🗓️Combining Multiple Embeddings for Better Retrieval Outcomes

How Qyver Represents Complex Data More Effectively

Retrieving high-quality results from LLM-powered vector searches on complex, embedded data isn’t easy. The traditional approach looks like this:

1. Embed all entity data into a single vector using a model (e.g., Hugging Face, proprietary models).

2. Run a vector search to find the top X nearest neighbors.

3. Rerank results using:

• Additional contextual filters

• Information not captured in the embedding (e.g., filtering out out-of-stock items)

This method has serious limitations:

• Data loss: A single embedding struggles to capture multiple attributes, reducing relevance.

• Inefficient reranking: Reranking with cross-encoders adds latency, slowing down production.

• High computational cost: Comparing all results pairwise to refine rankings is resource-intensive.

Is there a way to achieve high-quality retrieval without data loss and latency overhead? Yes!

---

A Smarter Approach: Multimodal Embeddings with Qyver

Instead of embedding all entity data into a single vector, Qyver Spaces allow you to:

• Embed each attribute separately based on its modality.

• Concatenate these embeddings into a multimodal vector.

• Capture the full complexity of data without losing information.

This results in:

• Better quality retrieval—more relevant, complete search results.

• No need for expensive reranking—eliminates computational overhead.

• Faster processing—retrieval is 10x faster, reducing latency from hundreds to tens of milliseconds.

Example: Searching Multimodal Data Efficiently

Consider a simple dataset where each paragraph has a text body and a like count:

• Paragraph 1: "Glorious animals live in the wilderness." → likes: 10

• Paragraph 2: "Growing computation power enables advancements in AI." → likes: 75

Using Qyver’s Spaces, we represent each attribute individually:

@schema
class Paragraph:
    id: IdField
    body: String
    like_count: Integer

Instead of embedding everything into a single vector, we use separate Spaces:

body_space = TextSimilaritySpace(
    text=paragraph.body, model="sentence-transformers/all-mpnet-base-v2"
)

like_space = NumberSpace(
    number=paragraph.like_count, min_value=0, max_value=100, mode=Mode.MAXIMUM
)

These Spaces are then combined into a single index:

paragraph_index = Index([body_space, like_space])

---

Efficient Querying Without Reranking

Once indexed, we query for relevance efficiently:

query = (
    Query(paragraph_index)
    .find(paragraph)
    .similar(body_space.text, Param("query_text"))
)

result = app.query(
    query,
    query_text="What makes the AI industry go forward?",
)

Expected Output:

body	like_count	id
Growing computation power enables advancements in AI.	75	paragraph-2
Glorious animals live in the wilderness.	10	paragraph-1

Why is this better?

1. Structured & unstructured data are embedded separately but stored in a single searchable index.

2. No need for reranking—Qyver Spaces ensure relevance is captured upfront.

3. ️Faster & more accurate—reduces processing overhead while improving retrieval precision.

---

Scaling Beyond Simple Use Cases

Qyver is designed for real-world complexity, handling:

• E-commerce recommendations → Blending product descriptions, user preferences, and purchase history.

• Search & discovery → Multimodal indexing for text, images, and metadata.

• LLM-powered applications → Enhancing RAG (Retrieval-Augmented Generation) with structured embeddings.

---

Final Thoughts

Instead of relying on reranking, filtering, or expensive post-processing, Qyver optimizes embeddings upfront—leading to faster, higher-quality retrieval.

Want to try it? Check out the notebook and see the power of multimodal embeddings in action.

Like what we’re doing? Give us a star!