omidzamani/dspy-embedding-retrieval

DSPy Embedding Retrieval

Goal

Build semantic retrieval over an application-owned text corpus with dspy.Embedder and dspy.Embeddings.

Basic Hosted Embedder

import dspy

corpus = [
    "DSPy programs are composed from modules.",
    "MIPROv2 optimizes instructions and demonstrations.",
    "RLM explores large contexts with a sandboxed REPL.",
]

embedder = dspy.Embedder("openai/text-embedding-3-small")
search = dspy.Embeddings(corpus=corpus, embedder=embedder, k=2)

result = search("Which optimizer tunes prompts?")
print(result.passages)
print(result.indices)

Use in RAG

class LocalRAG(dspy.Module):
    def __init__(self, retriever):
        super().__init__()
        self.retriever = retriever
        self.answer = dspy.ChainOfThought("context: list[str], question -> answer")

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

Custom Local Embeddings

Wrap any callable that accepts list[str] and returns a 2D numeric array:

from sentence_transformers import SentenceTransformer
import dspy

model = SentenceTransformer("sentence-transformers/static-retrieval-mrl-en-v1")
embedder = dspy.Embedder(model.encode)
search = dspy.Embeddings(corpus=corpus, embedder=embedder, k=5)

Scores, FAISS, and Persistence

Use dspy.EmbeddingsWithScores when downstream logic needs similarity thresholds or reranking.

For corpora at or above the brute_force_threshold default of 20_000, DSPy builds a FAISS index. Install FAISS first:

pip install faiss-cpu

Persist the index when embedding the corpus is expensive:

search.save("./retrieval-index")
loaded = dspy.Embeddings.from_saved("./retrieval-index", embedder=embedder)

Related Skills

• Build a complete pipeline: dspy-rag-pipeline
• Design typed context fields: dspy-signature-designer
• Harden caches: dspy-production-deployment

Best Practices

1. Evaluate retrieval quality separately from answer quality.
2. Keep corpus chunking deterministic and versioned.
3. Persist expensive indexes.
4. Use EmbeddingsWithScores when debugging relevance.
5. Measure memory and latency before enabling FAISS for large corpora.

Official Documentation

• Embedder API: https://dspy.ai/api/models/Embedder/
• Embeddings API: https://dspy.ai/api/tools/Embeddings/