How to Intermediate · 4 min read

How to use DSPy with Pinecone

Q: How to use DSPy with Pinecone

Use dspy to define your AI tasks declaratively and combine it with pinecone for vector storage and retrieval by initializing both clients and implementing a retrieval function that queries Pinecone vectors within your dspy workflow. This enables seamless integration of LLM predictions with Pinecone's vector search capabilities.

Quick answer

Use dspy to define your AI tasks declaratively and combine it with pinecone for vector storage and retrieval by initializing both clients and implementing a retrieval function that queries Pinecone vectors within your dspy workflow. This enables seamless integration of LLM predictions with Pinecone's vector search capabilities.

PREREQUISITES

Python 3.8+
OpenAI API key (free tier works)
Pinecone API key
pip install dspy pinecone-client openai

Setup

Install the required packages and set environment variables for your OpenAI and Pinecone API keys.

Install packages: pip install dspy pinecone-client openai
Set environment variables: OPENAI_API_KEY and PINECONE_API_KEY

bash

pip install dspy pinecone-client openai

Step by step

This example shows how to define a simple dspy signature for question answering, initialize Pinecone to store and query embeddings, and integrate Pinecone retrieval into the dspy prediction workflow.

python

import os
import dspy
from openai import OpenAI
import pinecone

# Initialize OpenAI client for DSPy
client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])

# Configure DSPy to use OpenAI
lm = dspy.LM("openai/gpt-4o-mini", api_key=os.environ["OPENAI_API_KEY"])
dspy.configure(lm=lm)

# Initialize Pinecone client
import pinecone
pc = pinecone.Pinecone(api_key=os.environ["PINECONE_API_KEY"])
index_name = "example-index"

# Create or connect to Pinecone index
if index_name not in pc.list_indexes():
    pc.create_index(index_name, dimension=1536)  # dimension must match embedding size
index = pc.Index(index_name)

# Define DSPy signature
class QA(dspy.Signature):
    question: str = dspy.InputField()
    answer: str = dspy.OutputField()

# Define a retrieval function using Pinecone

def embed_text(text: str) -> list[float]:
    response = client.embeddings.create(model="text-embedding-3-small", input=text)
    return response.data[0].embedding

# Define a DSPy predict function that uses Pinecone retrieval
@dspy.Predict(QA)
def qa_with_retrieval(question: str) -> str:
    # Embed the question
    q_vector = embed_text(question)
    # Query Pinecone for top 3 relevant documents
    results = index.query(vector=q_vector, top_k=3, include_metadata=True)
    # Combine retrieved texts
    context = "\n".join([match['metadata']['text'] for match in results['matches']])
    # Prompt with context
    prompt = f"Context:\n{context}\n\nQuestion: {question}\nAnswer:" 
    # Use DSPy LM to generate answer
    answer = lm(prompt)
    return answer

# Example usage
if __name__ == "__main__":
    # Example: Upsert some documents to Pinecone
    docs = [
        ("doc1", embed_text("Python is a programming language."), {"text": "Python is a programming language."}),
        ("doc2", embed_text("Pinecone is a vector database."), {"text": "Pinecone is a vector database."}),
        ("doc3", embed_text("DSPy is a declarative LLM programming library."), {"text": "DSPy is a declarative LLM programming library."})
    ]
    index.upsert(items=docs)

    # Run the QA
    result = qa_with_retrieval(question="What is Pinecone?")
    print("Answer:", result)

output

Answer: Pinecone is a vector database.

Common variations

You can adapt this integration by:

Using async DSPy calls with async def and await for embedding and LM calls.
Changing the embedding model to a different OpenAI or custom model.
Using other Pinecone index configurations or namespaces for multi-tenant setups.
Integrating DSPy with other vector databases by replacing the Pinecone client calls.

Troubleshooting

If you see errors like Index does not exist, ensure you have created the Pinecone index with the correct dimension matching your embedding model.

If embeddings are empty or incorrect, verify your OpenAI API key and model name.

For connection issues, check your Pinecone environment and API key validity.

✅

Key Takeaways

Use dspy to declaratively define AI tasks and integrate with Pinecone for vector search.
Initialize both OpenAI and Pinecone clients with environment API keys for secure access.
Embed queries with OpenAI embeddings and retrieve relevant vectors from Pinecone before LLM generation.
Adjust embedding dimensions and Pinecone index settings to match your use case.
Handle common errors by verifying index existence, API keys, and environment configurations.

Verified 2026-04 · openai/gpt-4o-mini, text-embedding-3-small

Verify ↗