Dense Search Over Your Own Embeddings, and What Hybrid Needs

from openai import OpenAI
from mixpeek import Mixpeek

openai = OpenAI(api_key="your-openai-key")
mvs = Mixpeek(api_key="your-mvs-key")

NAMESPACE = "my-namespace"

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

# The one step a plain BYO upsert does not do for you: if you will ever want BM25
# over these documents, the TEXT payload index has to exist before the first upsert.
mvs.namespaces.create(
    namespace_name=NAMESPACE,
    payload_indexes=[{"field_name": "content", "type": "text"}],
)

# Upsert documents with your own dense vectors. Vectors are named; "dense" is the
# name you will search by. The text goes in payload so it is stored with the document.
documents = [
    {"text": "FastAPI uses Pydantic v2 for data validation and serialization", "topic": "python"},
    {"text": "Express.js middleware handles request/response transformations", "topic": "node"},
    {"text": "FastAPI supports async/await natively with Starlette ASGI", "topic": "python"},
    {"text": "Django ORM provides database abstraction with QuerySet API", "topic": "python"},
]

mvs.namespaces.documents.upsert(
    namespace=NAMESPACE,
    collection_id="col_your_collection",
    documents=[{
        "document_id": f"doc-{i}",
        "vectors": {"dense": embed(doc["text"])},
        "payload": {"content": doc["text"]},
        "metadata": {"topic": doc["topic"]},
    } for i, doc in enumerate(documents)],
)

# Dense search by raw vector. feature_uri MUST name your vector ("dense"); the
# default is a text_extractor index and would silently return 0 results.
query_text = "FastAPI Pydantic validation"
results = mvs.features.search(
    collection_identifiers=["col_your_collection"],
    feature_uri="dense",
    query={"input_mode": "vector", "vector": embed(query_text)},
    top_k=5,
)

for doc in results:
    print(f"{doc['score']:.3f} | {doc['metadata'].get('topic', '')} | {doc['payload']['content'][:80]}")