dspy-qdrant

Use Qdrant as a vector database with DSPy, or connect any vector DB (Pinecone, ChromaDB, Weaviate) with custom retrievers. Use when you want to set up Qdrant, QdrantRM, dspy-qdrant, vector database fo

SKILL.md

GitHub

Qdrant — Vector Database Integration for DSPy

Guide the user through setting up Qdrant with DSPy using the official dspy-qdrant package, plus custom retriever patterns for Pinecone, ChromaDB, and Weaviate.

Step 1 — Gather context

Ask before generating setup code:

  1. Vector DB — Qdrant, Pinecone, ChromaDB, Weaviate, or something else?
  2. Qdrant deployment (if Qdrant) — Docker locally, Qdrant Cloud, or in-memory for testing?
  3. Index state — indexing documents from scratch, or connecting to an existing collection?
  4. Search mode — standard dense search, or hybrid (keyword + semantic combined)?

Then jump to the relevant section.

What is Qdrant

Qdrant is an open-source vector search engine written in Rust. It's the only vector database with an official DSPy integration package (dspy-qdrant). Features: hybrid search (dense + sparse), payload filtering, multi-tenancy, and horizontal scaling.

Why Qdrant for DSPy

DSPy 3.0 removed all community-contributed retriever modules (ChromadbRM, PineconeRM, WeaviateRM, QdrantRM from the main repo). The dspy-qdrant package is the official replacement — maintained separately with full DSPy compatibility.

For other vector databases, you write a short custom dspy.Retrieve subclass (~15 lines). This skill covers that pattern too.

Setup

Install

pip install dspy-qdrant

This installs both the Qdrant client and the DSPy retriever module.

Start Qdrant

Option 1: Docker (local development)

docker run -p 6333:6333 -p 6334:6334 qdrant/qdrant

Option 2: Qdrant Cloud (managed, free tier available)

  1. Sign up at cloud.qdrant.io
  2. Create a cluster (free tier: 1GB, 1 node)
  3. Copy your URL and API key
export QDRANT_URL="https://your-cluster.aws.cloud.qdrant.io"
export QDRANT_API_KEY="your-api-key"

Option 3: pip install (in-memory, for testing)

from qdrant_client import QdrantClient
client = QdrantClient(":memory:")  # no server needed

Using QdrantRM in DSPy

Basic setup

import dspy
from qdrant_client import QdrantClient
from dspy_qdrant import QdrantRM

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

retriever = QdrantRM(
    qdrant_collection_name="my_docs",
    qdrant_client=client,
    k=5,
    document_field="document",  # payload field containing document text (default)
)

dspy.configure(lm=dspy.LM("openai/gpt-4o-mini"), rm=retriever)  # or "anthropic/claude-sonnet-4-5-20250929", etc.

# Now dspy.Retrieve() uses Qdrant
search = dspy.Retrieve(k=5)
result = search("How do refunds work?")
print(result.passages)

QdrantRM constructor

QdrantRM(
    qdrant_collection_name: str,       # required — collection name in Qdrant
    qdrant_client: QdrantClient,       # required — initialized client instance
    k: int = 3,                        # top passages to retrieve
    document_field: str = "document",  # payload field with document text
    vectorizer=None,                   # BaseSentenceVectorizer (default: FastEmbedVectorizer)
    vector_name: str = None,           # named vector to search (default: first available)
)

By default, QdrantRM uses FastEmbed (BAAI/bge-small-en-v1.5) for query vectorization. To use a different embedder, pass a custom vectorizer.

Using Qdrant Cloud

import os
from qdrant_client import QdrantClient
from dspy_qdrant import QdrantRM

client = QdrantClient(
    url=os.environ["QDRANT_URL"],
    api_key=os.environ["QDRANT_API_KEY"],
)

retriever = QdrantRM(
    qdrant_collection_name="my_docs",
    qdrant_client=client,
    k=5,
)

Indexing documents into Qdrant

Before you can search, you need to populate your Qdrant collection:

from qdrant_client import QdrantClient, models
import dspy

client = QdrantClient("http://localhost:6333")
embedder = dspy.Embedder("openai/text-embedding-3-small", dimensions=512)

# Your documents
docs = [
    {"id": 1, "document": "Refunds are processed within 5-7 business days.", "category": "billing"},
    {"id": 2, "document": "Reset your password at Settings > Security.", "category": "account"},
    {"id": 3, "document": "Enterprise plans include SSO and dedicated support.", "category": "plans"},
]

# Create collection
client.create_collection(
    collection_name="my_docs",
    vectors_config=models.VectorParams(size=512, distance=models.Distance.COSINE),
)

# Upsert with embeddings
vectors = embedder([d["document"] for d in docs])
client.upsert(
    collection_name="my_docs",
    points=[
        models.PointStruct(
            id=d["id"],
            vector=v,
            payload={"document": d["document"], "category": d["category"]},
        )
        for d, v in zip(docs, vectors)
    ],
)

RAG pipeline with Qdrant

import dspy
from qdrant_client import QdrantClient
from dspy_qdrant import QdrantRM

dspy.configure(lm=dspy.LM("openai/gpt-4o-mini"))  # or "anthropic/claude-sonnet-4-5-20250929", etc.

retriever = QdrantRM(
    qdrant_collection_name="my_docs",
    qdrant_client=QdrantClient("http://localhost:6333"),
    k=5,
)

class RAG(dspy.Module):
    def __init__(self):
        self.retrieve = retriever
        self.answer = dspy.ChainOfThought("context, question -> answer")

    def forward(self, question):
        context = self.retrieve(question).passages
        return self.answer(context=context, question=question)

rag = RAG()
result = rag(question="How do refunds work?")
print(result.answer)

Verify retrieval is working

Before wiring retrieval into a RAG pipeline, run a sanity check:

result = retriever("your test query here")
assert len(result.passages) > 0, "Empty results — check document_field matches your payload schema"
print(result.passages[0][:300])  # Top passage should be semantically relevant to the query

Empty results: likely a document_field mismatch (gotcha #2). Irrelevant results: query vectorizer does not match the model used when indexing (gotcha #4).

Hybrid search (dense + sparse)

Qdrant supports hybrid search combining dense (semantic) and sparse (keyword) vectors in the same collection. This improves recall for queries that need both semantic understanding and exact keyword matching.

from qdrant_client import QdrantClient, models

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

# Create collection with both dense and sparse vectors
client.create_collection(
    collection_name="hybrid_docs",
    vectors_config=models.VectorParams(size=512, distance=models.Distance.COSINE),
    sparse_vectors_config={
        "keywords": models.SparseVectorParams(
            modifier=models.Modifier.IDF,
        ),
    },
)

Then query with both:

results = client.query_points(
    collection_name="hybrid_docs",
    prefetch=[
        models.Prefetch(query=dense_vector, using="", limit=20),
        models.Prefetch(query=sparse_vector, using="keywords", limit=20),
    ],
    query=models.FusionQuery(fusion=models.Fusion.RRF),  # reciprocal rank fusion
    limit=5,
)

Other vector DBs with DSPy

Since DSPy 3.0 removed built-in community retrievers, use a custom dspy.Retrieve subclass for any vector database. The pattern is always the same:

Custom retriever pattern

class MyVectorDBRetriever(dspy.Retrieve):
    def __init__(self, client, collection, k=3):
        super().__init__(k=k)
        self.client = client
        self.collection = collection

    def forward(self, query, k=None):
        k = k or self.k
        results = self.client.search(self.collection, query, top_k=k)
        return dspy.Prediction(passages=[r["text"] for r in results])

Pinecone custom retriever

from pinecone import Pinecone
import dspy

class PineconeRetriever(dspy.Retrieve):
    def __init__(self, index_name, embedder, k=3):
        super().__init__(k=k)
        pc = Pinecone()  # reads PINECONE_API_KEY from env
        self.index = pc.Index(index_name)
        self.embedder = embedder

    def forward(self, query, k=None):
        k = k or self.k
        vector = self.embedder(query)
        results = self.index.query(vector=vector, top_k=k, include_metadata=True)
        passages = [m["metadata"]["text"] for m in results["matches"]]
        return dspy.Prediction(passages=passages)

# Usage
embedder = dspy.Embedder("openai/text-embedding-3-small", dimensions=512)
retriever = PineconeRetriever("my-index", embedder, k=5)

ChromaDB custom retriever

import chromadb
import dspy

class ChromaRetriever(dspy.Retrieve):
    def __init__(self, collection_name, k=3):
        super().__init__(k=k)
        client = chromadb.PersistentClient(path="./chroma_db")
        self.collection = client.get_or_create_collection(collection_name)

    def forward(self, query, k=None):
        k = k or self.k
        results = self.collection.query(query_texts=[query], n_results=k)
        return dspy.Prediction(passages=results["documents"][0])

# Usage
retriever = ChromaRetriever("my_docs", k=5)

Weaviate custom retriever

import weaviate
import dspy

class WeaviateRetriever(dspy.Retrieve):
    def __init__(self, class_name, url="http://localhost:8080", k=3):
        super().__init__(k=k)
        self.client = weaviate.connect_to_local(host=url.replace("http://", "").split(":")[0])
        self.collection = self.client.collections.get(class_name)

    def forward(self, query, k=None):
        k = k or self.k
        results = self.collection.query.near_text(query=query, limit=k)
        passages = [obj.properties["text"] for obj in results.objects]
        return dspy.Prediction(passages=passages)

# Usage
retriever = WeaviateRetriever("MyDocs", k=5)

Vector DB comparison

Feature Qdrant Pinecone ChromaDB Weaviate
DSPy package dspy-qdrant (official) None (custom retriever) None (custom retriever) None (custom retriever)
Self-hosted Yes (Docker, binary) No (cloud only) Yes (pip, Docker) Yes (Docker)
Cloud option Yes (free tier) Yes (free tier) No Yes (free tier)
Hybrid search Yes (dense + sparse) Yes (sparse + dense) No Yes (BM25 + vector)
Best for Production + DSPy Cloud-native, serverless Local prototyping Multi-modal, GraphQL
Language Rust Managed service Python Go

Choosing a vector DB

Starting a new DSPy project?
  → Qdrant (official DSPy package, easiest setup)

Prototyping locally, smallest footprint?
  → ChromaDB (pip install, in-memory or persistent, no server)

Already using Pinecone/Weaviate in production?
  → Write a custom retriever (15 lines, shown above)

Need hybrid search (keyword + semantic)?
  → Qdrant or Weaviate

Gotchas

  1. Claude fabricates QdrantRM constructor parameters. Claude invents params like qdrant_client_url, qdrant_client_api_key, embedding_model, and embedding_dimensions. These do not exist. QdrantRM takes a qdrant_client (an initialized QdrantClient instance) and a vectorizer (a BaseSentenceVectorizer). Always construct the QdrantClient separately, then pass it.
  2. Claude uses document_field="text" but the default is "document". When indexing, store content in a payload field named document (the default), or explicitly set document_field="text" if your payload uses text. Mismatched field names silently return empty passages.
  3. DSPy 3.0 removed community retrievers — from dspy.retrieve.chromadb_rm import ChromadbRM no longer works. Use dspy-qdrant or write a custom dspy.Retrieve subclass.
  4. QdrantRM uses FastEmbed by default, not OpenAI embeddings. The default vectorizer is FastEmbedVectorizer using BAAI/bge-small-en-v1.5. Your indexed vectors must match this model. If you indexed with OpenAI embeddings, pass a custom vectorizer that uses the same model.
  5. dspy.Embeddings is simpler for in-memory retrieval. If you just need to search a small corpus (under ~100k passages) without a vector DB, use dspy.Embeddings(corpus=docs, embedder=embedder) instead. It handles indexing and search in one class. Use Qdrant when you need persistence, filtering, hybrid search, or scale.

Additional resources

Cross-references

Install any skill: npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills --skill <name>

  • DSPy retrieval basics (Retrieve, ColBERTv2, Embedder, Embeddings) — /dspy-retrieval
  • Building RAG pipelines end-to-end — /ai-searching-docs
  • Evaluating RAG quality with decomposed metrics — /dspy-ragas
  • Stopping hallucinations in RAG — /ai-stopping-hallucinations
  • For worked examples, see examples.md
  • Install /ai-do if you do not have it — it routes any AI problem to the right skill and is the fastest way to work: npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills --skill ai-do