have your OpenAI API key set as an environment variable
# os.environ["OPENAI_API_KEY"] = "sk-..."

# Fallback for API key or local Qdrant for quick demo without remote setup
if os.environ.get("OPENAI_API_KEY") is None or os.environ.get("OPENAI_API_KEY") == "":
    print("Warning: OPENAI_API_KEY not set. Using a dummy key for example purposes. This will fail if you try to use OpenAI models.")
    os.environ["OPENAI_API_KEY"] = os.environ.get("OPENAI_API_KEY", "sk-dummy-key")

# Initialize Qdrant client (local in-memory for quick start, or connect to a server)
# For persistent storage, use QdrantClient(path="./qdrant_data") or connect to a running Qdrant instance
client = QdrantClient(location=":memory:") # In-memory Qdrant instance

# Create a QdrantVectorStore instance
vector_store = QdrantVectorStore(client=client, collection_name="my_documents")

# Configure LlamaIndex settings (important for v0.10+)
Settings.embed_model = OpenAIEmbedding()

# Create a dummy document
documents = [
    Document(text="LlamaIndex is a data framework for LLM applications."),
    Document(text="Qdrant is a vector similarity search engine."),
    Document(text="Combining LlamaIndex with Qdrant enables powerful RAG systems."),
]

# Build the VectorStoreIndex
index = VectorStoreIndex.from_documents(documents, vector_store=vector_store)

# Query the index
query_engine = index.as_query_engine()
response = query_engine.query("What is LlamaIndex?")
print(f"Response: {response}")

response = query_engine.query("What is Qdrant used for?")
print(f"Response: {response}")

view raw JSON →