Qdrant — Module 14.2

How This Lesson Fits the Module & Volume

Qdrant closes Module 14.2: a Rust-based open-source vector database known for fast filtered search, rich JSON payloads, and clean APIs (REST/gRPC). Many production RAG stacks pair Qdrant with LangChain or LlamaIndex.

Use this lecture to synthesize the local vs managed and filter/scale trade-offs across FAISS → Chroma → Pinecone → Milvus → Weaviate → Qdrant before orchestration frameworks in 14.3.

Learning Objectives

By the end of this lesson, students should be able to:

Create Qdrant collections with distance metrics and point payloads.
Upsert points and query with Filter conditions on payload fields.
Explain payload indexes for filter performance at scale.
Compare Qdrant Cloud vs self-hosted Docker/K8s.
Select among Module 14.2 stores using a decision table.
Bridge vector storage into Module 14.3 orchestration for full RAG apps.

Definition

Qdrant is an open-source vector similarity search engine that stores vectors with arbitrary JSON payloads, supports advanced filtering during ANN search, and exposes REST and gRPC APIs for high-performance retrieval workloads.

Module 14.2 Comparison Snapshot

Store	Sweet spot	Watch-outs
FAISS	Local ANN / research	No DB features
Chroma	DX + local RAG	Scale ceiling
Pinecone	Managed SaaS	Cost / lock-in
Milvus	Huge self-host fleets	Ops complexity
Weaviate	Hybrid + modules	Schema/module learning
Qdrant	Filtered search + payloads	You still size cluster

Collections, Points, Payloads

from qdrant_client import QdrantClient
from qdrant_client.models import (
    Distance, VectorParams, PointStruct, Filter, FieldCondition, MatchValue
)

client = QdrantClient(url="http://localhost:6333")

client.recreate_collection(
    collection_name="rag_chunks",
    vectors_config=VectorParams(size=384, distance=Distance.COSINE),
)

client.upsert(
    collection_name="rag_chunks",
    points=[
        PointStruct(
            id=1,
            vector=emb_1,
            payload={"text": "Reset MFA in Security settings.", "product": "auth", "tenant": "acme"},
        ),
        PointStruct(
            id=2,
            vector=emb_2,
            payload={"text": "Download invoices from Billing.", "product": "billing", "tenant": "acme"},
        ),
    ],
)

hits = client.search(
    collection_name="rag_chunks",
    query_vector=query_emb,
    limit=5,
    query_filter=Filter(
        must=[
            FieldCondition(key="tenant", match=MatchValue(value="acme")),
            FieldCondition(key="product", match=MatchValue(value="auth")),
        ]
    ),
)
for h in hits:
    print(h.id, h.score, h.payload["text"])

Filtering and Scale

Payload design

Stable keys

Payload index

Speed filters

Filtered ANN

Search + constraints

RAG context

text from payload

Qdrant applies filters during search rather than only as a post-pass—critical when tenancy must not leak neighbors. At scale, index frequently filtered fields; unindexed payload scans hurt latency.

Strengths

Excellent filtered vector search
Flexible JSON payloads
Strong OSS + cloud story
Popular RAG framework adapters

Tradeoffs

Cluster sizing still on you (self-host)
Payload bloat increases storage
Need discipline on filter indexes
Not a full document warehouse

Common Misconception

“Storing full PDFs in every payload is fine.” Keep chunk text lean; large blobs belong in object storage with IDs in payload. Fat payloads raise RAM/disk and slow snapshots.

Knowledge Check

Short Answer: What does a Qdrant point contain? Answer: Id, vector, and optional JSON payload.
True/False: Qdrant supports filtering during similarity search. Answer: True.
Multiple Choice: Distance.COSINE is set on: (a) the LLM, (b) collection vector params, (c) CSS. Answer: (b).
Short Answer: Why index payload fields used in filters? Answer: To keep filtered search latency low at scale.
True/False: Qdrant is only available as closed-source SaaS. Answer: False—it is open source (with cloud option).
Multiple Choice: Multi-tenant isolation often uses: (a) payload filters / collections, (b) random seeds, (c) dropout. Answer: (a).
Short Answer: When prefer Pinecone over self-hosted Qdrant? Answer: When you want managed ops / less cluster work.
Short Answer: What Module follows vector DBs? Answer: 14.3 LangChain & orchestration frameworks.
Multiple Choice: Oversized payloads mainly hurt: (a) ethics boards, (b) storage/memory and ops, (c) tokenization rules only. Answer: (b).
True/False: Frameworks like LangChain commonly integrate Qdrant as a vector store. Answer: True.

Key Takeaways

Qdrant excels at vector search with rich, filterable payloads.
Design lean payloads and index filter keys for production latency.
Use the Module 14.2 table to pick local, managed, or distributed stores.
Storage alone is not an app—orchestration comes next.
Next module: LangChain wires models, retrievers, and tools into pipelines.

Trainer’s Guide

Capstone mini: Teams pick FAISS, Chroma, or Qdrant for a 2-tenant FAQ bot and justify filters, ops, and cost.

Transition: Preview how the same Qdrant collection becomes a LangChain VectorStore in 14.3.

Recap: Qdrant finishes the vector DB tour with filtered payloads. Continue to LangChain.